NONDESTRUCTIVE EVALUATION OF FRESH CHESTNUT INTERNAL
QUALITY USING X-RAY COMPUTED TOMOGRAPHY (CT)
By
Irwin R. Donis-González

A DISSERTATION
Submitted to
Michigan State University
In partial fulfillment of the requirements
For the degree of
Biosystems Engineering – Doctor of Philosophy
2013

ABSTRACT
NONDESTRUCTIVE EVALUATION OF FRESH CHESTNUT INTERNAL DECAY
USING X-RAY COMPUTED TOMOGRAPHY (CT)
By
Irwin R. Donis-González
Internal decay is an important quality attribute in chestnuts (Castanea
spp.). Worldwide, internal decay is mainly caused by microorganism attack and
physiological cell breakdown. It is problematic for the industry, and impacts
consumer satisfaction, shelf life, and proper storage. Currently, destructive
techniques

can

be

employed

to

evaluate

fresh

chestnut

internal

quality. However, clearly not all produce can be evaluated. In commercial
situations, decayed chestnuts are eliminated by their proclivity to float in water.
Nonetheless, performance significantly varies between species and throughput,
making this floating method unreliable for sorting purposes. Thus, the overall
objective of the study is to develop the methods to nondestructively visualize and
automatically classify fresh chestnuts, based on their internal quality, using X-ray
CT imaging.
In this study, medical grade computed tomography (CT) was used to
obtain transversal two-dimensional (2D) images from fresh chestnuts (cv.
‘Colossal’ and Chinese seedlings). If the information obtained by the CT
scanning of fresh chestnuts is to be used in an industrial setting for in-line
sorting, automated interpretation of CT images is essential. For this purpose: (1)
Chestnut CT image quality was optimized by studying the combined effect of

image acquisition parameters (voltage – 120 kV, current – 170 mA and slice
thickness – 2.5 mm) using response surface methodology, (2) effective image
visualization techniques to infer fresh chestnut internal quality attributes were
established, and (3) an image analysis algorithm for the automatic classification
of CT images obtained from 2848 fresh chestnuts, during the harvesting years
from 2009 to 2012, was developed and tested.
The CT imaging system provided high-resolution and high-contrast images
of the internal structure and components of fresh chestnuts. Approximately 50
original CT image slices (stack) were obtained per chestnut, from three different
planes (angular orientations) across the longitudinal (Z) (XY-plane-slice),
horizontal (YZ-plane-slice) and vertical (XZ-plane-slice) axes. From this image
stack, 6 secondary CT images per chestnut sample, including mean and
maximum intensity value images for each of the planes were extracted.
Thereafter, a total of 1194 grayscale intensity, and textural features were
extracted from the 6 secondary CT images per sample. Ultimately, 86, 155 and
126 features were found to be effective in designing a quadratic discriminant
analysis classifier with an overall performance accuracy of 85.9 %, 91.2 % and
96.1 % for 5, 3 and 2 classes, respectively.
This study provides a powerful tool to accurately visualize and sort
chestnuts based on their internal quality, leading to the improved marketability of
attractive, safe, high quality chestnuts. Results show that this method is accurate,
reliable, and objective and it is applicable to an automatic noninvasive in-line CT
sorting system.

Copyright by
IRWIN R. DONIS-GONZÁLEZ
2013

To those close and far away that I try to tell, every time I can, that I love. You
might never know how much your love, encouragement, moral support and
sacrifice has meant to me through the development of this dissertation. It has not
been easy, but this is an accomplishment from all of us.
With lots of love this is for you…

v

ACKNOWLEDGMENTS

I owe a debt of gratitude to many people for their support and contributions
to this dissertation.
First, I would like to thank my advisor, Professor Daniel E. Guyer, who has
been closely working on this project, and contributed excellent guidance and
moral support throughout. None of this could have been done without Professor
Guyer’s enthusiasm, willingness to help, and great knowledge on the topic.
Professors Dennis W. Fulbright, Anthony Pease and Renfu Lu, who have been
intrinsically involved in the project and have made innumerable useful
suggestions, experimental design, development and editorial comments. All have
helped me to be a better scientist and a better person.
I would also like to thank Professor Ajit K. Srivastava, Dr. Dirk J.L. Colbry,
Dr. Fernando Mendoza, Dr. Diwan P. Ariana, Dr. Akira Mizushima, Dr. Sanghyup
Jeong, and Professor Randolph Beaudry; highly known scientist from diverse
departments at Michigan Sate University, for their technical and knowledge
support, during the experimental design and development of diverse studies.
I am particularly grateful to all who provided financial support during my
career and all stages of the study at Michigan State, including: Professor Dennis
W. Fulbright through the Ernie and Mabel Rogers Endowment, the Department of
Biosystems and Agricultural Engineering, the Michigan Chestnut Grower
Incorporation, the Galen and Ann Brown scholarship, the Merle and Catherine
Esmay Scolarship, Michigan State University and the Generating Research and

vi

Extension to meet Economic and Environmental Needs (GREEEN) grant
program.
Finally, I would like to express my deepest appreciation to everybody who
participated in the study and also offered countless hours of moral support,
including Sara Stadt, Mario Mandujano, James Burns, Gabriel A. LeivaValenzuela, Frank Barthel and fellow students. Also, Rex Miller, Mark Seller, and
Meg Willis-Redfern for technical support using the CT scanner, and the Michigan
State Veterinary Teaching Hospital for providing the CT scanner used for the
experiments. Without their assistance and help provided, I could obviously never
have performed this study.
Perhaps the largest share of my appreciation goes to my family and
friends. Without their loving faith, this study would not have been possible, nor
worth doing.

vii

TABLE OF CONTENTS

LIST OF TABLES …………………………………………………………................. xi
LIST OF FIGURES .............................................................................................. xii
KEY TO ABBREVIATIONS .............................................................................. xviii
CHAPTER 1. INTRODUCTION AND LITERATURE REVIEW ............................. 1
1.1 Introduction ..................................................................................................... 1
1.2 Chestnut fruit .................................................................................................. 6
1.3 World chestnut industry .................................................................................. 9
1.4 Electronic sorting technologies ..................................................................... 13
1.5 Basic principles of X-ray computed tomography (CT) .................................. 15
1.6 Application of CT in non-medical industries .................................................. 26
1.7 Objectives and Hypothesis ........................................................................... 27
CHAPTER 2. APPLICATION OF RESPONSE SURFACE METHODOLOGY
(RSM) TO SYSTEMATICALLY OPTIMIZE FRESH CHESTNUT COMPUTED
TOMOGRAPHY (CT) IMAGE QUALITY ............................................................. 29
2.1 Abstract ........................................................................................................ 29
2.2 Introduction ................................................................................................... 30
2.3 Materials and methods ................................................................................. 31
2.3.1 Chestnut sample collection, reference cylinders, and Quality Assurance
(QA) phantom CT imaging scans ....................................................................... 31
2.3.2 RSM Box-Behnken design for CT image quality optimization.................... 34
2.3.3 Visual based fresh chestnut quality evaluation and SNR (Y1) ................... 39
2.3.4 Volume accuracy (Y2) using Teflon® reference cylinders ......................... 42
2.3.5 HCSR (Y3) and LCD (Y4) calculation using a standard QA Phantom........ 42
2.3.6 Optimized 2D CT image quality validation from an independent data set . 43
2.4 Results.......................................................................................................... 46
2.4.1 RSM Box-Behnken design for CT image quality optimization.................... 46
2.4.2 SNR (Y1) optimization ............................................................................... 52
2.4.3 Volume accuracy (Y2) optimization............................................................ 56
2.4.4 HCSR (Y3) and LCD (Y4) optimization ...................................................... 60
2.4.5 Optimized CT image quality attributes and validation ................................ 67
2.5 Discussion .................................................................................................... 73
2.6 Conclusions .................................................................................................. 75
CHAPTER 3. RELATION OF COMPUTED TOMOGRAPHY (CT) HOUNSFIELD
UNIT MEASUREMENTS AND INTERNAL COMPONENTS OF FRESH
CHESTNUTS...................................................................................................... 77
viii

3.1 Abstract ........................................................................................................ 77
3.2 Introduction ................................................................................................... 77
3.3 Materials and methods ................................................................................. 78
3.3.1 Sample collection and preparation ............................................................ 78
3.3.2 In vivo CT imaging scans........................................................................... 80
3.3.3 Visual based fresh chestnut quality and internal component assessment . 80
3.3.4 HU-value inference using training data set ................................................ 83
3.3.5 Chestnut categories prediction using an independent testing data set ...... 87
3.4 Results.......................................................................................................... 88
3.4.1 HU-value and category threshold inference using training data set ........... 88
3.4.2 Chestnut categories prediction using an independent testing data set ...... 92
3.5 Discussion .................................................................................................... 97
3.6 Conclusion .................................................................................................... 98
CHAPTER 4. POSTHARVEST NONINVASIVE ASSESSMENT OF FRESH
CHESTNUT INTERNAL DECAY USING COMPUTED TOMOGRAPHY (CT)
IMAGES ............................................................................................................ 99
4.1 Abstract ........................................................................................................ 99
4.2 Introduction ................................................................................................. 100
4.3 Materials and methods ............................................................................... 104
4.3.1 Sample collection and preparation .......................................................... 104
4.3.2 CT image preprocessing:......................................................................... 105
4.3.3 CT image re-slicing .................................................................................. 105
4.3.4 Individual chestnut CT image cropping .................................................... 106
4.3.5 Contrast enhancement ............................................................................ 111
4.3.6 CT image segmentation (Binary mask): .................................................. 112
4.3.7 Visual based fresh chestnut quality and internal component assessment:
.................................................................................................................. 113
4.3.8 Feature extraction: ................................................................................... 113
Basic intensity features: ................................................................................ 114
Haralick textural (Tx) features:...................................................................... 115
Intensity local binary pattern (LBP) textural features: ................................... 118
Intensity Gabor textural features:.................................................................. 120
Contrast features: ......................................................................................... 122
4.3.9 Feature selection: .................................................................................... 124
4.3.10 Classification (training and validation): .................................................. 128
4.4 Results........................................................................................................ 138
4.5 Discussion .................................................................................................. 142
4.6 Conclusions ................................................................................................ 147
CHAPTER 5. DISCUSSION AND FINAL REMARKS ....................................... 149
APPENDICES .................................................................................................. 160
APPENDIX A. Estimation of SNR, Volume accuracy, High contrast spatial
resolution (HCSR), Low Contrast Detectability (LCD), and digital quality
assessment (DQA) ........................................................................................... 161
APPENDIX B. CT imaging in other fresh agricultural commodities .................. 168
ix

REFERENCES ................................................................................................. 176

x

LIST OF TABLES

Table 1. Scanning parameters for the CT – General Electric, BrightSpeed™ RT
16 Elite (GE Healthcare, Buckinghamshire, England, Great Britain) .................. 34
Table 2. Box-Behnken experimental design variables ....................................... 36
Table 3. Design matrix (un-coded factors) in respect to each response variables
........................................................................................................................... 38
Table 4. ANOVA table for Signal-to-noise Ratio (SNR), Volume Accuracy (VA),
High Contrast Spatial Resolution (HCSR), and Low Contrasts Detectability (LCD)
non-linear response models ............................................................................... 48
Table 5. Stationary maximized points in original units for each non-linear
response variable model .................................................................................... 49
Table 6. Regression coefficients and P valuesa for each non-linear response
variable model .................................................................................................... 51
Table 7. Acquired HU value samples from different chestnut components and
air, based on regions of interest (ROIs) .............................................................. 84
Table 8. Haralick textural (Tx) features (Haralick, 1979) .................................. 117
Table 9. Main selected features (seventy five) for the Quadratic discriminant
analysis (QDA) classifier using sequential forward selection (SFS) in combination
with the Fisher discriminant objective function (J(W)) for the 5-, 3- and 2-class
classifiers. ......................................................................................................... 139
Table 10. Classifier performance using selected features (m) with a 4-folds
validation .......................................................................................................... 141

xi

LIST OF FIGURES

Figure 1. World chestnut (Castanea spp.) production in 2011 ............................ 2
Figure 2. Chestnut fruit morphology. A. Chestnut fruit longitudinal cut, B. Spiny
burr attached to chestnut tree with chestnuts. For interpretation of the references
to color in this and all other figures, the reader is referred to the electronic version
of this dissertation................................................................................................. 6
Figure 3. X-ray computed tomography (CT) A. Traditional CT working principal,
B. Measuring arrangement of the GE (GE Healthcare, Buckinghamshire,
England, Great Britain) BrightSpeed™ RT 16 Elite CT used in these studies, C.
Schematic representation of a CT system scanning (5 chestnuts per row)
containing several 2D XY-plane CT images (slices). .......................................... 18
Figure 4. CT imaging and data acquisition A. Transmission measurements
through the objects at numerous angles (black arrows) around the object, B.
Back-projected (reverse) transmission measurement onto 2D XY-plane CT
image matrix, C. Back-projection reconstruction from projection values (P), for a
simple image of four voxels. ............................................................................... 19
Figure 5. (a) 3D virtual cross-sections representation of a chestnut at three
planes (different angular orientations) along the horizontal (X), vertical (Y), and
longitudinal (Z) axes (Figure not to scale). (b) Original series of acquired CT XYplane-slices. (c) Reconditioned (re-sliced) CT YZ-plane-slices. (d) Re-sliced CT
XZ-plane-slices. .................................................................................................. 24
Figure 6. (a) Fresh chestnuts samples, and Teflon® reference cylinders, (b)
Testing phantom, (c) 16-bit CT images of fresh chestnut samples for Signal-tonoise Ratio (SNR) calculation, and Teflon® reference cylinders for volume
accuracy (mm-3) estimate. (d) 8-bit CT image of QA phantom used to infer High
Contrast Spatial Resolution (HCSR). (e) 8-bit CT image of QA phantom used to
infer Low Contrast Detectability (LCD). .............................................................. 32
Figure 7. (a) Example of chestnut 16-bit CT image slice used for Signal-to-noise
Ratio (SNR) calculation, from a region containing fresh uniform chestnut tissue,
using reference color image. (b) Example of Teflon® reference cylinder 8-bit CT

image used to estimate volume accuracy (mm-3) with binary image after
segmentation (simple global threshold of 135). (c) Example of cropped 8-bit CT
image used to calculate High Contrast Spatial Resolution (HCSR) with binary
image after segmentation (simple global threshold of 134). Each pattern consists
of five bars and spaces called line pairs (lp). The sizes of the patterns are

xii

equivalent to 1.6 mm, 1.3 mm, 1.0 mm, 0.6 mm, and 0.5 mm, respectively. (d)
Example of cropped 8-bit CT image used to determine Low Contrast Detectability
(LCD) with binary image after segmentation (simple global threshold of 94).
Image displays various sized Holes (H). The diameter of each H equal 10.0 mm,
7.5 mm, 5.0 mm, 3.0 mm, and 1.0 mm. .............................................................. 41
Figure 8. Example of images used for Digital Quality Assessment (DQA). (a)
Reference color image. (b) 16-bit optimized CT gray scale image. (c) Binary
image of whole chestnut after segmentation, applying a simple global threshold
of 400 HU to Fig. 8(b). (d) Binary image of transition points and pellicle after
employing a Sobel filtering method to detect edges in Fig. 8(b). (e) Binary image
of healthy tissue after segmentation, using a simple global threshold of 1050 HU
to Fig. 8(b). Figure is partially presented in color. ............................................... 45
Figure 9. Surface plots of SNR versus (a) slice thickness (ST - mm) for current
(C - mA), (b) slice thickness (ST - mm) for voltage (V - kV), and (c) current (C mA) for voltage (V - kV). ..................................................................................... 54
Figure 10. (a) Surface plot of volume accuracy (VA - mm-3) versus slice
thickness ( ST - mm) for current (C - mA). (b) Interaction plot for volume accuracy
(VA - mm-3) versus slice thickness (ST - mm) for current (C - mA). Data points
followed by the same lower case letter within the same slice thickness are not
significantly different at P = 0.05 (ANOVA with post-hoc Tukey multiple
comparison of means). Surface plots of volume accuracy (VA - mm-3) versus (c)
slice thickness (ST - mm) for voltage (V - kV), and (d) current (C - mA) for voltage
(V - kV). .............................................................................................................. 58
Figure 11. Surface plots of HCSR versus (a) slice thickness (ST - mm) for
current (C - mA), (b) slice thickness (ST - mm) for voltage (V - kV), and (c)
current (C - mA) for voltage (V - kV). .................................................................. 62
Figure 12. Surface plots of LCD versus (a) slice thickness (ST - mm) for current
(C - mA), (b) slice thickness (ST - mm) for voltage (V - kV), and (c) current (C mA) for voltage (V - kV). ..................................................................................... 64
Figure 13. Example of Gray-scale CT image quality for low (-1), medium (0), and
high (+1) factor combinations. Note that low (-1) and high (+1) level combinations
are not part of the Box-Behnken design, but CT images were acquired for
visualization and validation purposes. ................................................................ 66
Figure 14. Box-plots showing experimental results for CT image quality attributes
(SNR, VA (mm-3), HCSR, LCD) from a completely independent validation data
set, using optimized scanning parameters, as seen in Table 5. The median of
each experimental quality attribute is represented as a thick horizontal black line,
upper and lower quartiles as a box with the maximum and minimum

xiii

measurements as lines protruding from these. The mean experimental response
for each quality attributes is symbolized as a black solid dot (.). Predicted
optimized response using non-linear response surface polynomial models (Table

2) is symbolized as a hollow circle (). ............................................................... 69
Figure 15. Relationship between Subjective Quality Rating (SQR) values
obtained from a 5-experts panel and the Digital Quality Assessment (DQA) from
their corresponding CT image validation set, obtained through optimized
scanning parameters. Black solid line represents the least-square linear
regression line (n = 110). The two dashed lines indicate 95 % CI of the linear
regression. The two dotted lines indicate 95 % prediction intervals of the linear
regression. .......................................................................................................... 72
Figure 16. Cross-sectional XY-plane 2D CT images and color raw image slices
(RGB) of chestnuts. (A) shows a healthy chestnut (SL1). (B) shows a partially
decayed chestnut (SL2). (C) shows a completely decayed chestnut (SL3). (D)
shows a partially decayed chestnut (SL2), illustrating the HU-values-profile taken
at the grey line (PL), typifying the HU variation within components in the same
chestnut. ............................................................................................................. 82
Figure 17. Box-plots showing the HU values from ROIs obtained using a training
data set. The median is represented as a thick horizontal black line, upper and
lower quartiles as a box with the maximum and minimum measurements as lines
protruding from these. Box-plots followed by the same letter and enclosed by the
same rectangle are not significantly different between each other at P = 0.05
(ANOVA) (Tukey multiple comparison of means). Solid circles joined by a dashed
black line (---) show the relationship between the mean of the HU values for each
ROIs and subsequently each category. .............................................................. 89
Figure 18. (a) Category prediction using estimated thresholds (horizontal black
lines) with an independent testing data set. HU threshold values are indicated
above each threshold line. The selected HU value threshold range for air, void
spaces, decay tissue, pellicle tissue, and healthy tissue are -1024 to -975, -974
to -500, -400 to -210, -209 to 24, and 25 to 300; respectively. (b) Confusion
matrix corresponding to the category prediction accuracy (overall accuracy rate =
90.6%). (c) Representative images, containing automatically generated color
labeled images from their corresponding CT images, based on HU-value
threshold intervals. RGB raw image slices are also included as a quality
reference. ........................................................................................................... 95
Figure 19. Procedure used to generate the pattern classification algorithm to
categorize chestnut quality using CT images. .................................................. 103
Figure 20. Mean intensity values of pixels in all of the: (a) YZ-plane-slice of the
whole board containing chestnuts, (b) YZ-plane-slices for the first chestnut row,
(c) XY-plane-slices for the first chestnut row, (d) XZ-plane-slices for the first
xiv

chestnut row. Mean intensity values of pixels in final secondary CT cropped
chestnut for the (e) YZ-, (f) XY- and (g) XZ-plane-slices. Maximum intensity
values of pixels in final secondary CT cropped chestnut for the (h) YZ-, (i) XYand (j) XZ-plane-slices. From (a) - (d) the beginning of the first chestnut row (Z1), the end of the row (Z2-), the left side (X1-), the right side (X2-), the bottom (Y1) and the top (Y2-) of the first chestnut -spatial-location-values are shown in red.
......................................................................................................................... 108
Figure 21. (a) Partially decayed fresh raw chestnut slices. (b) Original secondary
mean CT image (YZ-plane-slice), (c) adjusted secondary mean CT image (YZplane-slice), (d) final contrast enhanced secondary mean CT image (YZ-planeslice). (e) Segmented CT image from final contrast enhanced secondary mean
CT image (YZ-plane-slice) (binary mask). (f) Original secondary maximum CT
image (YZ-plane-slice), (g) adjusted maximum secondary maximum CT image
(YZ-plane-slice), and (h) final contrast enhanced secondary maximum CT image
(YZ-plane-slice). ............................................................................................... 110
Figure 22. (a) Healthy and (b) partially decayed chestnuts with their
corresponding secondary maximum CT image (XY-plane-slice) (55 pixels x 65
pixels). (1) Local Binary Patter (LBP) transformations using different pixel
comparison (d,h) was applied to the secondary maximum CT image. (2)
Secondary maximum CT image Gabor filtered transformed images (Irs(i,j)) at
different scales (r) and orientation (s) for an θ = 45 are included. (3) Example of
Haralick textural (Tx) features, contrast and intensity features obtained from
included secondary maximum CT image. For visual reference, three-color fresh
raw image slices of the evaluated chestnut are included. ................................. 126
Figure 23. (a) Sample distribution used to train and validate the 5-class
classifier. Figure also contains an example for each of the 5 categorical classes,
representing the quality index levels. (b) Quadratic discriminant analysis (QDA)
classifier performance using validation set with 4-folds, in relation to the number
of selected features (m). Black line represents the classification mean
performance, dotted black line (---) represent 95 % confidence intervals for the
validation pool. (c) Validation confusion matrix corresponding to the chestnut
quality class prediction using 25 % of samples with 86-m (overall accuracy rate
equal to 85.9 %). .............................................................................................. 130
Figure 24. (a) Sample distribution used to train and validate the 3-class
classifier. Figure also contains an example for each of the 3 categorical classes,
representing the quality index levels. (b) Quadratic discriminant analysis (QDA)
classifier performance using validation set with 4-folds, in relation to the number
of m. Black line represents the classification mean performance, dotted black line
(---) represent 95 % confidence intervals for the validation pool. (c) Validation
confusion matrix corresponding to the chestnut quality class prediction using 25
% of samples with 155-m (overall accuracy rate equal to 91.2 %). .................. 132

xv

Figure 25. (a) Sample distribution used to train and validate the 2-class
classifier. Figure also contains an example for each of the 2 categorical classes,
representing the quality index levels. (b) Quadratic discriminant analysis (QDA)
classifier performance using validation set with 4-folds, in relation to the number
of m. Black line represents the classification mean performance, dotted black line
(---) represent 95 % confidence intervals for the validation pool. (c) Validation
confusion matrix corresponding to the chestnut quality class prediction using 25
% of samples with 126-m (overall accuracy rate equal to 96.1 %). Figure is
partially presented in color. ............................................................................... 134
Figure 26. (a) Ultrafast Rossendorf electron beam X-ray tomograph (ROFEX)
scanner working principle. (b) Experimental setup of the ROFEX-scanner. ..... 152
Figure 27. Visual preliminary results of CT images (lower row) obtained using
the ROFEX-scanner with its corresponding color raw image slices (upper row).
(a) Healthy, (b) partially decayed (rotten) and (c) completely decayed chestnuts.
(d) 3D reconstruction of two chestnuts, showing a rotten section in one of the
chestnuts (white arrow). ................................................................................... 153
Figure 28. (a) Commercially available in-line traditional X-ray sorter, and (b) its
corresponding 2D X-ray images. ...................................................................... 157
Figure 29. Typical unit operations in mechanized packinghouse (Thompson et
al., 2002). Vision of future integration of in-line Computed Tomography (CT) for
sorting fresh agricultural products*. *: Flowchart in red are the proposed sections
that will be potentially improved. ....................................................................... 159
Figure 30. Color raw images of (a) chestnuts, (b) pineapples, (c) tart cherries
and (d) pickling cucumbers showing the regions of interest (ROIs) for each fresh
commodity. ....................................................................................................... 169
Figure 31. (a) Color raw image slices, cross-sectional 2D CT images acquired
using the GE BrightSpeed™ RT 16 Elite CT scanner, and 3D reconstruction of
pineapples (b) Black dots showing the HU-values for the mean of 100 data points
per each pineapple ROI (n = 100). Values followed by the same letter are not
significantly different between each other at P = 0.05 (ANOVA) (Tukey multiple
comparison of means). Vertical bars represent the standard deviation (SD) of
each ROI. ......................................................................................................... 170
Figure 32. (a) Color raw image slices, cross-sectional 2D CT images acquired
using the GE BrightSpeed™ RT 16 Elite CT scanner, and 3D reconstruction of
tart cherries (b) Black dots showing the HU-values for the mean of 100 data
points per each tart cherry ROI (n = 100). Values followed by the same letter are
not significantly different between each other at P = 0.05 (ANOVA) (Tukey
multiple comparison of means). Vertical bars represent the SD of each ROI. .. 172

xvi

Figure 33. (a) Color raw image slices, cross-sectional 2D CT images acquired
using the GE BrightSpeed™ RT 16 Elite CT scanner, and 3D reconstruction of
pickling cucumbers (b) Black dots showing the HU-values for the mean of 100
data points per each pickling cucumber ROI (n = 100). Values followed by the
same letter are not significantly different between each other at P = 0.05
(ANOVA) (Tukey multiple comparison of means). Vertical bars represent the SD
of each ROI. ..................................................................................................... 174

xvii

KEY TO ABBREVIATIONS

2D – Two-dimensional.
3D – Three-dimensional.
ANOVA – Analysis of variance.
CGI – Michigan Chestnut Grower Inc.
CI – Confidence intervals.
CS – Chestnut species. Co - ‘Colossal’ (Castanea sativa x C. crenata) and Ch Chinese seedlings (C. mollissima).
CT – Computed Tomography.
dH2O – Distilled water.
DQA – Digital Quality Assurance.
df – Degrees of freedom.
FAO – United Nations Food and Agriculture Organization.
Gy - Grays.
HCSR – High Contrast Spatial Resolution.
HU – Hounsfield Units.
IAEA – International Atomic Energy Agency.
kV – Kilovolt.
L and U – Lower and Upper category thresholds.
LCD – Low Contrast Detectability.
mA – Milliamperage.
MRI – Magnetic Resonance Imaging.

xviii

MS – Mean square.
NIR – Near infrared.
PB – Peg-board containing chestnut samples attached to it. Boards are used to
keep samples in place while scanning. PB1 contained Chinese seedlings and
PB2 ‘Colossal” chestnuts.
PDA – Potato-dextrose-agar.
PL – Profile-line (HU values).
Px – Pixel.
QA – Quality Assurance.
RGB – Sub-index representing images that are originally printed in color. RGB is
a universally accepted system for color representation in television, video sets,
and CRT displays; using red, green and blue primaries. It is not a perceptually
uniform color space (i.e. differences between colors in the 3D RGB space do not
correspond to those perceived by humans).
ROI, ROIs – Region of interest, Regions of interest.
SCF – Scientific Committee on Food of the European Commission.
SD – Standard deviation.
SL – Internal disorder severity levels. Where, SL1 contains chestnuts with no
internal disorders – Healthy, SL2 chestnuts which are partially disordered and
SL3 represents the group of chestnuts, that are completely disordered.
SS – Sum of squares.
SU – Sampling unit. Where, SU1 = 1-pixel per repetition, and SU2 = Mean HU
value from a 4 mm2 (6.6 pixel2) square region per repetition.

xix

RSM – Response Surface Methodology.
SD – Standard Deviation.
SNR – Signal-to-noise Ratio.
SQR – Subjective Quality Rating.
VA – Volume Accuracy, expressed in mm-3.
WHO – World Health Organization.

xx

CHAPTER 1. INTRODUCTION AND LITERATURE REVIEW
1.1 Introduction
Chestnut (Castanea spp.) is one of the most popular nut-bearing trees in
the countries of Asia, including China, Republic of Korea, and Japan; as well as
the Mediterranean region (Ridley, 1999). Worldwide, from 1998 to 2011, chestnut
production increased by approximately 92,000 kg per year (Food and Agriculture
Organization Statistics Division (FAOSTAT), 2013). As observed in Fig. 1, in
2011, chestnuts grown in Asia and Europe accounted for more than 95% of the
worldwide production, with chestnut production increasing in Australia, New
Zealand, Chile and Bolivia among others (Grau and France, 1999; Klinac et al.,
1999; Ridley, 1999; Fulbright and Mandujano, 2000; Mencarelli, 2001).
In the United States, chestnuts are uncommon; people are more likely to
be familiar with the unrelated and poisonous horse chestnut (Aesculus spp.)
(Fulbright and Mandujano, 2000), than with any of the edible, sweet chestnut
species, including C. dentata, C. mollissima, C. crenata, C. sativa, C. seguinii, C.
pumila, and C. henryi (Anagnostakis et al., 1998; Fulbright and Mandujano,
2000; Miller, 2003). This is partially due to chestnut blight (Cryphonectria
parasitica, Murril and Barr), a fungus disease that virtually eliminated the oncewidespread American chestnut (C. dentata) during the first half of the twentieth
century (Merkel, 1905; Metcalf and Collins, 1909; Gravatt and Marshall, 1926).
Over the past twelve years, with the availability of blight resistant hybrids, as well
as tolerant European and Chinese cultivars, commercial orchards, and wild forest

1

tree populations have significantly increased in the United States, including
Michigan (Fulbright and Mandujano, 2000; Jacobs, 2007).

Source: Food and Agriculture Organization Statistics Division (FAOSTAT), 2013
Figure 1. World chestnut (Castanea spp.) production in 2011
Before the early 19th-century, especially in the European Mediterranean
region and in Asia, chestnut consumption was predominant in rural areas, and
was considered to be a food for the poor. Currently, chestnuts are mainly
consumed by roasting the fresh product (Harte et al., 2003), but also are used as
a cooking ingredient in diverse culinary applications, including pastries, and
specialty dishes (Kelley and Behe, 2002; De la Montaña Miguelez et al., 2004;
Blackwell, 2006; Borges et al., 2008). Additional to the fresh product, small

2

portions of chestnuts are consumed peeled. Mechanically peeled and
subsequently frozen chestnuts are usually sold as a value-added product to
extend the chestnut market beyond seasonal sales, providing the opportunity to
expand markets and their utilization (Guyer et al., 2003; Gao et al., 2008).
Positively, for the United States, Vossen (2000) claimed that the chestnut
industry has a great potential, and a small rise in domestic consumption could
generate a net revenue of up to 800 million US$ annually. Studies performed in
Wisconsin and California, between the years 2000 and 2004, evaluating
consumer preference for chestnuts, reflected that the majority of the interviewed
had never tasted a chestnut, but were interested in exploring them as a new food.
Overall quality and nutritional value were listed as the most important attributes
influencing purchase and consumption decisions (Gold et al., 2004).
Chestnut quality is measured not only by external factors such as color,
shape, size, surface blemish, and surface mold growth, but also by internal
disorders and freshness, which are very important for consumer acceptance. It is
important to note that the external appearance usually is not altered, at least
initially by internal disorders, making disordered chestnuts difficult to detect
without destructive evaluation. Internal disorders usually are the result of
anatomical and physiological changes within the tissue, such as moisture loss,
chemical conversion, discoloration, senescence, microorganism attack, cell
breakdown (physiological decay), and insect injury (Upchurch et al., 1993).
In fresh chestnuts, microorganism attack is the most important internal
disorders determining internal (kernel) quality. Penicillium sp., Aspergillus sp.,
Fusarium sp., Phomopsis castanea, Acrospeira mirabilis, and Sclerotinia
3

pseudotuberosa (syn. Ciboria batschiana, S. batschiana; anamorphic from
Rhacodiella castanea, syn. Myrioconium castanea) have been persistently
identified worldwide, including in the United States, as microorganisms that play
an important role as casual agents in postharvest microbiological decay (Ridé
and Gudin, 1960; Paglietta and Bonous, 1979; Ellis and Ellis, 1985; Montealegre
and Gonzalez, 1986; Vettraino et al., 2005; Jerimini et al., 2006; Sieber et al.,
2007; Donis-González, 2008; Donis-Gonzalez et al., 2010). Other issues, like
chestnut internal kernel breakdown (IKB) also threatens the industry, especially in
Michigan. IKB is defined as physiological kernel decay, near the end of the
harvesting season (September), where no fungal pathogens are linked with the
condition. Symptomatic IKB chestnuts develop properly and no apparent issue
can be observed on the chestnut shell. It has been hypothesized that IKB might
be related to pollination incompatibility (Long, 2012). In some cases, these
problems can lead to a completely unmarketable product. A certain amount of
affected product may be sold in less demanding markets, nonetheless this
practice results in important economic and potential market losses due to
consumer rejection.
In addition to quality concerns and product appearance, product safety is
also a significant issue. Some fungi present in chestnuts are capable of secreting
substances that are potent, acute toxins, or carcinogens to both animals and
humans. These toxic agents are called mycotoxins, and their impact on domestic
animals in terms of decreased growth rate, abnormal reproduction, and early
death has long been recognized. The most important mycotoxins are aflatoxins,
deoxynivalenol (DON), zearalenone, fumonisin B1, T-2 toxin, and ochratoxin A
4

(Adams and Moss, 2000; Bullerman and Bianchini, 2007). These substances
have been predominately isolated from grains such as wheat and corn, but can
also been found in chestnuts (Overy et al., 2003; Donis-González et al., 2010a)
as well as in other products including rice (Tanaka et al., 2007), grapes (Varga et
al., 2007), beer, wine (Mateo et al., 2007), vegetable oil (Schollenberger et al.,
2008), peanuts, pecans (Lillard et al., 1970), and processed foods (Adams and
Moss, 2000; Bullerman and Bianchini, 2007).
Currently, random and destructive sampling techniques can be employed
to evaluate internal chestnut quality. However, in commercial situations,
physiologically disordered, microbiologically decayed, and empty or damaged
chestnuts, are eliminated by their proclivity to float in tap water, as healthy
chestnuts tend to sink. Nevertheless, performance varies significantly between
cultivars and throughput, proving that this floating method is unreliable for sorting
purposes. Errors associated with the method might be caused by the large
density variations among individual chestnuts and the presence of a perfectly
healthy chestnut containing void spaces, thick shell, or pellicle embedded in the
healthy kernel. Additionally, the technique is slow, cumbersome, and a significant
amount of healthy product is also being discarded. (Donis-González, 2008).
All these issues regarding postharvest mold, decay, safety concerns, and
inability to sort chestnuts appropriately, reflect the need to develop a reliable
technique for chestnut internal disorder detection, which is fast, practical and
appropriate to apply for in-line sorting systems.

5

1.2 Chestnut fruit
Many nuts, as well as chestnuts, are wrapped in a papery or spiny husk
called the involucre, more commonly known as a bur (Fig. 2). This structure is
usually mistaken as the fruit, but it is actually a whorl of modified leaves around
the flower or flowers. Inside the involucre, depending on the species or cultivar,
one

to

three

chestnuts

can

be

found.

In

chestnuts,

the

ovule

Figure 2. Chestnut fruit morphology. A. Chestnut fruit longitudinal cut, B.
Spiny burr attached to chestnut tree with chestnuts. For interpretation of
the references to color in this and all other figures, the reader is referred to
the electronic version of this dissertation
or ovules (inside ovary) that develop into a seed or several seeds are called
kernels, structures, which will swell and give rise to what is known as the shell
(fruit called an achene or pericarp). The shell has an external brown wall with a

6

woody and shiny appearance. Two large cotyledons, forming the kernel, surround
the embryo, which contains the radicle, hypocotyls and epicotyls. Between the
shell and the kernel, is a thin brown papery-like structure, commonly called a
pellicle, which is botanically known as a seed coat, testa or episperm (Mencarelli,
2001; Miller, 2003). Usually there is only one kernel per shell, but in certain cases
there may be two or more, commonly referred to as a multiple embryo chestnut.
In different regions of Europe, including Italy and Spain, the term “marron”, or
"marrone" denotes a single kernel nut, while the term “chatâigne”, “castaño” or
“castaña” denotes double or even more kernels per shell. Therefore, from a
botanical point of view, a chestnut is an achene type fruit, containing one or more
seeds, known as a kernel with creamy, yellow-colored edible cotyledons, covered
by a membrane called pellicle or episperm (Mencarelli, 2001; Miller, 2003).
When chestnuts begin to ripen in late summer or autumn, the bur changes
color from light green to yellow-brown and releases the chestnuts. Sometimes the
bur opens on the tree, releasing the chestnuts, but the bur can also drop and
open on the ground (Anagnostakis et al., 1998; Mandujano et al., 1998; Miller,
2003; Willis et al., 2007).
Unlike edible nuts such as almonds, hazelnuts, and walnuts, the chestnut
kernel is slightly hard, rich in carbohydrates, mainly starch, but also sucrose,
glucose and fructose (40 – 90 %), containing high fiber (14 - 19 %), protein (6 10 %) and low lipids (0.4 - 10 %), which are 90 % unsaturated fatty acids
(nutrients are expressed as the percentage of dry weight). After harvest,
chestnuts water content is relatively high (> 50 %) (Anagnostakis and Devin,
1999). Furthermore, chestnuts are a source of vitamin A, calcium, iron, fiber with
7

and antioxidants (Biomhoff et al., 2006; Gao et al., 2008). Biomhoff et al. (2006)
as well as Anagnostakis and Devin (1999) reported that all these characteristics
are highly beneficial for human health and are needed for proper nutrition and
protection of animal cells, ranking chestnuts as a healthy food.
After harvest, respiration of most seeds is usually characterized by a low
and constant rate without any peak, reflecting a non-climacteric respiration
pattern. The same relative pattern can be found in chestnuts, but when compared
to other seeds and to most other nuts, chestnuts have a higher respiration rate.
Subsequently, water loss and starch conversion to sugars is high (Kader, 2002;
Willis et al., 2007). Based on several respiration rate studies (Harte et al., 2003),
chestnuts were comparable to blueberries (Perkins-Veasie, 2004) and iceberg
lettuce (Smyth and Cameron, 1998; Kader, 2002), exhibiting a lower respiration
rate than cut broccoli (Talasila et al., 1994), peas, asparagus, sweet corn, and
mushrooms (Kader, 2002).
In addition to physiological, and microbiological decay problems arriving
from the field, the most common changes that occur in chestnuts after harvest
are: moisture loss, starch conversion to sugar, fungal decay, cell breakdown, and
insect damage (Wells, 1980). Only starch conversion to sugar, commonly known
as "curing", positively affects chestnut quality, enhancing the flavor, sweetness,
and acceptability of the product (Harte et al., 2003). The other three major factors
have a negative affect on quality and storage potential of the final product
(Paglietta and Bonous, 1979; Wells, 1980; Montealegre and Gonzalez, 1986;
Harte et al., 2003; Jerimini et al., 2006). Because quality can only be maintained,
sometimes monitored or detected, and not improved after harvest (Kader, 2002;
8

Willis et al., 2007; Kincaid et al., 2008), efforts have focused on increasing the
storability of chestnuts by reducing microbes, insects, and maintaining the quality
at harvest. Vossen (2000), reported that during three months storage,
temperatures between -2 C to 0.5 C were recommended, playing an important
role in inhibiting fungal decay, increasing storage, and reducing moisture loss. In
Michigan, in addition to the recommended storage temperature (Vossen, 2000), a
postharvest treatment has recently been incorporated and has proven to maintain
chestnut quality during storage. Treatment consists in dipping fresh chestnuts for
5 min in water containing 2700 ppm hydrogen dioxide + 200 ppm peracetic acid
(Storox™, BioSafe Systems, Glastonbury, CT, USA) (Donis-González, 2008;
Donis-González et al., 2010b). Nonetheless, as mentioned before, because there
is no reliable commercially available sorting method, chestnuts that are already
decayed at time of harvest, are still being stored among healthy product. The
storing of decayed chestnuts inevitably increases costs for the chestnut industry.
Furthermore, decayed, bad quality commodity additionally contaminates healthy
chestnuts, causing them to decay during storage. Therefore chestnut sorting
might be considered pre-storage as well as post-storage, before the product is
sent to the client.
1.3 World chestnut industry
According to the Food and Agriculture Organization Statistics Division
(FAOSTAT), in 2013 at least 25 countries produced chestnuts. Reports from
1998 to 2011, showed that Chestnut production and harvested area have
experienced a continuous average growth. Chestnut production worldwide in

9

2011 was estimated to be over 2,015,000 metric tons (2.01 x 109 kg) distributed
as following: China,  84 %; Republic of Korea,  3 %; Italy, Turkey, and Bolivia,
 2-3 % each; France, Portugal, Spain, and Greece,  0.2-1 % each; and the
United States, Australia, New Zealand, Chile, Peru, Brazil, Albania, among
others, less than 0.1 % each (Fig. 1).
China is the largest low-cost producer and exporter of chestnuts with an
estimated production of 1,700,000 metric tons (1.7 x 109 kg). Approximately one
third of their chestnut production is exported to Japan. Locally, most chestnuts
are consumed fresh or roasted with an unspecified amount used in Chinese
cuisine to develop a broad variety of dishes (Vossen, 2000).
Republic of Korea, the second largest producer, yields significantly less
chestnuts compared to China per year (55,780 metric tons = 55,780,000 kg),
approximately half of which are exported to Japan and another countries like the
United States ( 1 to 3 %) (Vossen, 2000).
Japan is the largest chestnut importer and is among the largest consumer
group, even though it is not the biggest producer. In the Republic of Korea and
Japan, local or imported chestnuts are primarily stored under refrigerated
conditions of 4 – 7 C, and consumed fresh, boiled, or as an ingredient in diverse
dishes. Some are also stored dry or peeled for further use.
Within Europe, Italy is the largest chestnut producer with approximately
57,493 metric tons (57,493,000 kg) and leads the world in production of delicacyprocessed chestnut-based products such as marrone glacé (preserved chestnuts
in sugared liquor). In Europe, the use of dry chestnuts and chestnut flour in
10

cooking has recently declined, but the popularity of these products is increasing
elsewhere, especially in the United States. These value-added products, such as
processed, dried, peeled, and frozen chestnut have reached more than 8.00 US$
per pound, prompting moves to expand the chestnut industry worldwide (Vossen,
2000). Europe’s second largest producer is Portugal, with up to 21,990 metric
tons (21,990,000 kg) per year. France is one of the biggest importers of
chestnuts in Europe, mostly from Italy, Spain, and Portugal.
Recently, the United States, Australia, New Zealand, Chile, and other
countries in the Southern Hemisphere have begun to produce chestnuts and
have established economically sustainable industries, mainly for export
(Mandujano et al., 1998; Fulbright and Mandujano, 2000; Vossen, 2000). The
United States has at least 5,000 acres of chestnut tree plantations. Of these,
approximately 1,500 acres are less than 10-year-olds. Chestnut trees are mainly
in California, Michigan, Oregon, Washington, Iowa, Idaho, Nebraska and Ohio
(Vossen, 2000). These commercial plantations are primarily from the cultivar
‘Colossal’ (European-Japanese hybrid = C. sativa x C. crenata). This cultivar has
been broadly used because it produces large nuts (up to 30 g), high yields (> 55
kg per tree), and is commercially sold at nurseries (Fulbright and Mandujano,
2000; Miller, 2003). Other cultivars such as ‘Dunstan Hybrid’ (a patented seedling
of a third generation cross between American and Chinese chestnut hybrids = C.
dentata x C. mollissima), ‘Skookum’, ‘Layeroka’, ‘Myoka’, ‘Skioka’, and ‘Eaton’,
which have apparent Chinese characteristics, have also been planted (Miller,
2003).

11

In recent years chestnut average production has been increasing (Vossen,
2000). As an example, in 2003 approximately 1,500 kg of chestnuts were
harvested by the Chestnuts Grower Incorporation (CGI) in Michigan, while in
2007, a total of 22,000 kg were harvested (Blackwell, 2006). An important
chestnut state is California, where the oldest commercial plantations are found.
Vossen (2000) indicated that most of the early chestnut orchards in California
were established by immigrants during the Gold Rush and are mostly seeds from
European chestnuts.
Michigan’s chestnut harvest starts in mid-September and proceeds
through the first week of November, but may start and subsequently end slightly
later, in northern or colder locations. Product is currently primarily sold through
CGI, a producer owned and controlled marketing cooperative. Most chestnuts are
sold from Thanksgiving through Christmas, via sales to specialty ethnic markets,
retail stores, food processors, restaurants, holiday festivals, farmers’ markets and
individual consumers. Fresh and frozen peeled chestnuts make up 60 % and 30
%, of outlet sales respectively. The remaining product (10 %) is sold as flour,
breading, and new dehydrated products such as chips and slices, and puree
(Blackwell, 2006).
Since freshness and microbial quality are sales factors in chestnut,
domestic production has a definite advantage over imported chestnuts. But to
maintain this advantage, local quality monitoring, sorting, storage conditions, and
postharvest management strategies are all key strategies in preventing the
chestnuts from molding and decaying. Regardless of poor quality and lack of
freshness, mainly due to long-term storage and transport of imported chestnuts,
12

the United States annually imports between 5 and 15 million kg of fresh
European and Chinese chestnuts, at a retail price of approximately 40 million
US$ (4-8 US$ per pound).

This indicates that the United States chestnut

consumption is high and economically significant. Nonetheless, in order for the
United States to expand its markets, replace imported chestnuts, and fulfill all its
local needs today, more than 20,000,000 m2 of production area would be
required. Furthermore, if chestnuts are marketed efficiently and domestic
consumption increases by only 0.22 kg per capita, the United Sates would
require over 202,000,000 m2 of mature production area to meet the demand.
Following this growth, the industry would be worth more than 300 million US$ per
year, but may be worth more than 800 million US$ annually, if the increase in
consumption is even higher (Vossen, 2000).
1.4 Electronic sorting technologies
Presently, noninvasive techniques mainly using color computer vision
systems are employed to determine external quality attributes (peel color,
external defects and shape) in fresh vegetables, nuts and fruits (Brosnan and
Sun, 2004; Mery and Pedreschi, 2005; Blasco et al., 2007; Gomes and Leta,
2012; Moreda et al., 2012). In addition, techniques based on optical, magnetic
resonance imaging (MRI), near-infrared (NIR), vibration, sonic and ultrasonic,
have also been explored for non-destructive determination of internal quality
attributes of a variety of agricultural and food products (Milczarek et al., 2009;
Cubero et al., 2010; Lorente et al., 2011). Internal quality attributes which have

13

been explored include apple firmness and soluble solids content (Peng and Lu,
2008), peach firmness (Lu and Peng, 2006), tart cherries pit presence (Qin and
Lu, 2005), tomato mechanical damage (Milczarek et al., 2009) and pickle internal
defects (Ariana and Lu, 2010). However, because of the morphological
characteristics of some commodities, like chestnuts, which contain an external,
relatively thick (1.0-2.0 mm), bright and shiny shell (> 70 gloss units), it is not
possible to use traditional color or even NIR sorting methods available in the
industry to evaluate internal quality traits. Currently, only destructive techniques
can be reliably employed to evaluate internal components of a variety of fresh
agricultural commodities, including chestnuts (Donis-González et al., 2013).
Clearly invasive techniques can’t be applied to all produce and, thus there is a
need to develop an in vivo in-line nondestructive tool capable of better assessing
fresh agricultural commodity internal attributes. This will enable the fresh
commodity industry to offer better quality products, therefore increasing customer
satisfaction, increase market share profit growth and have less customer
complaints (Garvin, 1984). One commodity which could extremely benefit from
such technologies is the chestnut, due to its propensity toward internal decay, as
described in Donis-González (2010).
Though not commercially available to the agricultural field, computed
tomography (CT) is an important noninvasive diagnostic tool broadly used in the
medical field and material sciences (Bushberg et al., 2002). Despite extensive
research efforts and off-line application studies, a real-time in-line CT inspection
system for internal quality attributes of fresh produce is not commercially
available. However, because of X-ray capabilities of visualizing internal properties
14

of objects through thick matter, as well as recent advances in high-performance
computing systems, new detector technologies including modern graphical
processing unit (GPU) computing capabilities (Pratx and Xing, 2011) and highperformance X-ray tubes, non-medical applications are gaining attraction. New
detector technologies offer real-time imaging, equipment cost decreases,
extended

or

continuous

operation,

and

significant

reduction

in

image

reconstruction time (Butz et al., 2005; Hanke et al., 2008).
1.5 Basic principles of X-ray computed tomography (CT)
X-rays are short wave radiations (10 - 0.01 nm) with energy between 1.92
x 10-17 and 1.92 x 10-14 J, which can penetrate matter. The wavelengths are
shorter than those of visible light rays (390 - 700 nm) and ultraviolet rays (10 400 nm), and longer than gamma rays (< 0.01 nm). X-rays are generated by
bombarding electrons on a metallic anode (X-ray tube) (Bushberg et al., 2002).
CT is an imaging modality where an X-ray tube is rotated around an object(s) and
the attenuation is recorded on a detector (Fig. 3). Other equipment may contain a
rotating stage in front of a fixed (non-moving) X-ray tube and detector (Bushberg
et al., 2002).
Data acquisition in a CT involves making X-ray transmission measurement
through the object(s) at various angles around the object(s), as seen in Fig. 4a.
Each X-ray that is acquired in CT is a transmission measurement through the
object(s) along a line, where the detector measures X-ray intensity (It). The un-

15

attenuated X-ray beam intensity (Io) is also measured during the scan by a
reference detector. The relationship between It and Io is given by:

I 𝑡 = 𝐼 𝑜 𝑒 −μt

(1-1)

where t is the object(s) thickness and μ is the average linear attenuation
coefficient along the X-ray path. It and Io are equipment dependent values, but
the product μt (parameter of interest), is related to the internal structure of the
object(s) along a given X-ray. When Eq. 1-1 is rearranged, the measured values
It and Io can be used to calculate μt, as seen in Eq. 1-2.

𝐼𝑜
μt = 𝑙𝑛 ( )
𝐼𝑡

(1-2)

where ln is the natural logarithm (constant approximately equal to 2.718281828),
t is canceled out, and the attenuation coefficient (μ) for each X-ray is used in the
CT reconstruction algorithm. This computation, which is a preprocessing step
performed before two-dimensional (2D) CT image reconstruction, reduces the
dependency of the CT image of the equipment-dependent parameters, resulting
in a 2D CT image that depends primarily on the object(s) density (Bushberg et
al., 2002). After preprocessing of the raw data, a CT reconstruction algorithm is
used to produce the 2D CT images. There are multiple reconstruction strategies,
however, filtered back-projection reconstruction is the most widely used. The
back-projection method uses planar projection data sets (preprocessed

16

sinogram) to build up the 2D CT image in the computer by reversing the
acquisition steps, as seen in Fig. 4b. During acquisition, a detector integrates μ
along a known path of each X-ray beam. During back-projection reconstruction, μ
for each X-ray is smeared along the X-ray path in the image of the object(s). In
addition to μ for each X-ray path, the reconstruction algorithm also records the
acquisition angle and position in the detector array. The back-projection algorithm
begins with an empty image matrix (512 x 512 pixels), and as the data from a
large number of X-ray paths (180 up to 1000) are back-projected onto the 2D CT
image matrix, a 2D CT image slice is generated (XY-plane-slice). In other words,
μ is added to each pixel in a line through the image corresponding to the X-ray
paths, as exemplified in Fig. 4c (van-Daatselaar et al., 2004; Goldman, 2008).

17

Figure 3. X-ray computed tomography (CT) A. Traditional CT working
principal of conveying objects through system, B. Measuring arrangement
of the GE (GE Healthcare, Buckinghamshire, England, Great Britain)
BrightSpeed™ RT 16 Elite CT used in these studies, C. Schematic
representation of a CT system scanning (5 chestnuts per row) containing
several 2D XY-plane CT images (slices).
18

Figure 4. CT imaging and data acquisition A. Transmission measurements
through the objects at numerous angles (black arrows) around the object,
B. Back-projected (reverse) transmission measurement onto 2D XY-plane
CT image matrix, C. Back-projection reconstruction from projection values
(P), for a simple image of four voxels.
19

In CT the difference in physical density of materials is visualized by
changes in image intensity and it is expressed in ‘Hounsfield-Units’ (HU) (or ‘CTnumber’). Hounsfield-Units represent the X-ray attenuation capabilities of a
specific material. In a 2D CT X-ray image, the HU(x,y) in each pixel (x,y), of the
image is generated through a combination of X-ray projection images and by
using Eq. 1-3,

𝐻𝑈( 𝑥,𝑦) = 1,000

𝜇(𝑥,𝑦) − 𝜇 𝑤𝑎𝑡𝑒𝑟
𝜇 𝑤𝑎𝑡𝑒𝑟

(1-3)

where (x,y), is the floating point number of the (x,y)-pixel before the XY-planeslice reconstruction, water is the attenuation coefficient of water (approximately
0.195), and the HU(x,y) is the Hounsfield Unit observed in the final 2D CT image.
Therefore, objects with a low-density like air, at standard temperature and
pressure, have a low-HU (-1000 HU), and high-density materials like bone will
have a high-HU (up to 3000 HU). In general, a HU-value equal to 0 stands for the
density of distilled water (1.0 g cm-3); values in the positive range represent
materials with a mass density above 1.0 g cm-3; and values in the negative range
stand for those below 1.0 g cm-3 (Bushberg et al., 2002).
In the case of several objects conveyed through a CT system, arranged in
rows, as seen with chestnuts in Fig. 3b, a single scanning of the CT system
consists of a block of 3D data stored as voxels. Voxels (volume elements), have
20

the same in-plane dimensions as pixels (2D image elements), but also include
the slice thickness (d) dimension (Bushberg et al., 2002) as observed in Fig. 3c.
However, not the entire block of data is acquired at once. Instead, each XYplane-slice is processed as the objects (chestnuts), previously arranged in rows
and attached to the boards, are passing though the CT scanner. A XY-plane-slice
is analogous to a virtual cross-section of the chestnuts imaged through the CT
scanner. Therefore, the imaging procedure is done one XY-plane-slice (crosssection) at a time, starting with t1- and ending with tn-XY-plane-slice, as outlined
in Fig. 3c. In addition, newer helical technology and multi-slice detectors are
available to allow multiple image slices to be obtained simultaneously while
constantly moving the samples through the imaging aperture (Bushberg et al.,
2002).
The originally acquired CT XY-plane-slices; containing images from
several chestnuts per row, moving through the Z-axis (longitudinal direction) are
stored in memory using a digital imaging and communications in medicine
(DICOM) standard format. DICOM images are composed exclusively of shades
of gray (16 bits), varying from black (low-density like air) at the weakest intensity
to white at the strongest (high-density materials like bone). When visualized in
traditional DICOM visualization software’s, like the Osirix Imaging Software
V3.6.1 (http://www.osirix-viewer.com/) image pixels are expressed in HU. On the
other hand, when imported into other software’s like MATLAB (2012a, The
MathWorks, Natrick, MA, USA) (http://www.mathworks.com), DICOM images
contain 65536 shades of gray (intensity values) per pixel, ranging from 0 to

21

65535, which are mapped to the original HU-values. In addition, DICOM files
contain metadata that can easily be accessed, providing detailed information
about the image, such as the size, dimensions, modality used to create the
image, and equipment settings used to capture the image. For comprehensive
information about the standard the reader can refer to the official DICOM web site
(http://medical.nema.org/).
There are multiple advantages of CT compared to traditional 2D projection
X-ray imaging. First, CT completely eliminates the superimposition of images of
structures within the samples and outside the region of interest. This is because
in 2D X-ray imaging, only one projection image (X-ray transmission through
sample) is acquired per object of interest. In contrast, depending on the CT
scanner type, a CT image is acquired at different angles by reconstructing
information obtained from 180 up to 1000 X-ray 2D transverse projection images
(van-Daatselaar et al., 2004; Goldman, 2008). Second, because of the intrinsic
contrast and high resolution of CT, physical density differences between
materials as low as 0.5 % can be easily differentiated (Bushberg et al., 2002).
Thirdly, the data from one CT imaging procedure can be reconditioned to be
observed in various planes (different angular orientations) as seen in Fig. 5, or
even observed volumetrically by creating a 3D image and merging the
information from several 2D slices. Using a chestnut as a model in Fig. 5, each
chestnut contains between 8 to 17 XY-plane-slices representing virtual crosssections of a chestnut along the longitudinal (Z) axis. This number of slices is

22

dependent on chestnut physical size and d, which in this case d = 2.56 mm. CT
image slices can be viewed in different planes, such as 90 from the XY-plane
toward the longitudinal axis (Z) – YZ-plane-slice, and 90 toward the horizontal
axis (X) – XZ-plane-slice. This process of reconditioning CT images so that they
can be observed in various planes (different angular orientations) is commonly
known as re-slicing (Bushberg et al., 2002), As before, depending on chestnut
size and d, each chestnut contains 8 to 17 YZ-plane-slices and XZ-plane-slices.

23

Figure 5. (a) 3D virtual cross-sections representation of a chestnut at three planes (different angular orientations)
along the horizontal (X), vertical (Y), and longitudinal (Z) axes (Figure not to scale). (b) Original series of acquired
CT XY-plane-slices.

24

Figure 5 (cont’d).

(c) Reconditioned (re-sliced) CT YZ-plane-slices. (d) Re-sliced CT XZ-plane-slices.

25

1.6 Application of CT in non-medical industries
CT has proven to be an invaluable diagnostic tool for numerous
applications in the medical field and material sciences (Bushberg et al., 2002),
ranging from tumor diagnosis and segmentation (Stadler et al., 2004; Deglint et
al., 2007) to osteoporosis screening (Bushberg et al., 2002) to differentiation of
encountered foreign bodies in corpses (Bolliger et al., 2009).
In agriculture and animal science fields, off-line adaptation of CT
technology has proven to be an accurate nondestructive descriptor of inner
properties. For example, water content of oak (Quercus robur) and spruce (Picea
abies) wood was accurately determined by Fromm et al. (2001) and an
estimation of lean meat content and quality was made in pigs (Sus spp.) by
Furnols et al. (2009). In agriculture, has been shown to be an accurate descriptor
of internal characteristics.
Postharvest internal evaluation of several nuts, fresh vegetables, and fruits
has been recently shown to be achievable as well. Jha et al. (2010) described the
potential of nondestructive techniques, including CT, to measure mango
(Mangifera indica) internal quality (i.e. size, shape, pulp and moisture), and
Kotwaliwale et al. (2006) differentiated nutmeat from shell components in fresh
pecans (Carya illinoinensis). Barcelon et al. (1999b; 1999a) used CT to study
internal quality of peaches (Prunus persica) and mango, including fruit density,
moisture content, soluble solids, and acidity. In addition, studies done by
Sornsrivichai et al. (2000) determined pineapple fruit (Ananas comosus)

26

translucency and ripeness. Moreover, using high resolution CT, which is a
sampling tool that combines less than 1 to 2 mm thick CT image slices with a
high spatial frequency reconstruction algorithm to generate CT images that show
small details (Kelly et al., 2003), Mendoza et al. (2007) measured pore space in
apples (Malus domestica), Lammertyn et al. (2003) obtained three dimensional
(3D) spatial distribution of ‘Conference’ pear (Pyrus sp.) core breakdown, and
Verboven et al. (2008) studied 3D gas exchange pathways in pears and apple
1.7 Objectives and Hypothesis
The comprehensive objective of the study is to develop the methods to
nondestructively visualize and classify fresh chestnuts, based on their internal
quality, using X-ray CT imaging. This study will provide a powerful tool to sort
chestnuts based on their kernel quality, leading to the improved marketing of
attractive, safe, high quality chestnuts.
The specific objectives of the research are:
1. To determine the combined effect of image acquisition parameters,
which include voltage (kV), current (mA), and slice thickness (mm) on optimizing
CT image quality (signal to noise ratio, volume accuracy, high contrast spatial
resolution and low contrast detectability), using response surface methodology
(CHAPTER 2).
2. To establish effective image visualization techniques that will infer
internal quality attributes of fresh chestnuts (healthy and decayed tissue, pellicle,
void spaces and air), using X-ray CT images (CHAPTER 3).

27

3. To apply image pre-processing techniques (cropping and image
enhancement), CT image segmentation, image feature extraction, feature
selection, multivariate discrimination algorithms and artificial neural network
(ANN) classifiers, which will automatically categorize fresh chestnuts based on
their internal quality, using X-ray CT images (CHAPTER 4).
The hypothesis of this research is that by measuring multiple X-ray CT
image features, using different image analysis techniques, in combination with
multivariate statistical discrimination algorithms and ANNs classifiers, it will be
possible to categorize chestnuts based on their internal quality. The long-term
goal is to develop a system that automatically classifies chestnuts by their
internal quality, after harvest. This nondestructive classification method will
eventually replace the questionable water floating technique, currently used
worldwide, to separate decayed from healthy chestnuts.

28

CHAPTER 2. APPLICATION OF RESPONSE SURFACE METHODOLOGY
(RSM) TO SYSTEMATICALLY OPTIMIZE FRESH CHESTNUT COMPUTED
TOMOGRAPHY (CT) IMAGE QUALITY
2.1 Abstract
The objective of this chapter is to describe a method that was developed to
systematically and efficiently obtain a model of CT scanning factor levels, which
returns optimized high quality chestnut (Castanea spp.) CT images. Chestnut
two-dimensional CT images were used to describe this optimization procedure,
considered to be a critical step in the development of a fast, nondestructive
technique, capable of assessing fresh internal quality attributes and components
of chestnuts, and other agricultural commodities. Response Surface Methodology
(RSM), using a three-factor, three-level Box-Behnken statistical design, was used
to optimize the factors affecting image quality, which include X-ray voltage,
current, and slice thickness. Response variables representing image quality were
digitally and automatically inferred from fresh chestnut image signal-to-noise
ratio, Teflon® cylinder reference volume accuracy, a Quality Assurance (QA) high
contrast spatial resolution phantom, and a QA low contrast detectability phantom.
Second-order RSM prediction models for each response variable reflected a
combined maximized CT image quality at a voltage, current, and slice thickness
equal to 120 kilovolts, 170 milliamps, and 2.5 millimeters respectively. The
experiment yielded optimal chestnut CT images that can accurately reflect
internal decay of fresh chestnuts with an overall accuracy rate equal to 96 %,

29

taking as reference data the subjective quality rating of five trained chestnut
experts.
2.2 Introduction
Even though image quality is an important factor that has been addressed
in the medical and veterinary fields (Mayo et al., 1994; Ford et al., 2003; Du et al.,
2007) little is known about how to statistically, routinely, systematically, and
reliably optimize CT parameters.
Response Surface Methodology (RSM) is a compilation of mathematical
and statistical methods suitable to optimize, develop and improve processes
(Myers et al., 2008), which can be helpful to quantitatively and routinely adjust
parameters that influence CT image quality. RSM has never been applied to
optimize CT image quality, but other engineering fields heavily rely in this
technique to optimize processes, including the optimization of protease
production (Braga et al., 2010), monitoring of ball bearings (Patil et al., 2010),
and ultrasonic-stimulated solvent extraction (Wang et al., 2011). The RSM makes
usage of factorial designs when a few significant parameters or factors are
involved in optimization (Myers et al., 2008), including three-level complete
factorial design, central composite design, equiradial designs, and Box-Behnken
design among others (Myers et al., 2008). The Box-Behnken design is a
balanced incomplete three-level block design, useful to fit second-order response
surfaces, and in certain cases, it is as powerful as a three-level complete factorial
design, only that it requires 15 runs instead of 27 without repetitions (Myers et al.,
2008).

30

Therefore, the aim of this study is to describe RSM using a Box-Behnken
design as a reliable method to accurately and systematically optimize CT
scanning parameters of utmost importance, which include voltage, current, and
slice thickness, essential to obtain good quality chestnut CT images as described
in theory by Renaudin et al. (1993), Mayo et al. (1994), Sharma et al. (2006),
Goldman (2007), Sellakumar et al. (2007) and Arnold et al. (2010). This research
offers a tool that will methodically yield useful, and high quality CT images in any
field related with CT. Optimized chestnut CT images are obtained without the
biased aid of human intervention as traditionally done in practice (General
Electric Company, 2007), and without using unnecessary scanning runs, which
are time consuming and computationally inefficient. More specifically, this method
reliably yielded maximized quality CT images, which accurately reflect inner
components of chestnuts.
2.3 Materials and methods
2.3.1 Chestnut sample collection, reference cylinders, and Quality Assurance
(QA) phantom CT imaging scans
A total of 50 physiologically mature and apparently healthy Chinese
chestnuts (C. mollissima) were collected by hand, directly from chestnuts trees in
a Michigan orchard. Following common postharvest procedures, samples were
immediately submerged for 300 s in 5 L of room temperature distilled water
containing 2700 μL L-1 hydrogen dioxide plus 200 μL L-1 peracetic acid

31

(Storox®1, BioSafe Systems, Glastonbury, CT, USA) with the objective of
reducing mold contamination, and stored at 4 °C. During this process, the
majority of decayed, empty or damaged chestnuts were eliminated by their
proclivity to float, as healthy chestnuts tend to sink (Donis-González, 2008;
Donis-González, 2008; Donis-Gonzalez et al., 2010). As it was important to
obtain uniform healthy tissue matter for the signal-to-noise estimation, after 20 d,
9 of the apparently healthiest chestnuts were attached to a rectangular acrylic
transparent board (2.75 mm x 300 mm x 24 mm), using approximately 50 g of
transparent silicone per chestnut, as seen in Fig. 6a.

Figure 6. (a) Fresh chestnuts samples, and Teflon® reference cylinders, (b)
Testing phantom, (c) 16-bit CT images of fresh chestnut samples for Signalto-noise Ratio (SNR) calculation, and Teflon® reference cylinders for
volume accuracy (mm-3) estimate. (d) 8-bit CT image of QA phantom used
to infer High Contrast Spatial Resolution (HCSR). (e) 8-bit CT image of QA
phantom used to infer Low Contrast Detectability (LCD).

1

Storox® is a registered Trademark BioSafe Systems.

32

To avoid sample confusion individual chestnuts were numbered. Parallel to
the chestnuts, one row containing three Teflon®2 (Polytetrafluoroethylen)
reference cylinders (38 mm diameter x 13 mm height) (Applied Plastics
Technology Inc., Bristol, RI) were also attached to the board, using approximately
85 g of transparent silicone per Teflon® cylinder (Fig. 6a).
CT scans were performed in a GE BrightSpeed®3 RT 16 Elite, multidetector CT instrument (General Electric Healthcare, Buckinghamshire, United
Kingdom) to the board containing chestnuts and Teflon® reference cylinders,
resulting in a set of 2D CT images as seen in Fig. 6c, as well as to a standard QA
phantom (Series QA-Phantom, GE BrightSpeed®, General Electric Healthcare,
Buckinghamshire, United Kingdom) (Fig. 6b), resulting in a set of 2D CT images
(e.g. Fig. 6d and Fig. 6e), as specified by General Electric Company (2007). Even
though, not used to its fullest capabilities, because of fast CT scanning protocols,
the CT instrument contains a Volara®3 digital data acquisition system (24-bit),
which has the capability of acquiring images with a 0.65 mm isotropic image
resolution (smallest detected structure), meaning that the spatial resolution in the
horizontal-vertical plane (XY) and that in the longitudinal planes (XZ and YZ) can
be the same (Tsukagoshi et al., 2007). Scanning parameters are in Table 1.
2

Teflon® is a registered Trademark of E.I. du Pont de Nemours and Company.

3

BrightSpeed® and Volara® are registered Trademarks of General Electric

Healthcare.

33

Table 1. Scanning parameters for the CT – General Electric, BrightSpeed™
RT 16 Elite (GE Healthcare, Buckinghamshire, England, Great Britain)
Units
Parameter
Voltage (kV)

80, 100, 120

Current (mA)

10, 90, 170

Slice thickness (pixel size in the Z-axis) (mm) – d

0.625, 2.5, 5

Pixel size (mm) in the X- and Y-axes

0.73

Resolution in the XY-plane (pixels mm-1)

1.37

Pixel area in the XY-plane (mm2)

0.53

Pixel area in the XZ- and YZ-planes (mm2)

1.82

Voxel volume (mm3)

0.46a, 1.82b, 3.65c

Pitch (table movement – mm : rotation)

17.5:1

Time per rotation (s)

1.2b,c, 1.7a

Scan time (s) per 1000 mm of board-length

48b,c, 96a

Reconstruction matrix

512 x 512

Field of view (FOV) (mm)

500

Original image intensity resolution

16-bit

CT images per 1000 mm of board-length (tn)

1600a, 400b, 200c

Resolution in the XZ- and YZ-planes (pixels mm-1) 0.625a, 2.5b, 5.0c

a

Slice thickness (d) = 0.625 mm, bd = 2.5 and cd = 5.0 mm.

2.3.2 RSM Box-Behnken design for CT image quality optimization
Box-Behnken three-factor, three-level statistical design was used to
optimize and evaluate main effects, interaction effects, and quadratic effects of
the CT scanning parameters on image quality. In this study, image quality is
characterized by four of the most important CT image quality measurements,
including fresh chestnut image Signal-to-noise Ratio (SNR), reference cylinders
Volume Accuracy (VA), High Contrast Spatial Resolution (HCSR), and Low
34

Contrast object Detectability (LCD) using a standard QA phantom (Goldman,
2007). Definition and description of how each of the image quality attributes was
calculated is included in sections 2.3.3 through 2.3.5.
The Box-Behnken design is specified by a set of points at the midpoint of
each edge of a multidimensional cube and a center point triplicate, being
appropriate to construct a second order polynomial model of the combined effect
of scanning parameters (factors) to each response variable (Myers et al., 2008).
The computer-generated models for every response variable representing image
quality were developed using the language and environment for statistical
computing software R (V2.10.0, R Development Core Team, Vienna, Austria).
Models with coded-factors are specified as Eq. 2-1.
(2-1)

where Yn is each measured response or dependent variable representing image
quality analyzed separately, associated with each factor level combination; b0 is
the regression intercept; b1 through b33 are the regression coefficients; and X1,
X2, and X3 are the independent variables or factors. Levels for each independent
scanning variable, which represent values for each factor were coded as low (-1),
medium (0), and high (+1) (Myers et al., 2008) (Table 2 and 3). Levels were
selected based on equipment capabilities, prior preliminary experimentation,
chestnut morphology, and size (Miller, 2003). The selected dependent, coded,
and un-coded independent variables are shown in Table 3, showing how each
35

independent variable changes from one run to another. Un-coded voltage (X1),
current (X2), and slice thickness (X3) parameter combinations used to perform
CT scanning for each of the 15 runs are specified in Table 3. One run is
considered as the procedure where all the CT images in total are acquired from
the board containing fresh chestnuts, Teflon® reference cylinders, and the
standard QA phantom using the combination of specified parameters. The QA
phantom images were acquired separately, from the board.
Table 2. Box-Behnken experimental design variables
Levels (coded) / un-coded
Variables (units)

Notation

Low (-1)

Medium (0)

High (+1)

X1 - Voltage (kV)

V

80

100

120

X2 - Current (mA)

C

10

90

170

X3 - Slice thickness (mm)

ST

0.625

2.5

5.0

Dependent variables
(Response - Yn)
Y1 - Signal-to-noise Ratio

SNR

Indep. variables (Factors –
Xf)

Y2 - Volume Accuracy
(mm-3)

Y3 - High Contrast Spatial
Resolution
Y4 - Low Contrast
Detectability

VA
HCSR
LCD

After generating the RSM models presented in Table 2, the process was
individually optimized for each response Yn, obtaining an optimized model per

36

response variable. Optimization was performed to find the levels of voltage (X1),
current (X2), and slice thickness (X3) maximizing each Yn.

37

Table 3. Design matrix (un-coded factors) in respect to each response variables
Independent variables (Factors)
Dependent variables (Response)a
Run

a

Voltage
(kV)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

100
80
120
100
80
120
100
100
100
80
120
100
80
120
100

Slice
Current
thickness
(mA)
(mm)
10
90
90
170
10
10
90
90
90
170
170
10
90
90
170

0.625
0.625
0.625
0.625
2.5
2.5
2.5
2.5
2.5
2.5
2.5
5.0
5.0
5.0
5.0

SNR ± SD

Volume
accuracy ±
SD (mm-3)

HCSR ± SD

LCD ± SD

0.14 ± 0.19b
0.27 ± 0.54ab
0.52 ± 0.41ab
0.49 ± 0.38ab
0.17 ± 0.32ab
0.38 ± 1.14ab
0.59 ± 0.56ab
0.62 ± 0.57a
0.61 ± 0.52a
0.51 ± 0.33ab
0.66 ± 0.45a
0.39 ± 0.43ab
0.59 ± 0.52ab
0.69 ± 0.62ab
0.70 ± 0.59ab

1.56 ± 0.35bc
2.31 ± 0.47c
2.39 ± 0.53c
2.58 ± 0.36c
0.67 ± 0.30ab
0.68 ± 0.26ab
0.61 ± 0.25ab
0.62 ± 0.54ab
0.65 ± 0.22ab
0.63 ± 0.52ab
0.60 ± 0.14ab
0.61 ± 0.02b
0.43 ± 0.24ab
0.45 ± 0.04a
0.42 ± 0.01a

0.08 ± 0.01a
0.19 ± 0.02b
0.30 ± 0.04c
0.39 ± 0.06c
0.10 ± 0.02a
0.21 ± 0.03c
0.40 ± 0.08c
0.43 ± 0.09c
0.44 ± 0.07c
0.44 ± 0.13c
0.52 ± 0.03d
0.27 ± 0.08bc
0.39 ± 0.13c
0.50 ± 0.13cd
0.53 ± 0.15cd

0.16 ± 0.01a
0.21 ± 0.03b
0.28 ± 0.08b
0.24 ± 0.06b
0.20 ± 0.06ab
0.30 ± 0.02b
0.52 ± 0.04c
0.46 ± 0.09c
0.46 ± 0.08c
0.42 ± 0.08c
0.67 ± 0.15cd
0.32 ± 0.03b
0.47 ± 0.08c
0.80 ± 0.16d
0.72 ± 0.17d

ANOVA for each run at P = 0.05 and Tukey multiple comparison of means post-hoc test. Values in each run followed by
the same letter within dependent variable are not significantly different.

38

2.3.3 Visual based fresh chestnut quality evaluation and SNR (Y1)
Immediately after scanning, each fresh chestnut was transversely sliced in
4-sections using a sharp hand knife. Slice sizes varied depending on chestnut
size (5 to 7.5 mm thickness). All internal kernel faces between each slice (total of
6) were then qualitatively assessed for disorders. Slice internal faces, were
scanned using a 48-bit color, 9600 x 4800 dots-per-inch (DPI) charge-coupleddevice (CCD) scanner (Scan Maker S400, Microtek International Inc., China),
using the ScanWizard 5 (Microtek International Inc., China) standard image
acquisition software, yielding a tagged image file format (tiff) color image, with a
resolution of 816 x 1123 pixels. Color scanning was performed for record keeping
and to use as references, to accurately determine uniform healthy fresh tissue for
SNR calculation from the 2D chestnut CT images. Before every scan, the
scanner was thoroughly cleaned, using compressed air in combination with
wiping the scanning glass with delicate task wipes, which had been previously
soaked in mild non-streak glass cleaner. To avoid variability between images,
and to stabilize the intensity of the scanner lamp, the scanner was on for at least
15 min before scanning. It is important to mention that the scanner, which was
used in this study, is internally calibrated every time it is tuned on, so no
calibration

and/or

calibration

targets

are

required

(http://support.microtek.com/product_dtl_2.phtml?prod_id=38).
The SNR is defined as the measurement that compares the level of signal
in the fresh chestnut CT image to the level of background noise (graininess). It is
calculated as the ratio of the mean (μ) Hounsfield Unit (HU) value of a 9 pixels2
39

square region of interest (ROI) (5.42 mm2), as seen in Fig. 7a containing fresh,
healthy, uniform chestnut tissue to the standard deviation (SD) of the ROI, using
Eq. (A. 1) in the appendix section A. Higher values imply higher signal in
comparison to noise therefore yielding better quality images with less graininess
(Goldman, 2007).

40

Figure 7. (a) Example of chestnut 16-bit CT image slice used for Signal-tonoise Ratio (SNR) calculation, from a region containing fresh uniform
chestnut tissue, using reference color image. (b) Example of Teflon®
reference cylinder 8-bit CT image used to estimate volume accuracy (mm-3)
with binary image after segmentation (simple global threshold of 135). (c)
Example of cropped 8-bit CT image used to calculate High Contrast Spatial
Resolution (HCSR) with binary image after segmentation (simple global
threshold of 134). Each pattern consists of five bars and spaces called line
pairs (lp). The sizes of the patterns are equivalent to 1.6 mm, 1.3 mm, 1.0
mm, 0.6 mm, and 0.5 mm, respectively. (d) Example of cropped 8-bit CT
image used to determine Low Contrast Detectability (LCD) with binary
image after segmentation (simple global threshold of 94). Image displays
various sized Holes (H). The diameter of each H equal 10.0 mm, 7.5 mm, 5.0
mm, 3.0 mm, and 1.0 mm.

41

2.3.4 Volume accuracy (Y2) using Teflon® reference cylinders
Volume accuracy estimation was calculated from Teflon® reference
cylinders CT images, following Eq. (A. 2) found in the appendix section. Volume
accuracy (mm-3), is defined as the extent to which the estimated volume of each
of the Teflon® reference cylinders, using 8-bit CT images agrees with the mean
standard true volume (~ 14,743 mm3) of three reference cylinders (Fig. 7b).
Higher volume accuracy values represent better overall 3D information from a
stack of 2D CT images.
Volume accuracy estimate is assumed to be accurate, because the
reference cylinders were precisely molded using compression technology
(Applied Plastics Technology Inc., Bristol, RI) with pure, high HU (900), high
quality Teflon®.

In addition, after cylinders were precisely and carefully cut,

dimensions of each cylinder were confirmed using a digital caliper (ABSOLUTE
Digimatic Caliper Series 500, Mitutoyo, Singapore) with a resolution equal to 0.01
mm.
2.3.5 HCSR (Y3) and LCD (Y4) calculation using a standard QA Phantom
The HCSR and LCD are computed using a QA phantom, which is
optimized to be used in the CT imager used in the study (General Electric
Company, 2007); and applicable to this study, because its size and densities
resemble those of the commodity of interest (i.e. fresh chestnuts) (Miller, 2003;
Du et al., 2007).

42

In practice HCSR is estimated visually, with the aid of some quantitative
measurements of resolution (General Electric Company, 2007). In this study,
HCSR was measured using image analysis techniques (e.g. Fig. 7c) applying the
derived Eq. (A. 3) in the appendix section. The HCSR of a CT image is a
measurement that defines how well two high contrast objects placed close
together are distinguished, and how small can these objects be visualized (Du et
al., 2007; General Electric Company, 2007; Sande et al., 2010). In this study,
higher HCSR values represent images that can better distinguish between two
high contrast objects placed close together.
The LCD represents the ability of a CT image to discriminate objects that
vary slightly from their background (low contrast), defined by the smallest visible
object (Du et al., 2007; General Electric Company, 2007). Conventionally LCR is
determined visually as specified by General Electric Company (2007). In this
study, LCD was calculated without the help of a human observer using derived
Eq. (A. 4), as found in the appendix section, and using images as the one
exemplified in Fig. 7d (General Electric Company, 2007; Sande et al., 2010).
High LCD values represent images that can better discriminate small objects
(sensitivity) (Du et al., 2007; Goldman, 2007).
2.3.6 Optimized 2D CT image quality validation from an independent data set
A completely different set of 2D CT images were acquired with optimized
scanning parameters as summarized in Table 5, from 266 ‘Colossal’ (C. sativa x
C. crenata), 266 Chinese independent chestnut samples, three Teflon®
reference cylinders, and the QA phantom. CT image acquisition and processing
43

was the same as described in sections 2.3.1 and 2.3.2, except that in the case of
the chestnuts, scanning boards were larger (915 mm x 335 mm), and naturally
decayed chestnuts were not eliminated. After scanning, all chestnuts were sliced
and color scanned for record keeping, and to use as samples for CT image
quality validation, as described in section 2.3.3.
The purpose was to first determine experimental CT image quality
attributes (SNR, VA, HCSR, and LCD) from this independent set of 2D CT
images, using ten repeated measurements (n = 10: 5-‘Colossal’ and 5-Chinese)
for SNR, and three repeated measurements (n = 3) for reference cylinder VA,
HCSR, and LCD. These experimental values of response were then compared
with predicted values, using RSM models (Table 6) and optimized scanning
parameters (Table 5).
Second, CT images were used to study their effectiveness in reflecting
fresh chestnut internal decay levels, in comparison with color image slices
(optimized CT image quality validation). Fresh chestnut quality was elucidated,
using the Subjective Quality Rating (SQR) of a 5-expert panel of individuals
working closely with chestnut production and research. Panelists were
experienced in detecting, identifying, and quantifying quality attributes (decay) in
fresh chestnuts. Each expert was presented with 11 randomized chestnut color
image slices per cultivar (e.g. Fig. 7a and Fig. 8a), for subjective quality rating
(total of 110 - n). The SQR was expressed as the apparent percentage of decay
tissue in relation to the total area of each color image. A value equal to 100
represents a chestnut slice that is completely decayed, while 0 indicates that the

44

chestnut does not contain decay. The SQR values obtained from the panel of
experts were then linearly modeled in relationship to the Digital Quality
Assessment (DQA) from their corresponding 2D CT images (Fig. 8b to Fig. 8e).
The DQA was calculated using image analysis techniques as explained the
appendix section, by using Eq. (A.1). DQA was expressed as the percentage of
decay tissue in relation to the total area of each CT image. A value equal to 100
represents a chestnut slice that is completely decayed, while 0 describes a
chestnut that does not contain decay.

Figure 8. Example of images used for Digital Quality Assessment (DQA). (a)
Reference color image. (b) 16-bit optimized CT gray scale image. (c) Binary
image of whole chestnut after segmentation, applying a simple global
threshold of 400 HU to Fig. 8(b). (d) Binary image of transition points and
pellicle after employing a Sobel filtering method to detect edges in Fig. 8(b).
(e) Binary image of healthy tissue after segmentation, using a simple global
threshold of 1050 HU to Fig. 8(b). Figure is partially presented in color.
Calculations, image processing and analysis, from sections 2.3.1 through
2.3.6 were done in MATLAB (2009a, The MathWorks, Natrick, MA). Data
analysis for section 2.3.6 was performed using R (V2.10.0, R Development Core
Team, Vienna, Austria).

45

2.4 Results
2.4.1 RSM Box-Behnken design for CT image quality optimization
The RSM using Box-Behnken design proved to be a reliable method that
offered a clear perception in how the diverse factor level combinations of X-ray
tube voltage, current, and slice thickness resulted in different responses for each
of the CT image quality measurements (SNR, VA, HCSR, and LCD), as seen in
Table 3. The Box-Behnken design presented sufficient information to test the
lack-of-fit for each of the non-linear response models, as presented in the
Analysis of Variance Analysis (ANOVA) (Table 4). This affirms that this design is
as robust in terms of prediction performance, in comparison with a complete
factorial design or others (Myers et al., 2008).
Table 4 shows that all P values for the lack-of-fit criteria are non-significant
( ³ than 0.05), suggesting that the second order non-linear models can accurately
predict the effect of the independent variables (factors) for each of the response
variable (dependent) independently. Furthermore, the approach indicates that the
non-linear models are optimally designed to provide an ideal insight on the
behavior of a second-order response surface, including the inference of
maximized stationary points (optimized conditions) as presented in Table 5
(Myers et al., 2008). The relationship between the dependent and independent
variables was completely interpreted, displayed, and studied in depth, using a set
of response surface and interaction plots (Fig. 9-12), in combination with the
regression coefficients for each response model, as summarized in Table 6. The
values in Table 6 corresponding to the coefficients for each coded factor (voltage
46

(X1), current (X2), and slice thickness (X3)) are related to the effect of each of the
factors on the change to each response variable (Yn). In other words, coefficients
are used to study the mathematical relationship in the form of a polynomial
equation for each of the measured response (Yn), as seen in Eq. 2-1 (Myers et
al., 2008).

47

Table 4. ANOVA table for Signal-to-noise Ratio (SNR), Volume Accuracy (VA), High Contrast Spatial Resolution
(HCSR), and Low Contrasts Detectability (LCD) non-linear response models
Term

df

Linear

Sum of squares (SS)

Mean square (MS)

F-value

SNR

VA

HCSR

LCD

SNR

VA

HCSR

LCD

SNR

VA

HCSR

LCD

3

0.01

7.2

0.40

4717.0

0.00

2.39

0.14

1572.4

210.4

143.7

182.38

236.8

3

0.00

0.2

0.01

327.3

0.00

0.06

0.00

109.1

6.4

3.3

2.79

16.4

3

0.00

0.4

0.02

536.8

0.00

0.12

0.01

178.9

24.0

7.1

11.63

26.0

Res.
Err.

5

0.00

0.1

0.01

46.5

0.00

0.02

0.00

6.64

Lackof-fit

3

0.00

0.1

0.00

41.5

0.00

0.02

0.00

8.30

4.4

8.2

1.59

3.32

Pure
error

2

0.00

0.0

0.00

5.0

0.00

0.00

0.00

2.50

Quadr
Polyn.

48

Table 4. Cont’d

Term

df

Linear
Quadr
Polyn.

SNR

VA

HCSR

LCD

3

0.0a

0.0a

0.0a

0.0a

3

0.0a

0.1

0.1

0.0a

3

0.0a

0.0a

0.0a

0.0a

0.2

0.1

0.4

0.3

Res.
Err.

5

Lackof-fit

3

Pure
error
a

P-valuea

2

ANOVA for each dependent variable at P = 0.05

49

Table 5. Stationary maximized points in original units for each non-linear response variable model
Response
Factors
Eigenvalues
variable
SNR

Voltage

Current

Slice

Voltage

Current

Slice

113
(kV)

140
(mA)

-0.003
(kV)
0.757

-0.005
(mA)
0.085

-0.034

-0.038

-0.012
thickness
-0.041
(mm)
-0.066

-9.502

-7.497

-1.250

VA (mm-3)
HCSR

> 170a
170

LCD

120

117

4.4

Max

> 120a
113

> 170a
140

5.0

Min
Optimized
a

> 120a
113

5.0
thickness
< 0.625a
(mm)
5.0

120

170

< 0.625a
2.5

Maximum stationary point was not discerned.

50

Table 6. Regression coefficients and P valuesa for each non-linear response variable model
Term
SNR
Volume accuracy
HCSR
LCD
(code)

Coeff.b

p-value

-3)
Coeff. (mmp-value

Coeff.

p-value

Coeff.

p-value

Intercept (b0)

0.5970

0.000a

b
0.5985

0.000a

0.4238

0.000a

0.4893

0.000a

Voltage (X1)

0.1100

0.000a

0.0000

0.767

0.0458

0.001a

0.1695

0.000a

Current (X2)

0.1533

0.000a

0.0814

0.085

0.1465

0.000a

0.1243

0.000a

Slice thickness (X3)

0.0113

0.000a

-0.8775

0.000a

0.0868

0.000a

0.0837

0.000a

Voltage x Current (X1X2)

-0.0003

0.811

-0.0316

0.962

-0.0030

0.813

0.0584

0.125

Voltage x Slice thickness

-0.0019

0.202

-0.0279

0.387

0.0035

0.784

0.0531

0.123

Current
(X1X3) x Slice thickness

0.0002

0.896

-0.2595

0.009a

-0.0074

0.567

0.0243

0.058

(X2X3) 2 (X12)
Voltage

-0.0040

0.033a

0.0830

0.834

-0.0367

0.032a

-0.0446

0.114

Current2 (X22)

-0.0120

0.000a

-0.0170

0.219

-0.0652

0.002a

-0.0883

0.009a

Slice thickness2 (X32)

-0.0036

0.004a

0.7553

0.000a

-0.0366

0.033a

-0.0103

0.688

a

ANOVA for each term at P = 0.05, giving a statistical significance summary of the main dependent variable effects their

interactions. bNote that Non-linear coefficients are scaled to the coded factors, as specified in Eq. (2-1).

51

Coefficients with higher values represent a higher probability for statistical
interaction effects and quadratic relationship. A positive value signifies an effect
that favors the response, while a negative value indicates an antagonistic effect
(Myers et al., 2008).
2.4.2 SNR (Y1) optimization
Regression coefficients for the SNR (Table 6) indicate that the individual
effect of voltage, current, and slice thickness are statistically significant, since the
corresponding P values are ≤ 0.05. No interaction effects between any of the
factors appear to be statistically significant. Table 4 shows the result of ANOVA
for the CT image SNR, in addition it was determined that the R2 and R2 (adjusted
for df) are 99.04 % and 97.81 %, respectively. This indicates that the factors
remarkably explain the amount of variation in the observed values of SNR.
Additionally, it is clear that the first, second, and polynomial order terms
contribute significantly to the model (P value ≤ 0.05), so the canonical analysis or
the relationships between factors in the data set is relevant (Myers et al., 2008).
The stationary point is near the experimental region, and the three eigenvalues
are negative, indicating that the stationary point is a maximum SNR (Table 5);
clear evidence a set of optimal scanning conditions was inferred. Optimized
response for SNR was established to be at a voltage, current, and slice thickness
equal to 113 kilovolts (kV), 140 milliamperage (mA), and 5.0 mm respectively.
The surface response plots shown in Fig. 9 (a-c) indicate that SNR increases with
an increase in all factors (voltage, current, and slice thickness), consequently

52

increasing image quality. This can be visually confirmed in Fig. 13, where the
combination of low voltage, current, and slice thickness yield CT images with low
SNR, hence CT images have poor quality (i.e. high graininess) (Bushberg et al.,
2002; Du et al., 2007; Goldman, 2007). This graininess phenomenon is
consistent with the theoretical fact that specifies that by increasing voltage,
current, and slice thickness, a higher number of X-rays will be detected by the CT
equipment, yielding better quality images with a higher SNR, and therefore less
graininess (Bushberg et al., 2002; Du et al., 2007; Goldman, 2007).

53

Figure 9. Surface plots of SNR versus (a) slice thickness (ST - mm) for current (C - mA), (b) slice thickness (ST mm) for voltage (V - kV).

54

Figure 9 (cont’d).

(c) current (C - mA) for voltage (V - kV).

55

2.4.3 Volume accuracy (Y2) optimization
Regression coefficients for VA (Table 6) indicate that the individual effect
of slice thickness is the only statistically significant factor affecting volume
accuracy, since the corresponding P value is ≤ 0.05. It is also observed that the
interaction effect between slice thickness and current also has a significant
contribution (P value ≤ 0.05). Table 4 shows the result of ANOVA for the CT
image volume accuracy model that yielded an R2 and R2 (adjusted for df) of
98.51 % and 96.59 %, respectively. This indicates that the factors highly explain
the amount of variation in the observed values of volume accuracy. Additionally, it
is clear that the first, and polynomial order terms contribute significantly to the
model (P value ≤ 0.05), so the canonical analysis or the relationships between
factors in the data set is relevant (Myers et al., 2008). The stationary point is not
near the experimental region, and only one out of the three eigenvalues are
negative, indicating that the stationary point could not be determined (Table 5).
Therefore, clear evidence of a set of optimal scanning conditions could not be
inferred (Myers et al., 2008).
Even though optimized conditions could not be determined, the surface
response plots shown in Fig. 10 (a, c, and d) indicate that volume accuracy is
directly proportional to the voltage and the current, while volume accuracy is
inversely proportional to slice thickness. Consequently, the combination of high
slice thickness, low current, and voltage yield CT images with low volume
accuracy, hence CT images under these conditions offer low 3D information.

56

Additionally, by statistically slicing, therefore evaluating the slice thickness and
current interaction effect (Fig. 10-b), it can be seen that only low (0.625 mm) and
high (5 mm) slice thickness causes a significant change in volume accuracy
dependent of current. Medium (2.5 mm) slice thickness does not seem to
significantly change volume accuracy with the variation of current. These results
are consistent with other studies in the medical (Bushberg et al., 2002; Goldman,
2007) and agricultural (Mendoza et al., 2007) fields, were it has been found that
by reducing slice thickness, better overall 3D information or volume accuracy will
be digitally perceived from CT images.

57

Figure 10. (a) Surface plot of volume accuracy (VA - mm-3) versus slice thickness ( ST - mm) for current (C - mA).

(b) Interaction plot for volume accuracy (VA - mm-3) versus slice thickness (ST - mm) for current (C - mA). Data
points followed by the same lower case letter within the same slice thickness are not significantly different at P =
0.05 (ANOVA with post-hoc Tukey multiple comparison of means).

58

Figure 10 (cont’d).

Surface plots of volume accuracy (VA - mm-3) versus (c) slice thickness (ST - mm) for voltage (V - kV), and (d)
current (C - mA) for voltage (V - kV).

59

2.4.4 HCSR (Y3) and LCD (Y4) optimization
Regression coefficients for both HCSR and LCD (Table 6) indicate that the
individual effects of voltage, current, and slice thickness are statistically
significant (P values ≤ 0.05). No interaction effects between any of the factors
appear to be statistically significant.
Table 4 shows the result of ANOVA for HCSR, in addition it was
determined that the R2 and R2 (adjusted for df) are 98.83 % and 97.32 %,
respectively. This indicates that the factors highly explain the amount of variation
in the observed values of HCSR. Additionally, it is clear that the first, and
polynomial order terms contribute significantly to the model (P value ≤ 0.05), so
the canonical analysis (relationships between factors) in the data set is relevant
(Myers et al., 2008).
Table 4 displays the result of ANOVA for LCD, were the R2 and R2
(adjusted for df) equaled 99.17 % and 98.11 %, respectively. This indicates that
the factors remarkably explain the amount of variation in the observed values of
LCD. Additionally, it is clear that the first, second, and polynomial order terms
contribute significantly to the model (P value ≤ 0.05), so relationships between
factors is relevant (Myers et al., 2008).
The stationary points for both HCSR and LCD are within the experimental
region, and the three eigenvalues are negative, indicating that the stationary point
is a maximum (Table 5); therefore clear evidence of a set of optimal scanning
conditions was inferred. Optimized response for HCSR was established to be at a
60

voltage, current, and slice thickness equal to 113 kV, 170 mA, and 5.0 mm
respectively. Optimized response for LCD was established to be at a voltage,
current, and slice thickness equal to 120 kV, 117 mA, and 4.4 mm respectively.
The surface response plots shown in Fig. 11-12 (a-c) indicate that HCSR and
LCD increase with an increase in all factors (voltage, current, and slice
thickness). Analysis shows that increasing voltage, current, and slice thickness
will result in an increase of HCSR and LCD. This can be visually confirmed in Fig.
13, where the combination of low voltage, current, and slice thickness yield CT
images with low HCSR and LCD, hence CT images have difficulty distinguishing
between two high contrast objects placed close together, and the ability of
discriminating objects that vary slightly from their background is limited
(Goldman, 2007). It could be confirmed that both HCSR and LCD follow a similar
response trend as the SNR, where objects place close together and small lowcontrast objects are obscured by high noise or graininess (Fig. 13) (Goldman,
2007).

61

Figure 11. Surface plots of HCSR versus (a) slice thickness (ST - mm) for current (C - mA), (b) slice thickness (ST
- mm) for voltage (V - kV).

62

Figure 11 (cont’d).

(c) current (C - mA) for voltage (V - kV).

63

Figure 12. Surface plots of LCD versus (a) slice thickness (ST - mm) for current (C - mA), (b) slice thickness (ST mm) for voltage (V - kV).

64

Figure 12 (cont’d).

(c) current (C - mA) for voltage (V - kV).

65

Figure 13. Example of Gray-scale CT image quality for low (-1), medium (0), and high (+1) factor combinations.
Note that low (-1) and high (+1) level combinations are not part of the Box-Behnken design, but CT images were
acquired for visualization and validation purposes.

66

2.4.5 Optimized CT image quality attributes and validation
The study of the effect of independent variables on each of the response
variables individually facilitated the determination of optimized conditions to
acquire maximized quality images. Final optimized parameters for voltage,
current, and slice thickness were set at 120 kV, 170 mA, and 2.5 mm (Table 5).
These values were chosen, because analysis showed that increasing current and
voltage result in a rise of all response variables, including SNR, VA, HCSR, and
LCD. Even though it might seem tempting to exponentially increase voltage, and
current even higher (above experiment upper boundary), it would not be
recommended, due to the fact that it can be predicted that the gain in image
quality will be minimum in comparison with a significant increase in X-ray dose,
and decrease the usable life of the X-ray source, as specified by studies in
medical field (Boone et al., 2003; Woodford et al., 2007), and CT hardware
performance studies (Prokop, 2003; McCollough et al., 2006; Goldman, 2007).
Based on the previously discussed RSM models, it is not recommended to
decrease voltage, and current below 113 kV, and 140 mA respectively, because
a significant loss in image quality, due to a significant increase in CT image noise
would be observed.
In contrast, it could be seen that if slice thickness increases, volume
accuracy will decrease while SNR, HCSR, and LCD increase. Therefore a slice
thickness optimized value of 2.5 mm was chosen, with the objective of finding an
intermediate value, between maximized volume accuracy (slice thickness ≤ 0.625
mm), while keeping high SNR, HCSR, and LCD (slice thickness ≥ 4.4 mm). In

67

addition, setting scanning slice thickness at 0.625 mm or lower, will not only lower
2D CT image quality significantly, by reducing SNR, HCSR, and LCD; but will
increase scanning time (Table 1), a critical factor in the development of fast CT
scanning systems (Hampel et al., 2005; Du et al., 2007; Sellakumar et al., 2007;
Bierberle et al., 2009).
Final experimental optimized response for each image quality attribute;
from the independent CT image validation set, is summarized in Fig. 14. These
experimental values for optimized CT images indicated that the mean SNR, VA,
HCSR, and LCD equaled 0.69 ± 0.25 (SD), 0.43 ± 0.13 mm-3, 0.52 ± 0.03, and
0.74 ± 0.25 respectively. In comparison, the predicted response, using RSM
models (Table 6) and optimized scanning parameters (Table 5) equaled 0.74,
0.32, 0.54, and 0.75 for SNR, VA, HCSR, and LCD respectively (Fig. 14).
Experimental values are in close agreement to predicted values; hence, the RSM
prediction models are accurate.

68

Figure 14. Box-plots showing experimental results for CT image quality
attributes (SNR, VA (mm-3), HCSR, LCD) from a completely independent
validation data set, using optimized scanning parameters, as seen in Table
5. The median of each experimental quality attribute is represented as a
thick horizontal black line, upper and lower quartiles as a box with the
maximum and minimum measurements as lines protruding from these. The
mean experimental response for each quality attributes is symbolized as a
black solid dot (). Predicted optimized response using non-linear response
surface polynomial models (Table 2) is symbolized as a hollow circle ().

It could be visually confirmed in Fig. 13, that SNR is one of the most
important and determinant response variables that will enhance the detection of
ROIs in fresh chestnuts (e.g. Slight decayed tissue vs. healthy tissue in
chestnuts) and is highly related to the HCSR and LCR (Du et al., 2007; Goldman,
69

2007). SNR does not seem to have a relationship to the 3D overall volumetric
information, mainly because the scanned Teflon® reference cylinders (900 HU)
have a high contrast in relation to their background (Air = -1000 HU). Volume
accuracy might play a higher role if the difference in contrast between the
Teflon® reference cylinders and the background wouldn’t have been as high (i.e.
Fresh chestnut tissue). Nonetheless, it was observed that the smaller the slice
thickness the higher the observed volume accuracy.
The HCSR represents the detection of edges of structures and small
foreign objects when a significant difference in contrast exists (Du et al., 2007;
Goldman, 2007), for example to determine the presence of foreign objects in the
chestnuts or void spaces. Resulting optimized 2D CT images showed that the
smallest discernable line pair (lp) is the 1.0 mm bar pattern, which equals a
spatial resolution of 1.5 lp mm-1 (General Electric Company, 2007).
LCD results indicated that optimized CT images provide the ability to
detect structures as small as 3 mm diameter (General Electric Company, 2007).
More specifically for chestnut sorting purposes, this means that the smallest
discernable section of tissue with a slight contrast difference (e.g. decay versus
healthy), will be equal to 3 mm diameter in all directions, due to the intrinsic
isotropic image resolution of the scanner (Tsukagoshi et al., 2007). Therefore, a
section of decay, pellicle, void space, or other region of interest smaller than 3
mm is practically undetectable using optimized CT images.
Relationship between SQR values obtained from a 5-expert panel and the
DQA from their corresponding 2D CT image validation set, obtained through
70

optimized scanning parameters (Fig. 15) demonstrated a significant positive
correlation (R = 0.98, P < 0.05), with a coefficient of determination that accounts
for 96.6 % of total variability explained by linear regression. This indicated a high
match between the digital estimation of decay (DQA) and the human visual
perception of decay (SQR), ultimately reflecting good scanning parameter
optimization/selection.

71

Figure 15. Relationship between Subjective Quality Rating (SQR) values
obtained from a 5-experts panel and the Digital Quality Assessment (DQA)
from their corresponding CT image validation set, obtained through
optimized scanning parameters. Black solid line represents the leastsquare linear regression line (n = 110). The two dashed lines indicate 95 %
CI of the linear regression. The two dotted lines indicate 95 % prediction
intervals of the linear regression.

72

2.5 Discussion
The study applied RSM using Box-Behnken design as a systematic tool to
optimize CT scanning settings (voltage, current, and slice thickness), using CT
images of fresh chestnuts, Teflon® reference cylinders, and a specific QA
phantom, as imaging models for CT image quality (SNR, VA, HCSR, and LCD).
Currently, CT image quality assurance for different application is done by visually
assessing image appearance, in combination with statistical parameters, and
human aided quantitative measurements to validate visual tests, as described by
several scientists including Goldman (2007), Ledenius et al. (2009), Prokop
(2003), Arnold et al. (2010), and companies like General Electric Company
(2007). The issue is that image quality assurance mainly relies in repetitive
subjective measurements of quality, which is time consuming, prone to high
variability, and bias dependent on the human observer capabilities and expertise.
Contrarily, in this study, custom formulas were derived in an attempt to provide a
new automatic and objective way to measure CT image quality by mimicking the
procedure that is usually done by human raters.
If necessary, the same tool can be applied and slightly modified to
evaluate the effect of other determinant scanning factors critical for image quality,
additional scanning response attributes, and further applications that will
significantly advance the field of CT and especially fast CT scanning.
Determinant factors for image quality that were not evaluated in this research,
because of their minor effect in image quality in comparison to the studied factors
can also be studied, like the effect of X-ray detector spacing, image

73

reconstruction algorithms, detector failure, X-ray tube focal spot, and number of
projection images (Bushberg et al., 2002; Goldman, 2007). Other applications of
the proposed tool include the systematic optimization of parameters for micro-CT
system (Badr et al., 1997; Du et al., 2007), clinical applications (Badr et al., 1997;
Huda et al., 2000; Woodford et al., 2007; Wang et al., 2011), X-ray CT dose
reduction (Mayo et al., 1994; Boone et al., 2003; McCollough et al., 2006;
Goldman, 2007; Ledenius et al., 2009), ultra-fast CT parameter optimization
(Hampel et al., 2005; Bierberle et al., 2009), and potential image optimization for
in-line postharvest sorting of agricultural commodities (Barcelon et al., 1999b;
Barcelon et al., 1999a; Sornsrivichai et al., 2000; Butz et al., 2005).
Optimized high-resolution 2D CT images, as obtained from this study, can
accurately reflect subjective internal characteristics of fresh chestnuts with a high
accuracy rate (96.6 %), with high contrast between internal chestnut structures.
This information is useful to study chestnut optimum storage conditions, quality
standards (Mencarelli, 2001), the effect of mechanical harvesting, pre-harvest
treatments (Mandujano et al., 1998; Monarca et al., 2005; Sieber et al., 2007;
Donis-González, 2008; Donis-Gonzalez et al., 2010), fresh chestnut in vivo fruit
morphology for cultivar characterization (Ertan, 2007), and chestnut peelability
(Guyer et al., 2005). In addition, fast image processing, pattern recognition
(Wulfshohn et al., 1993; Kavdir and Guyer, 2007) and feature extraction (Kim and
Schatzki, 2000) applied to more than one 2D CT image per sample, will be a
requirement to develop in-line quality sorting algorithms and systems. Based on
the size of possible discerned structures using optimized CT images, which is
equal to objects larger than 3 mm (founded on LCD results) separated by at least
74

1 mm (established HCSR results) between them, it can be inferred that the
minimum number of acquired images per sample (i.e. ~ 8-17 CT images per
chestnut) are sufficient for the development of future quality-sorting algorithms,
critical to offer premium quality fresh chestnuts.
2.6 Conclusions
The RSM using Box–Behnken design has proved to be a successful
technique to assess the significance of three of the most important CT equipment
scanning parameters (voltage, current, and slice thickness), in combination with a
digital and automatic procedure of measuring CT image quality (SNR, VA, HCSR,
and LCD). It is a reliable method or tool that will yield useful, and high quality CT
images in any field related with CT, without the biased aid of human sight.
Second-order (polynomial) RSM prediction models for each response variable
reflected a combined maximized CT image quality at voltage, current, and slice
thickness equal to 120 kV, 170 mA, and 2.5 mm respectively, for this specific
postharvest application. More specifically, optimized scanning parameters
provided fresh chestnut CT images with high-resolution and high-contrast with a
high accuracy rate (96.6 %) between the digital estimation of decay (DQA), and
human visual perception of decay (SQR). In addition, experimental response
values from optimized CT images are in close agreement to optimized predicted
response values using non-linear response models; hence, the models are
accurate.
Results obtained in this experiment not only offer an advanced technique
that can statistically, routinely, systematically, and reliably optimize CT
75

parameters, but also progress in the study of fresh chestnut quality, as well as
the development of an in-line CT sorter for chestnuts and other agricultural
commodities.

76

CHAPTER 3. RELATION OF COMPUTED TOMOGRAPHY (CT) HOUNSFIELD
UNIT MEASUREMENTS AND INTERNAL COMPONENTS OF FRESH
CHESTNUTS
3.1 Abstract
In this study, a medical grade computed tomography (CT) was used to
obtain XY-plane 2D CT images from decayed and healthy fresh chestnuts, from
the hybrid cultivar ‘Colossal’ and Chinese seedlings. Attenuation coefficients,
referred to as Hounsfield-units (HU) or CT numbers, were acquired from different
2D CT image regions including air, and several chestnut components containing
decayed tissue, healthy tissue, various imperfections such as pellicle invagination
into healthy kernel, and void spaces. Results offer an in vivo accurate insight of
fresh intact chestnuts, and suggest that CT technology is appropriate for in-line
sorting. HU measurements can be used as a nondestructive predictor of fresh
chestnut internal components with a 90.6% overall accuracy rate.
3.2 Introduction
CT methods for accurate visualization, segmentation and inner component
identification of fresh chestnuts, which include the presence of decayed tissue,
pellicle, void spaces and healthy tissue, are not available. It is hypothesize, that
CT technology can be used as a tool to study fresh in vivo chestnut components.
Additionally, with appropriate equipment and classification algorithms, the
information gathered in this study will indicate if the technique is practical and
suitable to apply for in-line sorting systems to accurately determine fresh chestnut

77

quality. For that reason, the objective of this study was to evaluate the HU values
(Section 1.5) from different components of fresh chestnuts and their scanning
environment (air) from images of the hybrid cultivar ‘Colossal’ and from Chinese
chestnut seedling.
3.3 Materials and methods
3.3.1 Sample collection and preparation
A total of 100 kg of physiologically mature chestnuts (C. sativa x C.
crenata) cv. ‘Colossal’ and Chinese chestnut seedlings (C. mollissima), were
obtained from Chestnut Growers Inc. (CGI; Grand Haven, Michigan, USA). These
chestnuts were previously collected from seven commercial farms in Michigan.
Postharvest treatment (dip) was done as described in section 2.3.1. In addition,
physiological internal disorders were stimulated in 25% of the apparently healthy
chestnuts (sinkers) by submerging them for 300 s in 80 C dH2O. This procedure
induces chestnut kernel damage by heat-shock, causing it to subsequently
degrade during storage. Internal microbial decay was artificially induced in
another 25% group of sinkers. This was done by manually injecting 100 L of
Penicillium expansum spore suspension, containing 3.8 x 106 spores L-1, through
the shell and into the chestnut kernel, using a 1 mL sterile medical grade
tuberculin syringe with a 0.21 mm x 15.9 mm needle (BD™, NJ, USA). P.
expansum, used for the experiment, was previously isolated from infected
chestnuts and cultured onto a semi-selective medium containing potato-dextrose-

78

agar (PDA) (BD™), 20 μL L-1 streptomycin (Sigma-Aldrich, Mo, USA) and 50 μL
L-1 ampicilin (Sigma-Aldrich) in a Petri dish. Spores were collected from a two
week old fungal culture, growing at 25 °C. To prepare the inoculum, 15 mL of
sterile dH2O was added to the culture and the spores were gently removed from
the surface with a sterile bacteriological loop. Spore were enumerated with a
hemocytometer and then adjusted to the desired concentration, by dilution with
sterile dH2O. Previous steps provided four groups of chestnuts per species, 1)
Apparently healthy chestnuts (sinkers), 2) Apparently naturally damaged
chestnuts (floaters), 3) Heat-induced-damaged chestnuts, and 4) Penicilliuminoculated chestnuts. The purpose of dividing the ‘Colossal’ and ‘Chinese
seedling’ into four groups was to guarantee a uniform distribution of a diverse
range of internal characteristics needed for the experiment.
All chestnuts from each group were stored in mesh bags at 4 C. After 90
d, 40 chestnuts from each group were randomly picked and attached to two
rectangular wood pegboards (PB1 = 930 mm x 380 mm and PB2 = 1060 mm x
380 mm), in 20 rows containing 8 chestnuts (160 total), using approximately
0.005 kg of 100% transparent silicone per chestnut (General Electric, Waterford,
NY, USA). Each pegboard (PB1 and PB2) contained a unique chestnut species
(CS); PB1 contained Chinese seedlings and PB2 ‘Colossal’ chestnuts.
Additionally, to avoid sample displacement, individual chestnuts were numbered
and whole PBs containing chestnuts were wrapped using a 0.0033 mm stretch
(shrink) polyethylene wrap (Just Packing Supplies Inc., New York, NY, USA) (Fig.
3b). Immediately after, CT scans were conducted.
79

3.3.2 In vivo CT imaging scans
CT scanning was performed on whole PBs, placed on the GE
BrightSpeed™ RT 16 Elite, multi-detector CT instrument (General Electric
Healthcare, Buckinghamshire, United Kingdom), as seen in Fig. 3b. The
experimental conditions (scanning parameters) for the CT observations are
shown in Table 1, using the parameters corresponding to a slice thickness (d) of
0.625 mm. This d was used to have a higher resolution in the Z-axis. A total of
1505 and 1729 XY-plane 2D CT images (slices) from PB1 and PB2 were
obtained, respectively. Fewer images were acquired from PB1, because it was
shorter in length due to the fact that Chinese seedlings are smaller in comparison
to ‘Colossal’ chestnuts. Total scanning time for PB1 equaled 89 s, and PB2 102
s, which corresponds to a scanning time of approximately 0.6 s per chestnut.
3.3.3 Visual based fresh chestnut quality and internal component assessment
Raw color sliced internal faces were color scanned, as described in
section 2.3.3, for record keeping and to use as references to accurately
determine HU values from the XY-plane CT images as described in section 3.3.4.
All chestnuts were then categorized based on their internal disorder severity level
(SL), irrespective of their group; where SL1 contained all chestnuts that are
healthy, SL2 chestnuts which are partially disordered, and SL3 represented the
group of chestnuts which are completely disordered. A representative example
of cross-sectional (XY-plane) CT 2D images in the fresh state, which reflects the
distribution of healthy and disordered tissue among chestnut samples, is shown
in Fig. 16. In these images, healthy - SL1 (e.g. ACT), partially affected (decayed) 80

SL2 (e.g. BCT and DCT), and completely affected by internal disorder - SL3 (e.g.
CCT) chestnuts can be viewed. Parallel, at the right side of every CT image,
freshly sliced raw color images (RGB), which correspond to approximately the
same CT scanned slices, can also be observed (e.g. ARGB, BRGB, CRGB and
DRGB). These images proved useful when judging internal quality, and to
accurately resolve the HU values and ROIs on the CT images. Figure 16-DCT
also exemplifies an image of a partially decayed fresh chestnut, affected by P.
expansum, with an overlying HU-values-profile containing the HU values
acquired at the grey profile-line (PL). As visual reference a HU-standard-bar can
also be noticed.

81

Figure 16. Cross-sectional XY-plane 2D CT images and color raw image
slices (RGB) of chestnuts. (A) shows a healthy chestnut (SL1). (B) shows a
partially decayed chestnut (SL2). (C) shows a completely decayed chestnut
(SL3). (D) shows a partially decayed chestnut (SL2), illustrating the HUvalues-profile taken at the grey line (PL), typifying the HU variation within
components in the same chestnut.
82

3.3.4 HU-value inference using training data set
HU-values from eighteen regions of interest (ROIs) were acquired from
different internal tissues (components) of fresh chestnuts and their scanning
environment (air), including decayed tissue, healthy tissue, pellicle, and void
spaces from XY-plane 2D CT images of the chestnut sample. ROIs were
designed to include all possible combinations where different chestnut
components and air are present, as seen in Table 7.

83

Table 7. Acquired HU value samples from different chestnut components and air, based on regions of interest
(ROIs)
Examples

ROIs

In Fig. 16

CSb

SLc

Region (Categ.e)

Mean HU
valued

ID

SUa

A

SU1

A,B,C CT 5

Ch

-

Air

-1000.94

B

SU1

A,B,C CT 5

Co

-

Air

-1001.15

C

SU2

ACT 3

Ch

SL1

Void space

-830.21

D

SU2

ACT 3

Co

SL1

Void space

-826.64

E

SU2

BCT 3

Ch

SL2

Void space

-836.98

F

SU2

BCT 3

Co

SL2

Void space

-837.58

G

SU2

CCT 3

Ch

SL3

Void space

-825.48

H

SU2

CCT 3

Co

SL3

Void space

-847.88

I

SU1

BCT 4

Ch

SL2

Decayed

-273.35

J

SU1

BCT 4

Co

SL2

Decayed

-293.67

K

SU1

CCT 4

Ch

SL3

Decayed

-299.40

L

SU1

CCT 4

Co

SL3

Decayed

-313.73

M

SU2

ACT 2

Ch

SL1

Pellicle

-118.68

Thresholdf

84

L

U

-1024

-975

-974

-500

-499

-210

-209

24

Table 7. (cont’d)
Examples

ROIs

In Fig. 16

CSb

SLc

Region (Categ.e)

Mean HU
valued

ID
N

SU2

ACT 2

Co

SL1

Pellicle

-137.73

O

SU1

ACT 1

Ch

SL1

Healthy

138.99

P

SU1

ACT 1

Co

SL1

Healthy

115.82

Q

SU1

BCT 1

Ch

SL2

Healthy

125.56

R
a

SUa

SU1

BCT 1

Co

SL2

Healthy

Thresholdf
L
U

112.84

25

300

Sampled Unit (SU) to infer mean HU valued. SU1 = 4 mm2 (6.6 pixels2) square region/repetition (300), SU2 = 1-

pixel/repetition (300). bChestnut species (CS) = cv. ‘Colossal’ – Co, Chinese – Ch. cSeverity level (SL) = SL1, SL2 and

SL3 represents healthy, partially decayed and completely decayed chestnuts, respectively. eCategories were inferred after

performing ANOVA at P = 0.05 and Tukey multiple comparison of means test (Fig. 17). fLower (L) and upper (U)
thresholds for each category estimated from maximum and minimum HU.

85

The first two ROIs included the HU-values of air, defined as the scanning
environment, from different scans. Both ROIs, were acquired to evaluate the
repeatability between CT scans in the 2D CT reconstructed slices and to
determine the intrinsic difference between HU values acquired in a similar
medium. To confirm this, 300 repeated measurements, each including the mean
HU value of a 4 mm2 (6.6 pixel2) square region containing only air, from PB1
(Table 7, ID-A) and PB2 (Table 7, ID-B) at different transverse XY-plane 2D CT
images (slices) were obtained, as exemplified in Fig. 16 -ACT 5, -BCT 5 and -CCT
5. This first step, after acquiring the images was to determine if significant
variability between XY-plane 2D CT images (slices) occurred. This is important to
corroborate because it will indicate if observed differences between HU values
obtained from different ROIs are due to internal changes in physical and/or
chemical characteristics of chestnut samples within evaluated ROIs or if changes
are due to unknown problems in the equipment or image acquisition procedure.
Thereafter, HU values of another sixteen different ROIs from chestnuts
samples were acquired, based on CS, and disorder SL (Table 7). In total, three
hundred random repetitions per measurement from dissimilar XY-plane 2D CT
images were conducted for every ROI type. Depending in the ROIs size, each
measurement either included the mean HU value of a 4 mm2 square region, or
the HU value acquired from an individual pixel. The sampled region for each ROI
can be found in Table 7 under the sampled unit (SU) column. All HU values were
acquired using the Osirix Imaging Software V3.6.1, developed by Dr. Antoine
Russet's Software Team (http://www.osirix-viewer.com/). Due to artifacts that
86

occur near sample transitions (beam hardening effects), care was taken to avoid
sampling near changeover regions, between the different media types.
HU values comparing the ROIs were analyzed using one-factor analysis of
variance (ANOVA). Significance difference between ROI means was determined
using the Tukey post-hoc multiple comparisons of means test at the 95% familywise confidence level (P = 0.05) (Ott and Longnecker, 2001). Calculations were
performed using the language and environment for statistical computing software
R V2.10.0 (http://cran.r-project.org/).
3.3.5 Chestnut categories prediction using an independent testing data set
XY-plane 2D CT images were acquired from a completely different and
independent set of 266 ‘Colossal’ and 266 Chinese chestnuts, with the objective
of testing the accuracy of how well the five categories can be forecasted (blindly
classified), using only HU value measurements. From these CT images, 50
independent HU value random measurements per each of the 5 categories and
per CS were acquired, yielding a total of 500 independent testing data points.
Accuracy refers to the percentage of correct category predictions inferred by
acquired HU value, when compared with the true category label in the testing
data set (Shapiro and Stockman, 2001). For example, a misclassification would
occur if the sampled data point from pellicle tissue (e.g. Fig. 16-ACT 2) (true label)
would yield a HU value equal to -300, automatically and mistakenly assigning this
data point to the decay category, based on HU values thresholds summarized in
Table 7, and further explained in sections 3.4.1 and 3.4.2. Chestnut sample

87

handling, CT scanning imaging parameters and procedure, HU value tissue
sampling, and data handling followed the procedure described in section 3.3.4.
3.4 Results
3.4.1 HU-value and category threshold inference using training data set
CT image acquisition reliability and repeatability was confirmed by
comparing HU values acquired from air, considered as a homogeneous standard
medium. As expected and observed in Fig. 17, HU mean values of air including
ROIs identified as A (Chinese data set) and B (‘Colossal’ data set) are not
significantly different (P < 0.05) and fall under the same category (air). Analysis
supported that under the specified experimental conditions (Table 1), the CT
used in the study, has a HU value SD for a homogeneous standard medium
equal to 1.80 HU. In other words, the equipment will be able to accurately
discern between tissues only when the HU value difference is higher than 1.8
HU.

88

‘

‘

Figure 17. Box-plots showing the HU values from ROIs obtained using a training data set. The median is
represented as a thick horizontal black line, upper and lower quartiles as a box with the maximum and minimum
measurements as lines protruding from these. Box-plots followed by the same letter and enclosed by the same
rectangle are not significantly different between each other at P = 0.05 (ANOVA) (Tukey multiple comparison of
means). Solid circles joined by a dashed black line (---) show the relationship between the mean of the HU values
for each ROIs and subsequently each category.

89

Fig. 17 offers a clear outline of the changes and variation of HU values
associated with changes in internal characteristics within samples and between
different SL. Supporting these observations, data showed that positive values,
higher than 24.89 HU (density > 100.0 kg m-3), could be observed in areas
containing healthy tissue. Negative HU-values, between -2.52 and -1001.08 HU
(density < 100.0 kg m-3), correspond to the existence of pellicle, decayed and
void spaces. Similar results were described in other studies, where moisture
content and density changes in fruits (Barcelon et al., 1999b; Barcelon et al.,
1999a) and wood (Fromm et al., 2001), were reflected in significant HU value
differences.
Regions of interest that are not statistically different can be grouped in the
same categorical variables as seen in the boxes around the ROIs in Fig. 17 and
summarized in Table 7. Information from Fig. 17 was used to generate
preliminary thresholds separating each neighboring category. These thresholds
were developed by taking in account the minimum and maximum HU values for
each category. ROIs followed by the same letter and enclosed by the same
rectangle in Fig. 17 are encompassed in the same category since they are not
significantly different between each other at P = 0.05. Fig. 17 also includes solid
black dots located in the mean HU value per category, joined by a dashed line (--) to observe the change in HU values in between categories.
Analyzing the statistical difference among ROIs in Fig. 17 revealed five
categories, which separately categorize sampled ROIs containing air, void
spaces, decayed tissue, embedded pellicle and healthy tissue as summarized in
90

Table 7. Significant changes, spread and high variability of HU values, based on
the different ROIs can also be observed in Fig. 17, which indicated an increase in
HU value across the categories. From Fig. 17, it is shown that the HU values
obtained from the ROIs are independent of CS and SL, due to the fact that only
the sampled region, which reflects changes on internal characteristics or chestnut
internal components, determines the changes in categories. Seeing that it has
been repetitively demonstrated that HU values are highly correlated to matter
bulk density, as reported by several scientists, including Fromm et al. (2001) and
Bushberg et al. (2002), it can then be confirmed that the significant change in HU
value means between the categories, is due to the changes in tissue density
within the chestnut. Even though this result seems obvious, because of the
principals of CT X-ray technology, it is important to mention that it indicates that
the HU values can be used as an index to segregate and group chestnuts, not
only based on their internal characteristics or components, but also because the
technology is sensitive enough to accurately quantify chestnut kernel density
changes. Density is considered an important quality parameter for chestnut as
well as for other commodities, as previously expressed by Barcelon et al.
(1999a). By rapidly and accurately measuring commodity density, some other
problems might also be predicted such as insect damage, mechanical injuries,
and firmness.
By observing Fig. 17, it can be concluded that physical and/or chemical
changes in the chestnut kernel, some caused by internal disorders and others by
the presence of void spaces and pellicle embedded in the healthy kernel tissue,
result in the HU values declining in relationship with degraded tissue health. It
91

can also be seen that by comparing between categories and evaluating the mean
HU value in combination with the total HU value range for each of the categories,
a threshold can be set to accurately separate healthy tissue from pellicle as well
as healthy tissue from decayed or disordered tissue. Pellicle can somewhat be
separated from decayed tissue but can accurately be segregated from voidspaces and air. Decayed tissue can additionally be separated from both voidspaces and air. Finally, no well-defined threshold can be specified between air
and void-spaces.
3.4.2 Chestnut categories prediction using an independent testing data set
Threshold value testing, summarized in Fig. 18 was useful to determine
how well the predicted HU values can forecast chestnut components. Fig. 18a
shows the prediction of the categories from an independent testing HU values
data set, using the estimated thresholds. The selected HU value thresholds,
which include lower (L) and upper (U) HU values for all the categories, can also
be found in Table 7. Figure 18b includes the confusion matrix corresponding to
this prediction procedure. Each column of the matrix represents the occasions in
a predicted category, while each row represents the instances in an actual
category. This matrix is beneficial, because it visibly documents the misclassified
categories (Shapiro and Stockman, 2001). Additionally, Fig. 18c shows
representative images, containing a labeled image generated from its
corresponding CT image, based on HU value thresholds. In Fig. 18c, color
images are also included as a quality reference. Labeled images were
automatically generated using the region-growing tool in Osirix Imaging Software
92

V3.6.1 (http://www.osirix-viewer.com/). This tool is used in the medical field to
isolate or label structures in a CT image automatically after the user has indicated
lower and upper HU value bounds that will be labeled, as specified by Wang and
Smedby (2010). As an example, the healthy tissue is yellow labeled (white in
gray scale image), after setting the lower bound at 25 HU and the higher bound
at 300 HU. All other categories are comparably labeled, with respective threshold
intervals and varying colors.
Results from Figs. 18a and 18b, show that the overall accuracy rate for all
the categories prediction, using independent testing data is equal to 90.6 %. The
highest accuracy rates were obtained from the prediction of the air category, and
the healthy category, where 100.0 % of the samples can be separated from all of
the other categories. Void spaces can be separated from the other categories
with a 96.0 % accuracy rate, because 4.0 % of the data points were recognized
as air. The lowest accuracy rate is present when trying to differentiate between
decayed tissue and pellicle, with an average accuracy value equal to 78.6 %.
Here, 20.0 % of the decayed tissue was mistakenly predicted as pellicle, and
23.0 % of the pellicle was wrongly categorized as decayed tissue.
Fig. 18c, indicated that regardless of the SL, XY-plane 2D CT images
could be used to accurately categorize chestnut components and air, based on
selected thresholds. Positively, decayed tissue is always isolated from healthy
tissue and both categories are never misclassified. On the other hand, if closely
observed, it can be seen that slight misclassification can be observed between
decayed tissue and pellicle, as well as void spaces and air, especially if
categories are mixed within each other, as seen in Fig. 18c(SL2) and 18c(SL3).
93

In addition, Fig. 18c(SL1) clearly states that certain tissues could be
categorized incorrectly (healthy tissue as pellicle), due to artifacts and noise that
occur near category transitions; broadly known as beam hardening effects
(Bushberg et al., 2002).

94

Figure 18. (a) Category prediction using estimated thresholds (horizontal black lines) with an independent testing
data set. HU threshold values are indicated above each threshold line. The selected HU value threshold range for
air, void spaces, decay tissue, pellicle tissue, and healthy tissue are -1024 to -975, -974 to -500, -400 to -210, -209
to 24, and 25 to 300; respectively.

95

Figure 18 (cont’d)

(b) Confusion matrix corresponding to the category prediction accuracy (overall accuracy rate = 90.6%). (c)
Representative images, containing automatically generated color labeled images from their corresponding CT
images, based on HU-value threshold intervals. RGB raw image slices are also included as a quality reference.

96

3.5 Discussion
Results show that CT technology can be used as a technique that will be
able to visualize and measure macroscopic changes in chestnuts components.
By doing this, the technique might be able to detect possible negative effects of
mechanical harvesting and pre-harvest treatments, cultivar characterization and
morphology (chestnut peelability), physiological changes during storage, and
early pathogen development. In addition, the data presented in this study is
essential for developing classification algorithms to sort chestnuts based on their
internal characteristics. HU values obtained from XY-plane 2D CT images are
used as a reference to determine the presence of internal components of
chestnuts, which will be useful for developing future prediction models of chestnut
internal quality. Nonetheless, results clearly indicated, that in addition to raw HU
values, other methods related to image processing, pattern recognition
(Wulfshohn et al., 1993; Kavdir and Guyer, 2007), and feature extraction (Kim
and Schatzki, 2000), will be a requirement to aid the development of future
sorting algorithms, and will be necessary to accurately separate void-spaces from
air as well as pellicle from decayed tissue. Transferring this tool to different
commodity industries will enable them to control and promote the quality of their
products. In addition, for the chestnut and other fruits, vegetables, and nuts, this
application will provide the opportunity to select products according to their final
use.
Further research is needed, to develop sorting algorithms (see Chapter 4)
and to investigate optimal equipment settings for real-time sorting potential, as
97

well as determine hardware capabilities, sorting speed, and system cost
considerations (see Chapter 5).
3.6 Conclusion
CT imaging provides high-resolution and high-contrast images of the
internal structure and components of fresh chestnuts. Internal component
prediction accuracy equal to 90.6% was achieved by using HU values as
predictors. In future studies, sorting algorithm speed, equipment cost and
characteristics, as well as other methods related to image processing, feature
extraction and pattern recognition will be useful for the development of reliable
sorting algorithms and CT sorting systems. More studies must be pursued to
evaluate the accuracy of sorting algorithms and in-line classification (Chapter 4).
The main advantage of using HU values with the aid of other algorithms
will be that CT is a fast, non-invasive procedure that has the potential to be
adapted for quality evaluation and in-line sorting. In general, this study indicated
that CT has a high potential for nondestructively evaluating internal components
of fresh chestnuts, which are attributes associated with chestnut kernel quality.

98

CHAPTER 4. POSTHARVEST NONINVASIVE ASSESSMENT OF FRESH
CHESTNUT INTERNAL DECAY USING COMPUTED TOMOGRAPHY (CT)
IMAGES
4.1 Abstract
X-ray CT is an effective noninvasive tool to visualize fresh agricultural
commodities’ internal components and quality attributes, including those of
chestnuts. There is no reported procedure to automatically, effectively and
efficiently classify fresh commodities from a continuous in-line flow through a CT
system. If the information obtained by CT scanning of fresh agricultural
commodities is to be used in an industrial application (e.g. in-line sorting),
automated interpretation of CT images is essential. For this purpose, an image
analysis method (algorithm) for the automatic classification of CT images
obtained from 2848 fresh chestnuts (cv. ‘Colossal’ and Chinese seedlings),
during the harvesting years from 2009 to 2012, was developed and tested.
Classification accuracy was evaluated by comparing the classes obtained from
six secondary CT images per chestnut, derived from raw CT images, to their
internal quality assessment. Internal quality assessment was done by an
experienced human rater by visually and invasively rating fresh chestnut internal
decay severity (quality) into 5-, 3- and 2- classes.
After CT image preprocessing, cropping and segmentation, 1194 grayscale
intensity and textural features were extracted from the six resultant CT images
per sample. Relevant features were selected using a sequential forward selection
algorithm with the Fisher discriminant objective function. 86, 155 and 126
99

features were effective in designing a quadratic discriminant classifier with a 4fold training for the 3 different class scenarios. Performance accuracy of 85.9 %,
91.2 % and 96.1 % for 5, 3 and 2 classes was found, respectively. Results show
that this method is accurate, reliable, and objective in determining fresh chestnut
internal quality, and it is applicable to an automatic noninvasive in-line CT sorting
system.
4.2 Introduction
Researchers agree that the absence of sorting commodities in industry
using in-line CT systems is primarily due to their limitations in gathering useful
information, and optimum data handling, when operating at high-speed (Butz et
al., 2005). More specifically, if CT in-line systems were to be developed, little is
currently known about how to efficiently handle the high amount of acquired data,
while continuously scanning. In addition, there is no available procedure reported
to automatically, effectively and efficiently classify fresh agricultural commodities
from a continuous in-line flow of entities going through a CT scanning system.
Pattern recognition algorithms, which are an intrinsic and an important part
of computer vision systems, offer a mechanism of classifying commodities based
on their quality attributes, and can be applied to CT systems. In general, systems
are trained from patterns of interest extracted from a set of images, representing
different quality categories. A pattern or feature is represented by a group of
textural and image intensity features, which are able to define all of the quality
categories. The system then assigns a new image series to a specific quality
category or class (Duda et al., 2000). The first step consists of extracting a high
100

number of features (patterns) from the different categories of known images.
Features then must be selected by their capacity of correctly separating the
images into different categories, therefore training the system, and allowing it to
automatically classify new images. Classification is done using statistical and
clustering algorithms by assigning images per sample, to its corresponding class
(Duda et al., 2000; Mery and Soto, 2008). The reader may refer to complete
information regarding statistical pattern recognition methods, which have been
extensively described in several publications, including Jain et al. (2000), Duda et
al. (2000), Bishop (2007), and Holmström and Koistinen (2010). Steps used to
generate the pattern classification algorithm to categorize chestnut quality using
CT images are illustrated in Fig. 19, and in depth explained in the materials and
methods section.
Therefore, the aim of this study is to describe a reliable method that will
systematically, rapidly, and efficiently classify a fresh agricultural commodity (i.e.
chestnuts) using CT images, obtained from a commercially available medical
grade CT system. This classification method is unique because it automatically
uses a set of CT images to determine the overall quality of fresh chestnuts,
without the aid of a human sorter. This research offers a novel tool, which can
methodically classify chestnuts from high-throughput 2D CT images. In addition,
the developed algorithm can be used to study the potential development of an inline CT system, and can be accessed by scientists to automatically and efficiently
classify CT acquired images, in other domains. This automatic classification
algorithm is a critical step in the development of a fast, in-line, nondestructive CT
system,

capable

of

assessing

fresh
101

internal

quality

attributes

and components of individual chestnuts, other agricultural commodities, in
addition to other objects of interest (e.g. airplane parts).

102

Figure 19. Procedure used to generate the pattern classification algorithm
to categorize chestnut quality using CT images.
103

4.3 Materials and methods
4.3.1 Sample collection and preparation
A total of 1424 physiologically mature Chinese seedlings (C. mollissima)
and 1424 (C. sativa x C. crenata) cv. ‘Colossal’ chestnuts (Total of 2848) were
obtained from Chestnut Growers Inc. (CGI; Grand Haven, Michigan, USA).
Chestnuts were collected from seven commercial farms in Michigan. Equal
numbers of chestnuts were collected every year, from the 2009 through the 2012
harvesting season (September-October). Following common postharvest
procedures samples were treated, after each collection year, as described in
section 2.3.1. Samples were then handled as described in section 3.3.1, with the
objective of attempting to generate a uniform distribution of a diverse range of
internal characteristics (different classes of decay) needed for this study. Every
year, after approximately 90 storage days (4 C), chestnuts were similarly
attached (facing toward the same direction) to rectangular polyethylene boards
(915 mm x 335 mm x 2.8 mm), in 19 rows containing 7 chestnuts (133 total), as
described in section 3.3.1 and seen in Fig. 3b. CT scans were then conducted on
each board following the procedure described in section 3.3.2, only that in this
case, instead of using the parameters corresponding to a slice thickness (d) of
0.625 mm (visualization purposes), a d equal to 2.5 mm was used, as concluded
during the CT image optimization procedure in Chapter 1.

104

4.3.2 CT image preprocessing:
Image preprocessing (re-slicing, cropping and contrast enhancement),
image visualization, segmentation, feature extraction, statistical analysis, and the
automatic classification/validation for this study were done in MATLAB (2012a,
The MathWorks, Natrick, MA, USA) (http://www.mathworks.com), and in the
language and environment for statistical computing software R (V2.10.0, R
Development Core Team, Vienna, Austria) (http://cran.r-project.org/), using a
Macintosh environment on a Lion operating system with 2.53 GHz Intel Core 2Duo, 8 GB random access memory (RAM), 1067 MHz double data rate 3 (DDR3)
(Apple Inc., Cupertino, California, USA). Feature extraction, feature reduction,
and the automatic classification/validation for this study were performed by
partially

using

the

“Balu”

free

(http://dmery.ing.puc.cl/index.php/balu/),

toolbox
developed

for
by

pattern
the

recognition

Department

of

Computer Science at the Pontifical Catholic University of Chile (Santiago, Chile).
This toolbox contains more than 200 functions for image processing, feature
extraction,

feature

transformation,

feature

analysis,

feature

selection,

classification, clustering, performance evaluation, image sequence processing,
and more.
4.3.3 CT image re-slicing
Because of the intrinsic 3D characteristic of CT imaging, CT image slices
were re-sliced following the steps described in section 1.5 (Fig. 5). Depending on
chestnut physical size and d, each chestnut contains between 8 to17 XY-, YZ-

105

and XZ-plane-slices representing virtual cross-sections of a chestnut along the Z,
X and Y axes.
4.3.4 Individual chestnut CT image cropping
Chestnut rows and individual chestnuts were visually/manually cropped
from the overall CT data set containing the scanning table, volume of air and
other chestnuts, by determining their spatial location, as shown in Fig. 20. Pixels
in Fig. 20a (entire board XZ-plane-slice) correspond to the mean intensity value
of all pixels at the same planer location (x,z) in the XZ-plane-slices stack. In Fig.
20a, Z1- and Z2-spatial-location-values are manually determined to crop each
chestnut row, generating Fig. 20b (row XZ-plane-slice), 20c (row XY-plane-slice)
and 20d (row YZ-plane-slice). To crop each individual chestnut, X1-, X2-, Y1-, and
Y2-spatial-location-values are manually inferred, as seen in Fig. 20a, 20b, 20c
and 20d. Pixels in Figs. 20a-20d, correspond to the mean of all pixel values at the
same planer location ((x,y), (y,z) or (z,x)) within each plane in the chestnut stack.
For each chestnut that is re-sliced making up the three different image
stacks per plane, a total data set of approximately 50 raw CT image slices of
about 50 x 50 pixels each (depending on chestnut physical size) are generated.
For further analysis, data set dimensionality is then reduced from these original
50 raw CT images per chestnut sample to 6 resultant CT images per sample
(secondary CT image set – Fig. 19). To accomplish this reduction, pixels in final
cropped chestnuts, as seen in Fig. 20e (mean XZ-plane-slice), 20f (mean XYplane-slice) and 20g (mean YZ-plane-slice) are generated from the mean
106

intensity value of all pixel values at the same planer location ((x,y), (y,z) or (z,x))
in all of the XZ-, XY- or YZ slices per chestnut.

107

Figure 20. Mean intensity values of pixels in all of the: (a) YZ-plane-slice of the whole board containing chestnuts,
(b) YZ-plane-slices for the first chestnut row, (c) XY-plane-slices for the first chestnut row, (d) XZ-plane-slices for
the first chestnut row. Mean intensity values of pixels in final secondary CT cropped chestnut for the (e) YZ-, (f)
XY- and (g) XZ-plane-slices. Maximum intensity values of pixels in final secondary CT cropped chestnut for the
(h) YZ-, (i) XY- and (j) XZ-plane-slices. The beginning of the first (Z1-), the end (Z2-) of the row, the left (X1-), the
right (X2-) side, the bottom (Y1-) and the top (Y2-) of the first chestnut -spatial-location-values are shown in red.

108

An additional example of a resultant mean CT image slice, viewed from the
XZ-plane can also be seen in Fig. 21b. In addition, cropped chestnuts are also
used to generate the second 3 resultant images based on the maximum intensity
value of all pixel values at the same planer location ((x,y), (y,z) or (z,x)) in all of
the reconditioned XZ-, XY- and YZ-plane-slices per chestnut, as shown in Figs.
20h (maximum XZ-plane-slice), 20i (maximum XY-plane-slice), and 20j
(maximum YZ-plane-slice). An example of another resultant maximum CT image
slice, viewed from the XZ-plane is also included in Fig. 21f. Therefore, three
mean and three maximum intensity value CT images from the three different
planes (XY, XZ and YZ) are obtained per chestnut sample (total of 6).

109

Figure 21. (a) Partially decayed fresh raw chestnut slices. (b) Original secondary mean CT image (YZ-plane-slice),
(c) adjusted secondary mean CT image (YZ-plane-slice), (d) final contrast enhanced secondary mean CT image
(YZ-plane-slice). (e) Segmented CT image from final contrast enhanced secondary mean CT image (YZ-planeslice) (binary mask). (f) Original secondary maximum CT image (YZ-plane-slice), (g) adjusted maximum
secondary maximum CT image (YZ-plane-slice), and (h) final contrast enhanced secondary maximum CT image
(YZ-plane-slice).

110

4.3.5 Contrast enhancement
The most important characteristic in CT X-ray images is image contrast.
Image contrast is the difference in luminance of the grayscale intensity values
(separation between the darkest and brightest regions of an image) that makes
the representation of different structures in a CT image discernable (Wang et al.,
1983). This includes the different tissue types reflected in chestnut CT images
(Donis-Gonzalez et al., 2012). Contrast enhancement is a vital step in image
processing, and it is done to increase image quality (Zimmerman et al., 1998;
Jagannath et al., 2012). Multiple techniques exist, which are used to improve
digital image contrast, including morphological operations enhancement, as
described in Sreedhar and Panlal (2012).
In this study, a set of image preprocessing steps were implemented to
increase secondary CT image contrast, based on morphological enhancement as
described in Wirth et al. (2004). First, image contrast is improved in each of the
original secondary CT images (Fig. 21b and 21f) by creating an adjusted mean
and maximum secondary CT image as exemplified in Fig. 21c and 21g,
respectively. This first step is done by mapping the original outputted secondary
CT image grayscale values (Fig. 21b and 21f), so that 1 % of their pixels at low
and high intensities (2 % in total) are saturated (limited to lowest and maximum
pixel values). Second, a top-hat operation (high grayscale intensity regions) using
a 5-neighbors disk shaped flat structuring elements (SE) is applied to the
adjusted secondary CT images (Fig. 21c and 21g), generating the top-hat image.
Third, a bottom-hat operation (low grayscale intensity areas) using the same SE

111

as described in previous step is also applied to the secondary CT adjusted
images (Fig. 21c and 21g), creating the bottom-hat image. The form and size of
the SE is appropriate because of the natural shape of the chestnuts, reflected in
the CT images. Finally, contrast enhanced secondary CT images are produced
as illustrated in Fig. 21d (mean) and Fig. 21h (maximum) by adding the top-hat
image to its corresponding adjusted image and subtracting its linked bottom-hat
image (i.e. adjusted image + top-hat image – bottom-hat image). This stretches
the high intensity areas toward increased intensity, while low intensity regions are
stretched towards decreased intensity.
4.3.6 CT image segmentation (Binary mask):
Image segmentation is implemented to recognize the region of interest in
an image, which in this case is the chestnut in each CT image. CT image
segmentation was done by using the balanced histogram thresholding method,
as described in Anjos and Shahbazkia (2008). This is a simple histogram based
thresholding procedure broadly used for automatic image threshold. This
methodology assumes that the CT image is divided in two main classes, (1) the
foreground (chestnut CT image) and (2) the background (air). To do so, the
methodology finds the optimum threshold level (first minimum grayscale value in
mean CT image histogram) from a CT mean slice, which will divide the image
into the two classes. Example of a segmented image for one plane, using this
routine, can be found in Fig. 21f.

112

4.3.7 Visual based fresh chestnut quality and internal component assessment:
Raw color sliced internal faces were color scanned, as described in
section 2.3.3, for record keeping. All chestnuts were then qualitatively assessed
for disorders, void spaces, and embedded pellicle. All chestnuts were then
categorized based on their internal disorder severity level into five, three and two
categorical groups as described in section 4.1.8 and visualized in Figs. 23, 24,
and 25, respectively. An additional representative example of cross-sectional raw
color images in the fresh state is shown in Fig. 21a. In these images, fresh slices
of a partially decayed chestnut can be viewed.
4.3.8 Feature extraction:
In general, feature extraction involves algorithms that detect and isolate
desired information from segmented images, including CT images (Nixon and
Aguado, 2008). The idea is to use the extracted features to decide to which
quality class does a chestnut belong. The type of extracted features play an
intrinsic and very important role in classifying images appropriately (Jain et al.,
2000).
In this study, features were extracted from the six contrast enhanced 16-bit
secondary CT intensity images per chestnut. A total of 199 features were
extracted per secondary/resultant CT image, and then features from the six
resultant CT images were concatenated to form a feature vector (x) with 1194
components, as partially illustrated in Fig. 19. Extracted features per CT image
included: (1) 6 basic intensity features, (2) 26 Haralick textural (Tx) features, (3)

113

95 intensity local binary pattern (LBP) features, (4) 67 Gabor intensity textural
features, and (5) 5 contrast features.
 Basic intensity features:
Six standard features, describing overall image intensity information, were
derived from the segmented image region, for each of the contrast enhanced
secondary CT grayscale images. Standard features included the mean (𝜇) – Eq.
4-1, standard deviation (𝜎) – Eq. 4-2, kurtosis (k) – Eq. 4-3, skewness (s) – Eq. 44, mean gradient [first-order derivative (
Laplacian [second-order derivative (

𝑓 ′ (𝑥(𝑖,𝑗) ))]

𝑓 ′′ (𝑥(𝑖,𝑗) ))]

– Eq. 4-5, and mean

– Eq. 4-6 (Shapiro and

Stockman, 2001; Nixon and Aguado, 2008; Mery et al., 2011).
𝑛

1
𝜇 = ∑ 𝑥(𝑖,𝑗)
𝑛

(4-1)

𝑖=1

𝑛

1
2
𝜎 = √ ∑(𝑥(𝑖,𝑗) − 𝑥̅ )
𝑛

(4-2)

4
1 𝑛
∑ 𝑖=1(𝑥(𝑖,𝑗) − 𝑥̅ )
𝑘= 𝑛
2
2
1 𝑛
( 𝑛 ∑ 𝑖=1(𝑥(𝑖,𝑗) − 𝑥̅ ) )

(4-3)

𝑖=1

114

3
1 𝑛
∑ 𝑖=1(𝑥(𝑖,𝑗) − 𝑥̅ )
𝑠= 𝑛
3
2
1 𝑛
( 𝑛 ∑ 𝑖=1(𝑥(𝑖,𝑗) − 𝑥̅ ) )

𝑓 ′ (𝑥 𝑖𝑗 ) =

(4-4)

𝑥(𝑖,𝑗)+1 − 𝑥(𝑖,𝑗)
(( 𝑖, 𝑗) + 1) − (𝑖, 𝑗)

(4-5)

𝑓 ′ (𝑥(𝑖,𝑗) ) − 𝑓 ′ (𝑥(𝑖,𝑗)−1 )
𝑓 ′ ′(𝑥 𝑖𝑗 ) =
(( 𝑖, 𝑗) + 1) − (𝑖, 𝑗)
where,

(4-6)

𝑥(𝑖,𝑗) represents the grayscale value of pixels in each of the segmented

CT images, and n the total number of evaluated pixels (i x j).
 Haralick textural (Tx) features:
Twenty six Tx features were extracted from each of the contrast enhanced
secondary CT grayscale images in order to obtain information about their
intensity values distribution. To obtain these Tx features, first a co-occurrence
matrix (C) (Eq. 4-7) is computed per CT image, which represents the joint
probability distribution of intensity pairs of neighboring pixels.

𝐶Δx,Δy (n,m)
𝑖

𝑗

1, if 𝐼( 𝑖, 𝑗) = Ng and 𝐼( 𝑖 + Δ𝑥, 𝑗 + Δ𝑦) = Ng
= ∑∑{
0, otherwise

(4-7)

𝑖=1 𝑗=1

where, C is defined over an i x j CT image (I), parameterized by an offset (Δx,Δy).
This matrix is square with dimensions n (row) x m (column) (Eq. 4-8),

115

represented by the number of gray intensity level in the CT image (N g), which is
equal to 216 (65536) in a 16-bit CT image. As it can also be seen in Eq. 4-7, an
element (n,m) of C is generated by counting the number of times a Ng in the ipixel is adjacent to an Ng in the j-pixel. Since, adjacency in a 2D image can occur
in four directions or offsets - (Δx,Δy) (horizontal, vertical, left and right diagonals)
four C-matrices can be calculated.

𝑐(1,1) ⋯
⋮
⋱
𝐶=[
𝑐(𝑁 𝑔 , 1) ⋯

𝑐(1, 𝑁 𝑔 )
⋮
]
𝑐(𝑁 𝑔 , 𝑁 𝑔 )

(4-8)

Subsequently, basic Tx statistical features as described in Eqs. 4-9 – 4-23
were computed (Table 8). Both mean and range values for each of the Tx
features, for a mask containing five different neighboring pixels, were processed.
Since rotation invariance is an important criterion for features extracted from the
CT images, invariance was accomplished for each of these statistics, by
calculating the mean over the four directional C-matrices (4 offsets).

116

Table 8. Haralick textural (Tx) features (Haralick, 1979)
Tx feature
Formula
1. Angular
Second
Moment
2. Contrast

∑ ∑ 𝐶(𝑛, 𝑚)2
𝑛
𝑚
𝑁 𝑔 −1

𝑁𝑔

(4-9)

𝑁𝑔

∑ 𝑖 2 {∑ ∑ 𝐶 ( 𝑛, 𝑚)} , | 𝑛 − 𝑚| = 𝑖
𝑖=0
3.
Correlation

𝑛=1 𝑚=1

∑ 𝑛 ∑ 𝑚( 𝑛𝑚) 𝐶 ( 𝑛, 𝑚) − μ 𝑥 μ 𝑦
𝜎𝑥 𝜎𝑦
μ 𝑥 , μ 𝑦 , 𝜎 𝑥 and 𝜎 𝑦 are the means and
standard deviation of 𝑐 𝑥 and 𝑐 𝑦 , the partial probability
where

4. Sum of
Squares:
Variance
5. Inverse
difference
moment
6. Average
sum

density functions.

∑ ∑( 𝑛 − μ) 𝐶(𝑛, 𝑚)
𝑛

𝑚

∑∑
𝑛
2𝑁 𝑔

𝑚

(4-12)

1
𝐶(𝑛, 𝑚)
1 + (𝑛 − 𝑚)2

(4-13)

∑ 𝑖𝐶 𝑛+𝑚 (𝑛)
𝐶 𝑛+𝑚 ( 𝑖) is

8.
Variance
sum

(4-11)

2

𝑛=2
where n (row) and m (column) are the coordinates of
an entry in the C, as described in Section 2.6.2.

7. Entropy
sum (𝑓7 )

(4-10)

the
summing to n + m.
2𝑁 𝑔

probability

of

− ∑ 𝐶 𝑛+𝑚 ( 𝑛) log{ 𝐶 𝑛+𝑚 ( 𝑛)}

C

(4-14)

coordinates

(4-15)

𝑛=2
2𝑁 𝑔

∑ ( 𝑖 − 𝑓7 )2 𝐶 𝑛+𝑚 (𝑛)
𝑛=2

117

(4-16)

Table 8. (cont’d)
Formula

Tx feature
9. Entropy

(𝐻𝑋𝑌)

10.
Variance
difference
11. Entropy
difference

− ∑ ∑ 𝐶 ( 𝑛, 𝑚) log(𝐶 ( 𝑛, 𝑚))
𝑛

𝑚

(4-17)

𝑁 𝑔 −1

∑ 𝑛2 𝐶 𝑛−𝑚 (𝑖)

(4-18)

𝑛=0
𝑁 𝑔−1

− ∑ 𝑝 𝑛−𝑚 ( 𝑛) log{ 𝑝 𝑛−𝑚 ( 𝑛)}

(4-19)

𝑛=0
12.
Correlation
meas. - 1
13.
Correlation
meas. - 2

𝐻𝑋𝑌 − 𝐻𝑋𝑌1
max{ 𝐻𝑋, 𝐻𝑌}

(4-20)

(1 − exp[−2( 𝐻𝑋𝑌2 −

1
𝐻𝑋𝑌)])2

where HX and HY are the entropies of

𝐶 𝑛 and 𝐶 𝑚 .

(4-21)

𝐻𝑋𝑌1
= − ∑ ∑ 𝐶 ( 𝑛, 𝑚) log{ 𝐶 𝑛 ( 𝑛) 𝐶 𝑚 (𝑚)}
𝑛

(4-22)

𝑚

𝐻𝑋𝑌2
= − ∑ ∑ 𝐶 𝑛 ( 𝑛) 𝐶 𝑚 (𝑚) log{ 𝐶 𝑛 ( 𝑛) 𝐶 𝑚 (𝑚)}(4-23)
𝑛

𝑚

 Intensity local binary pattern (LBP) textural features:
Ninety-five LBP features (classic and semantic) were extracted from each
of the contrast enhanced secondary grayscale CT images to compute the
relationship between the intensity of each pixel with its eight neighboring pixels,
using the occurrence histogram (classic – cLBP), and a histogram clustered by
118

the pixel intensity value (semantic – sLBP). LBP approach introduces an operator
where a neighborhood of up to 36 pixels for the cLBP, and 56 pixels for the sLBP
of the images are thresholded in relation to the center pixel value, forming a new
binary sub-image (Ojala et al., 2002; Ahonen et al., 2009; Chai et al., 2013) (Eq.
4-24, Fig. 22). The rotations of the selected neighborhoods inside each image
increase the number of LBP features, giving a rotation invariant reliable
characterization of the image. The values of the pixels in the thresholded
neighborhood are multiplied by the binomial weights given to the corresponding
pixels. Finally, the values of the eight pixels are summed to obtain the LBP
(Pietikäinen et al., 2000) (Eq. 4-25). LBP are defined as:

𝑠( 𝑁𝑔0 ∙ 𝑁𝑔 𝑖 ) = {

1, 𝑖𝑓 𝑁𝑔 𝑖 > 𝑁𝑔0
0, 𝑖𝑓 𝑁𝑔 𝑖 ≤ 𝑁𝑔0

(4-24)

8

𝐿𝐵𝑃(𝑑,ℎ) = ∑ 𝑠( 𝑁𝑔0 ∙ 𝑁𝑔 𝑖 )2 𝑖−1

(4-25)

𝑖=1
where, 𝑁𝑔0 is the gray value of the center pixel in the circularly symmetric
neighborhood, and

𝑁𝑔 𝑖

takes the different eight pixels values from the

neighborhood. LBPs have been found to be influential features for texture
classification in images (Ojala et al., 2002; Ahonen et al., 2009). Example of
cLBP images with different neighborhood size operations for a maximum intensity
XY-plane-slice CT image of a partially decayed and a healthy chestnut can be
can be found in Fig. 22.

119

 Intensity Gabor textural features:
Sixty-seven Gabor features were extracted per contrast enhanced
secondary CT grayscale images. The general Gabor function is complex,
exponential and modulated by a Gaussian envelope function (Kumar and Pang,
2002).

𝑓( 𝑥, 𝑦)
1 𝑥2
𝑦2
=
exp (− ( 2 + 2 )) exp(2𝜋𝑗𝑢0 𝑥 )
2 𝜎𝑥
𝜎𝑦
(2𝜋𝜎 𝑥 𝜎 𝑦 )
1

(4-26)

where, 𝜎 𝑥 and 𝜎 𝑦 denote the Gaussian envelope along the 𝑥 and 𝑦 axes, and 𝑢0
defines the radial frequency. In the frequency domain, the Gabor function acts as
a multi-scale and multi-orientation band pass filter with a real and an imaginary
component, using the Gaussian function. The self-similar filter banks can be
obtained by dilations and rotation of

𝑓( 𝑥, 𝑦) through the generating function, as

seen in Eqs. 4-27 – 4-29:

𝑓𝑟𝑠 ( 𝑖, 𝑗) =∝−𝑃 𝑓( 𝑖′, 𝑗′)

(4-27)

where,

𝑖 ′ = ∝−𝑃 (𝑖( 𝑐𝑜𝑠𝜃 𝑟 ) + 𝑗( 𝑠𝑖𝑛𝜃 𝑠 ))

(4-28)

∝−𝑃 > 1; 𝑟 = 1,2, … , 𝑆; 𝑠 = 1,2 … , 𝐿.

(4-29)

𝑟 and 𝑠 represent the index for dilation (scale) and orientation, respectively. S is
the total number of dilatations (scales) and L is the total number of orientations,

120

𝜃 𝑟 and 𝜃 𝑟 are

the angle for each 𝑟 and 𝑠. In this study, S = 8 and L = 8 are

used, as proposed by Ng et al. (2005b). By applying the Gabor filter (fr) to an
mage

𝐼( 𝑖, 𝑗),

( 𝐼 𝑟𝑠 ( 𝑖,

𝑗)) is obtained as follows:

the magnitude information or magnitude filtered image

𝐼 𝑟𝑠 ( 𝑖, 𝑗) = {[ 𝑓𝑟𝑠 ( 𝑖, 𝑗) 𝑒 ∗ 𝐼( 𝑖, 𝑗)]2 + [ 𝑓𝑟𝑠 ( 𝑖, 𝑗)0 ∗
𝐼( 𝑖, 𝑗)]2 }

(4-30)

1
2

where “*” denotes a 2-dimensional convolution operation, and

𝑓𝑟𝑠 ( 𝑖, 𝑗)0

𝑓𝑟𝑠 ( 𝑖, 𝑗) 𝑒

and

represent the real (even), and imaginary (odd) parts of the Gabor

filter, respectively.
After the feature extraction, the 𝑘 extracted features were arranged in a 𝑘 –
vector:

𝑤 = [ 𝑤1 … 𝑤 𝐾 ] 𝑇 and normalized (Mery and Soto, 2008):

̃(𝑎,𝑏) =
𝑤

𝑤(𝑎,𝑏) − ̅̅̅ 𝑏
𝑤
𝜎𝑏

(4-31)

𝑎 takes values from 1 to the number of samples (scales), and 𝑏 takes values from
1 to the number of features (orientations).

̅̅̅ 𝑏
𝑤

and

𝜎𝑏

are the mean and

standard deviation of the b-th. Gabor feature. Gabor features, extracted from the
magnitude filtered images with different scales and orientations may be helpful
for extracting useful textural features from an image and important for
classification. Gabor features are useful in image processing applications such as
121

optical character identification, texture recognition in fruits/food, iris and
fingerprint recognition (Zhang, 2002; Ng et al., 2005a; Zhu et al., 2007).
Magnitude Gabor filtered maximum intensity XY-plane-slice CT images can be
found for a healthy and a decayed chestnut in Figs. 22a and 22b, respectively.
 Contrast features:
Image contrast is defined in section 2.3.3. In addition to statistical textural
techniques, which can be used to calculate contrast as described in Eq. 4-10,
contrast was calculated in each of the contrast enhanced secondary CT
grayscale images using additional techniques (5-methods) previously applied to
detect defects in aluminum casting X-ray images (Kamm, 1998), as seen in Eqs.
4-32 to 4-34:

𝐾1 =

𝐺− 𝐺𝑁
𝐺𝑁

(4-32)

𝐾2 =

𝐺− 𝐺𝑁
𝐺− 𝐺𝑁

(4-33)

𝐾3 = 𝑙𝑛 (

𝐺
)
𝐺𝑁

(4-34)

where, related to this study, G is the mean gray intensity value using an eroded
segmented region of the originally segmented CT image. This eroded segmented
region was obtained by applying an erosion algorithm (imerode in MATLAB) to
the original segmented region (whole chestnut) using a 15 x 15 disk SE. As in
contrast enhancement, this SE is appropriate because of the natural shape and

122

size of the original segmented chestnuts. GN is the mean gray intensity value of
the whole segmented region (whole chestnut).
Other methods of measuring contrast as proposed in Mery (2001), and
Mery and Filbert (2002) were also customized to this study. These other two
contrast measurements are calculated through inputting values into Eqs. 4-37
and 4-38, obtained from three consecutive steps by: (1) calculating a gray
intensity profile in the i-direction (P1) and in the j-direction (P2) in relationship to
the centroid of the whole segmented image towards the border of the segmented
image (background), (2) compute the first order function that contains the first
and last points in P1 and P2, therefore eliminating the bias ramp (approaching the
background) for each of the profiles (R1 and R2), and (3) generate the new
profiles (Q1 and Q2) without the background ramp, as seen in Eqs. 4-36 to 4-38:

𝑄1 = 𝑃1 − 𝑅1

(4-35)

𝑄2 = 𝑃2 − 𝑅2

(4-36)

𝐾𝜎 = 𝜎𝑄

(4-37)

𝐾 = 𝑙𝑛( 𝑄
where

𝑚𝑎𝑥

− 𝑄

𝑚𝑖𝑛 )

(4-38)

𝜎 𝑄 , 𝑄 𝑚𝑎𝑥 and 𝑄 𝑚𝑖𝑛 are the standard deviation, the maximum and

minimum value of Q, respectively. For a schematic representation and extended
information of the steps required to obtain
(2002).
123

𝐾𝜎

and K, refer to Mery and Filbert

4.3.9 Feature selection:
After feature extraction, it is necessary to select the best features to train
the classifier (Mery and Soto, 2008). The purpose of the feature selection step,
also known as feature reduction, is to obtain a smaller subset of features (m)
from the original data set (x), which will yield the highest classification rate
possible (Jain et al., 2000). High dimensionality increases time and space
requirements for processing data. Also, in the presence of irrelevant and/or
redundant features, classification methods tend to over-fit and become less
interpretable, especially when the number of features is much larger than the
number of samples. Feature selection algorithms usually involve maximizing or
minimizing an objective function (f), whose output can be calculated for the
generated m, therefore measuring their classification potential (effectiveness) and
working as a feedback signal to select the best features.
In this study, the sequential forward selection (SFS) technique (algorithm),
taking feature dependencies into account (eliminates features that are highly
correlated r ≥ |0.95|) (Silva et al., 2002), was selected as a search strategy (Jain
et al., 2000). SFS is one of the most widely used techniques, it is fast and starts
by selecting the best single feature in x (m = 1) then adds one feature at a time,
while eliminating features that are not relevant, and constantly monitoring the
classification effectiveness using different criterion functions. Three different
objective functions were evaluated in this study: (1) the Fisher score (J(W) – Eq.
4-39), (2) linear discriminant analysis (LDA – Eq. 4-40), and (3) quadratic
discriminant analysis (QDA) objective functions (Jain et al., 2000; Bishop, 2007).

124

̃
|𝑆 𝐵 |
𝐽( 𝑊 ) = arg max
̃
|𝑆 𝑊 |

(4-39)

where arg max stands for the “argument of the maximum” or the points of the
̃
̃
given argument for which J(W) reaches its maximum value. |𝑆 𝐵 | and |𝑆 𝑊 |
represent m-between-class (interclass) and m-within-class (intraclass) scatter
(covariance), respectively (Duda et al., 2000; Jain et al., 2000).
LDA is a subspace feature selection method (i.e. feature transformation),
allowing feature combination, which is based on J(W). It finds a linear
transformation (

𝑊 ∈ ℝ 𝑑x𝑚 ),

projecting the x-data into an m-dimensional

subspace in which the between-class (|𝑊
the within-class scatter (|𝑊

𝑇

̃ 𝑤 𝑊|)
𝑆

𝑇

̃ 𝐵 𝑊|) scatter is maximized while
𝑆

is minimized, as seen in Eq. 4-40. To

calculate the between and within scatter (covariance) it is assumed that both
have the same covariance matrix. (Duda et al., 2000; Quanquan et al., 2011).

|𝑊 𝑇 ̃ 𝐵 𝑊|
𝑆
𝐿𝐷𝐴 (𝑊) = arg max
|𝑊 𝑇 ̃ 𝑤 𝑊|
𝑆

(4-40)

Similar to LDA, QDA is a subspace feature selection method and can also
be used for classification/validation (Section 4.1.8). The difference is for QDA, the
between and within covariance matrices are assumed to be different between
each other (arbitrary) (Duda et al., 2000).

125

Figure 22. (a) Healthy and (b) partially decayed chestnuts with their corresponding secondary maximum CT
image (XY-plane-slice) (55 pixels x 65 pixels). (1) Local Binary Patter (LBP) transformations using different pixel
comparison (d,h) was applied to the secondary maximum CT image. (2) Secondary maximum CT image Gabor
filtered transformed images (Irs(i,j)) at different scales (r) and orientation (s) for an θ = 45 are included. (3)

126

Figure 22 (cont’d)

Example of Haralick textural (Tx) features, contrast and intensity features obtained from included secondary
maximum CT image. For visual reference, three-color fresh raw image slices of the evaluated chestnut are
included.

127

4.3.10 Classification (training and validation):
A supervised learning approach was used to train the pattern classification
algorithm (Duda et al., 2000). Supervised classes, known as labels, were based
on 5-, 3- and 2-categorical-groups, were each chestnut was invasively
categorized into 5-, 3- or 2-quality-classes, based on their internal disorder
(decay) severity level. An internally developed rating system was used, which
clearly and accurately defines the classes, taking in account invasive
standardized quality assessment methods, such as that described in UNECE
(2010). The rater is experienced in detecting, identifying, and quantifying quality
attributes (i.e. decay) in fresh chestnuts.
Quality classification per chestnut was expressed as the apparent
percentage of decay tissue in relation to the total area of each chestnut. In the
case of the 5-class-classification, a class-5 represents a chestnut that is
completely decayed, class-2 through class-4 represent chestnuts that are
partially to highly decayed, while class-1 is designated to chestnut with no decay
(healthy), as seen in Fig. 19 and Fig. 23a. Sample distribution, used to train and
validate the 5-class-classifier, can be observed in Fig. 23a. For the 3-classclassification, a class-3 represents a chestnut that is completely decayed, a
class-2 denotes partial to high decay, and class-1 indicates that the chestnut
does not contain decay, as seen in Fig. 19 and Fig. 24a. Sample distribution,
used to train and validate the 3-class-classifier is included in Fig. 24a. For the 2class-classification, a class-2 represents any chestnut with decay, while a class-1
shows that the chestnut does not contain decay, as seen in Fig. 19 and Fig. 25a.
128

Sample distribution, used to train and validate the 2-class-classifier, can be seen
in Fig. 25a.

129

Figure 23. (a) Sample distribution used to train and validate the 5-class classifier. Figure also contains an
example for each of the 5 categorical classes, representing the quality index levels.

130

Figure 23 (cont’d)

(b) Quadratic discriminant analysis (QDA) classifier performance using validation set with 4-folds, in relation to
the number of selected features (m). Black line represents the classification mean performance, dotted black line
(---) represent 95 % confidence intervals for the validation pool. (c) Validation confusion matrix corresponding to
the chestnut quality class prediction using 25 % of samples with 86-m (overall accuracy rate equal to 85.9 %).

131

Figure 24. (a) Sample distribution used to train and validate the 3-class classifier. Figure also contains an
example for each of the 3 categorical classes, representing the quality index levels.

132

Figure 24 (cont’d)

(b) Quadratic discriminant analysis (QDA) classifier performance using validation set with 4-folds, in relation to
the number of m. Black line represents the classification mean performance, dotted black line (---) represent 95 %
confidence intervals for the validation pool. (c) Validation confusion matrix corresponding to the chestnut quality
class prediction using 25 % of samples with 155-m (overall accuracy rate equal to 91.2 %).

133

Figure 25. (a) Sample distribution used to train and validate the 2-class classifier. Figure also contains an
example for each of the 2 categorical classes, representing the quality index levels.

134

Figure 25 (cont’d)

b) Quadratic discriminant analysis (QDA) classifier performance using validation set with 4-folds, in relation to
the number of m. Black line represents the classification mean performance, dotted black line (---) represent 95 %
confidence intervals for the validation pool. (c) Validation confusion matrix corresponding to the chestnut quality
class prediction using 25 % of samples with 126-m (overall accuracy rate equal to 96.1 %). Figure is partially
presented in color.

135

Using the optimized selected features obtained from section 4.3.9,
decision boundary lines, planes, and hyper planes were implemented using LDA
(Section 2.7, Eq. 41), QDA (Section 2.7), Mahalanobis distance (MD) (Eq. 4-41),
a two-layer artificial neural network (ANN), a three-layer ANN using a logistic
activation function, and a three-layer ANN using a Softmax activation function
following the applied procedure in Ren et al. (2006; 2010), Mery et al. (2010),
Leiva et al. (2011), and Donis-González et al. (2013). In general, this step
assigns the object (i.e. set of 6 secondary CT images per chestnut) to a specific
quality category (class).
LDA, and QDA were described in section 4.3.9. The MD measures the
difference between two points in the space defined by two or more variables
(features). The MD takes the correlations within a data set between the variable
into consideration, meaning that it depends on the covariance matrix of the
attribute and adequately accounts for the correlations. The MD is the distance
between an observation and the mean for each group in m-dimensional space,
defined by m-features and their covariance. A small value of MD increases the
probability of a set of features (new sample) to be closer to the class’s mean and
the more likely it is to be assigned to that group-class. The MD between a set of
features (m) and a class (μ) is defined as:

𝑀𝐷 = √(𝑚 −

𝜇) 𝑡

−1

∑ (𝑚 − 𝜇)

136

(4-41)

where

Σ-1 is the inversed covariance matrix. The MD does not depend upon the

scale on which the variables are measured (Duda et al., 2000).
When used for classification ANN, is usually a collection of neuron-like
processing units with weighted connections between the units. ANNs are
computer algorithms, which are highly parameterized statistical models that can
automatically find relations in the data without a predefined model. ANNs first
linearly transform the input feature vector by multiplying it with a weight matrix
(wji) then, the activation function (f(∙)) is applied to each coordinate of the
resulting vector to produce a value. Training an ANN involves determining the wji
that maximize the performance for a set of supervised training data, using a
specific f(∙) (Duda et al., 2000). In general, ANNs transform a feature vector in the
input space (input-i) to a vector in the output space (output-k).
A two-layer ANN was evaluated in this study, this type of ANN can only
implement a linear decision boundary or activation function (f(∙)). Multilayer
networks can be built by subsequent application of additional wji and additional
nonlinear f(∙). In addition to the simple linear two-layer ANN, two three-layer
ANNs were also evaluated in this study using two different f(∙): (1) logistic and (2)
Softmax. More information, in depth discussion, methodology description, and
followed steps for each of the applied the different classifiers can be found in
Duda et al. (2000).
Performance of each of the classifiers was measured as the correctly
classified chestnuts, using the set of secondary CT images, in reference to its
137

supervised categorical class (label). Classifier validation was implemented using
a 4-fold stratified technique, therefore yielding an average estimate of classifier
performance with 95% CI (Confidence Intervals) for the validation pool (Jain et
al., 2000). In the experiment, 75% of the samples were used for training and 25%
were used for validation repeated four times.
4.4 Results
In this study, the features were reduced from 1194 to 86 (5-class
classifier), 155 (3-class classifier) and 126 (2-class classifier) features, in order to
avoid overtraining. This feature reduction was implemented using different
algorithms. The SFS with the Fisher discriminant objective function (J(W))
method offered the most powerful features, yielding the best classifier
performance. Reducing overall feature space diminishes the computational time,
allowing for possible in-line implementation. The other feature selection methods,
using addition objective functions, as described in section 4.3.9 performed poorly
(Results not shown). The main selected features for all of the classifiers are
enumerated in Table 9. Example of image transformation (cLBP and Gabor) and
some of other selected features, obtained from a healthy and partially decayed
chestnut can be seen in Figs. 22a and 22b, respectively. Validation of
classification was carried out using a 4-fold stratified validation with 25% of the
samples for each fold (repetition). The best overall classifier is the QDA classifier.
Other classifiers were also tested, however the overall classification performance
was lower in all cases, as can be seen in Table 10.

138

Table 9. Main selected features (seventy five) for the Quadratic discriminant
analysis (QDA) classifier using sequential forward selection (SFS) in
combination with the Fisher discriminant objective function (J(W)) for the 5, 3- and 2-class classifiers.
n Selected feature (m) n Selected feature
n Selected feature
(m)
(m)
1
26 ̃(1,1) [YZ-max]
51 Tx(13,1)-Mean
̃(3,3) [XY-max]
𝑤
𝑤
[XY-max]
2 Tx(8,1)-Mean [XY27 cLBP(1,6) [XY-max] 52 ̃(3,3) [YZ-mean]
𝑤
mean]
3 Tx(9,1)-Mean [XY28 cLBP(1,2) [XY-max] 53 ̃(2,6) [XY-max]
𝑤
max]
4 Intensity-Std. Dev.
29 sLBP(1,48) [XY54 Contrast-Ks [XY[XY-max]
max]
max]
5
30 cLBP(1,5) [XY-max] 55 cLBP(1,3) [XỸ(5,8) [XY-max]
𝑤
max]
6 Tx(1,1)-Mean [XY31 sLBP(1,40) [XY56 Contrast-K [XYmax]
max]
max]
7 Intensity-Skewness
32 sLBP(1,5) [XY-max] 57 Tx(11,1)-Mean
[XY-mean]
[XY-max]
8 cLBP(1,9) [XY-max]
33 sLBP(1,41) [XY58 Tx(6,1)-Range
max]
[XY-max]
9
34 sLBP(1,25) [XY59 Tx(10,1)-Mean
̃(7,4) [XY-max]
𝑤
max]
[XY-max]
10 Tx(5,1)-Range [YZ35 ̃(6,7) [xY-max]
60 Tx(7,1)-Range
𝑤
max]
[XY-max]
11 Intensity-Mean [XY36 sLBP(1,40) [YZ61 cLBP(1,36) [XYmax]
mean]
max]
12 ̃(4,7) [XY-max]
37 cLBP(1,16) [XY62 sLBP(1,20) [XY𝑤
max]
max]
13 Mean Laplacian [XY- 38 cLBP(1,12) [XY63 Tx(6,1)-Mean [XYmax]
max]
max]
14 Tx(12,1)-Mean [XY39 cLBP(1,26) [XY64 Tx(7,1)-Mean [XYmean]
max]
max]
15 Tx(13,1)-Mean [XY40 ̃(4,4) [XY-max]
65 Tx(13,1)-Mean
𝑤
max]
[XZ-mean]
16 Tx(3,1)-Mean [XY41 ̃(4,6) [XY-max]
66 sLBP(1,54) [XY𝑤
max]
max]
17 sLBP(1,50) [XY-max] 42 Tx(13,1)-Mean [XY- 67 ̃(5,7) [XY-max]
𝑤
mean]

139

Table 9. (Cont’d)
18

̃(7,6) [XY-max]
𝑤

43

̃(6,5) [XY-max]
𝑤

19 sLBP(1,34) [XZ-max]
20 Tx(5,1)-Range [XYmean]
21 Intensity-Skewness
[XY-mean]
22 Intensity-Skewness
[XY-max]
23 Tx(12,1)-Mean [XYmax]
24 Tx(3,1)-Mean [XYmean]
25 LBP(1,1) [XY-max]

44
45

̃(2,3) [XY-max]
𝑤
̃(1,3) [XY-max]
𝑤

46

̃(1,4) [XY-max]
𝑤

47 Tx(4,1)-Mean [XYmax]
48 LBP(1,13) [XY-max]
49 sLBP(1,13) [XYmax]
50 LBP(1,24) [XY-max]

68 sLBP(1,45) [XYmax]
69 ̃(1,7) [XY-max]
𝑤
70 Tx(9,1)-Mean [XZmax]
71 sLBP(1,47) [XZmean]
72 Tx(5,1)-Mean [XZmean]
73 Mean Laplacian
[XY-mean]
74 LBP(1,32) [XYmax]
75 Tx(1,1)-Mean [XZmean]

Tx(k,p)-(Mean, Range): Haralick texture features, where k is the texture type, as
seen in section 2.6.2 – Table 8, and p is the number of neighbor pixels.
LBP(d,h): Classic local binary patterns sematic. sLBP(d,h): Sematic local binary
patterns. Where d is the number of compared pixels with h – neighboring pixels.
See section 2.6.3.
̃(a,b): Gabor filters, where a is the frequency number, and b is the number of
𝑤
orientations.
Between brackets [] are the different CT images used to extract features (XYmean = XY-plane-slice mean images, YZ-mean = YZ-plane-slice mean image,
XZ-mean = XZ-plane-slice mean image, XY-max = XY-plane-slice maximum
image, YZ-max = YZ-plane-slice maximum image, and XZ-max = XZ-plane-slice
maximum image). For reference, see sections 1.5 and 4.3.3.

140

Classifier

Table 10. Classifier performance using selected features (m) with a 4-folds validation
5-classes
3-classes
2-classes
Mean

UCI1

LCI2

Mean

UCI

LCI

Mean

UCI

LCI

QDA

85.9

87.2

84.6

91.2

92.1

90.3

96.1

96.9

95.3

ANN-Logistic

85.9

87.3

84.5

88.9

89.7

88.1

95.6

96.5

84.7

ANN-Softmax

85.6

87.0

84.2

86.7

87.5

85.9

95.9

96.6

95.2

LDA

84.6

86.3

82.9

91.0

91.9

90.1

96.5

97.4

95.6

MD

83.9

85.4

82.4

87.6

88.5

86.7

93.2

94.1

92.3

ANN-linear

81.6

82.9

80.2

86.6

87.5

85.7

93.3

94.4

92.2

m
1
2

86

155

Upper Confidence interval
Lower Confidence interval

141

126

Performance results for 5-, 3- and 2-classes, using the selected QDA
classifier, for increasing number of features (m), are included in Figs. 7b, 8b and
9b, respectively. It can be seen how performance increases in relation to an
increase in m and that with the 86-, 126- and 155-m the classifiers had a high
overall performance accuracy classification rate of 85.9 % (5-classes), 91.2 % (3classes) and 96.1 % (2-classes). Classifier performance slightly increases after
the reduced number of selected features (m), but not notably (> 0.5 %), so
additional features are not required to classify internal chestnut quality. In
addition, it is not recommended to use a higher number of features to avoid
classifier over-training, due to overall sample size, as recommended by Duda et
al. (2000).
Figs. 23c, 24c and 25c include the confusion matrix corresponding to the
overall QDA classifier performance. This matrix is beneficial, because it visibly
documents the misclassified classes (Shapiro and Stockman, 2001).
4.5 Discussion
By observing Table 9, it can be seen that the most important features for
classification mainly include textural features (approximately 88 %) including: (1)
LBP features, (2)Tx and contrast features, and (3) Gabor features at different
scale and orientation, acquired from the different CT images. Less influential,
about 12 % of the utmost important features involve the basic intensity features.
Examples of some of the extracted features, and image transformations to XYplane-slice CT images can be seen in Fig. 22.

142

Textural features are important because they cannot be attributed to a
single pixel value, but rather to several pixels in the image and their relationship.
It can be postulated that LBP textural features are among the most influential
features, because they are invariant to monotonic grayscalse changes; therefore
CT image noise and imaging variability do not play an imperative role (Ojala et
al., 2002). LBP grayscale textural features captured the local structure and
textural variations between decayed, healthy and partially decayed chestnuts
clearly reflected in the CT images and the generated LBP images, as reflected in
Fig. 22. In general, Tx and contrast features, which were also highly important,
described the pixel spatial variation and their relationship in the segmented
image. For example, CT images that contain pixels which are similar between
each other have a low variance sum (Tx(8,p)), while images that contain a high
variability of pixel intensity values have a high contrast (K) and high variance sum
(Tx(8,p)) (Haralick, 1979). Therefore, in this study, a uniform healthy or decayed
chestnut will represent pixels that are highly correlated (low variance sum Tx(8,p)), but a low overall image contrast (K) as observed in Fig. 22a. On the
other hand, CT images of chestnut that have a high variation between the pixels
(e.g. decayed tissue embedded in healthy tissue from partially decayed
chestnuts) will have a higher variance sum (Tx(8,p)) and contrast values (K) as
exemplified in Fig. 22b. Gabor textural features captured different image
information by means of combining different scaling and orientation factors.
Gabor features, extracted from magnitude Gabor filtered images (Irs(i,j)), have
143

been found to be appropriate for texture representation and discrimination (Zhu et
al., 2007), thus it is hypothesized that Gabor features offer information regarding
the development of chestnut decay, yielding a different feature value to chestnuts
with partial decay in comparison to chestnuts that are either completely decayed
or healthy. Examples, of several magnitude Gabor images (Irs(i,j)) generated by
applying 3 different Gabor filters to a maximum secondary CT images (XY-planeslice) of a healthy and partially decayed chestnut can be observed in Figs. 6a and
6b, respectively. These Irs(i,j) are useful to visualize how Gabor features can
differentiate between chestnuts that contain internal decay in comparison to a
healthy chestnut. It is clear that Gabor filtered secondary CT images of a
decayed chestnut present a higher level of textural attributes in comparison with
a healthy chestnut.
Even though simple image intensity features only account for roughly 12 %
of the most important features, these features represented and summarize the
overall quality appearance of the chestnuts in the CT images. It could be
quantified that decayed tissue has a lower intensity in comparison with healthy
tissue (See Fig. 22). However, possibly because of the high variation between
the CT images of the same class and CT image noise, simple intensity features
by themselves do not provide enough information and are not as sensitive to
accurately classify chestnuts, as it was initially hypothesized and briefly
discussed in Donis-González et al. (2012).
It was observed that the secondary contrast enhanced maximum intensity
value CT images (Figs. 20h-20j, Fig. 21f and Fig. 22), which summarized the
144

overall pixel values of all of the cropped chestnut CT image slices, offered better
features than the secondary contrast enhanced mean intensity value CT images
(Figs. 20e-20g and Fig. 21d). After feature selection, 80 % of the features were
extracted from secondary maximum intensity CT images, while only 20 % are
extracted from secondary mean intensity CT images. This might be the case,
because maximum images are less variable, have a higher contrast and can
easily discern between healthy and decayed tissue, as visually observed in Fig.
21. Fig. 21 shows that it is visually easier to distinguish the decay tissue in the
secondary maximum intensity CT image in comparison to the secondary mean
intensity CT image. In addition, secondary XY-plane-slices offer better features in
comparison to reconditioned secondary YZ-plane-slices and XZ-plane-slices. The
reason behind this is that image spatial resolution is higher in the XY-plane in
comparison to the images in the YZ- and XZ-planes (anisotropic voxels), as was
summarized in Table 1. Loss in image resolution in both planes comes from the
fact that images are re-sliced onto the Z-axis using CT images that have been
acquired every 2.5 mm (d), while pixel size in both the X and Y planes is equal to
0.73 mm, considerably smaller that d. It is hypothesized that this loss in
resolution could be emended by evaluating different CT scanning techniques and
CT image reconstruction algorithms, but more research is required to accurately
address this issue.
A confusion matrix is a specific table layout that allows visualization of the
performance of a classification algorithm. Each column of the matrix represents
the instances in a predicted class, while each row represents the instances in an
actual class. The name stems from the fact that the matrix makes it easy to see if
145

the system is confusing two classes (i.e. commonly mislabeling one as another).
It also reports the number of false positives, false negatives, true positives, and
true negatives. The diagonal of the matrix represents the true positives and true
negatives of each class (correct classification), while the other values represent
the false positives and false negatives (classification error) (Duda et al., 2000). In
this study, false negatives designate chestnuts that are classified as decayed,
while being healthy. In this scenario, healthy chestnuts might be unnecessarily
discarded creating direct profit losses. On the other hand, false positives reflect
chestnuts that are decayed and classified as healthy. This means that potentially
a client will obtain chestnuts that are decayed while expecting to receive a
healthy product. False negatives might generate customer complaints and might
create an indirect economic loss, which is difficult to directly calculate. By
analyzing the confusion matrix in Fig. 23c, it could be seen that classification
errors occur evenly across the different classes. This means that in general,
healthy chestnuts (class-1) and minimally decayed chestnuts (class-2) present
similar classification accuracy when trying to discriminate between chestnuts that
contain medium to high decay (class 3,4 and 5). In other words, chestnut quality
is similarly discerned regardless of the quality level. Similar results are seen in
the confusion matrices found in Figs. 24c and 25c for the 3- and the 2-class
classifiers, respectively.
Results from this study show that CT images, acquired using a medical
grade CT scanner, can be used as a technique that will be able to classify
chestnuts based on their internal quality. This technique objectively, rapidly, and
automatically classifies chestnut slices into up to 5-classes by measuring different
146

textural, and basic intensity features from six CT images. Most importantly,
chestnuts that are healthy can be accurately classified from decayed chestnuts
with a 96 % accuracy rate (2-class-classifier). This high classification rate can be
accomplished with a relatively low number of features in relation to the available
number of images, with appropriate feature selection (i.e. = SFS), and
classification techniques (i.e. = QDA). Therefore this study offers a tool, to
objectively forecast the overall quality of fresh chestnuts. It helps the research
community to understand which are the features that play an important role in the
ideal classification of CT images. In addition, it creates a general structure that
could be used as a reference tool in the development of an in-line noninvasive
quality sorter of fresh chestnut, and similar products, like peanuts, almonds, and
other fresh commodities (e.g. apples, cucumbers, cherries, pineapples) using a
fast CT systems (i.e. using digital cameras). Transferring this tool to different
commodity industries will enable them to control and promote the quality of their
products. In addition, for the chestnut and other fruits, vegetables, and nuts, this
application will provide the opportunity to select products according to their final
use.
4.6 Conclusions
The CT imaging system provides high-resolution and high-contrast images
of the internal structure and components of fresh chestnuts. After scanning and
cropping, approximately 50 original CT image slices (stack) were obtained per
chestnut, from three different planes (angular orientations) across the longitudinal
(Z) (XY-plane-slice), horizontal (YZ-plane-slice) and vertical (XZ-plane-slice)
147

axes. From this image stack, 6 secondary CT images per chestnut sample,
including mean and a maximum intensity value images for each of the planes
were extracted. Secondary extracted CT images were then preprocessed
(contrast

enhancement)

and

segmented,

using

a

balanced

histogram

thresholding method. Thereafter, a total of 1194 grayscale intensity, and textural
features were extracted from the segmented region in the 6 secondary CT
images per sample. SFS was carried out to reduce the dimensionality of the total
image-extracted features. Ultimately, 86, 155 and 126 features were found to be
effective in designing a QDA classifier with a 4-fold cross-validated overall
performance accuracy of 85.9 %, 91.2 % and 96.1 % for 5, 3 and 2 classes,
respectively.
Although the specific developed classification procedure might not be
applicable to other foods, agricultural commodities or other applications directly,
the methodology is broadly valid. Before applying the classification procedure to
other commodities or other applications (e.g. industrial), it will be necessary to
train it appropriately. These results showed that this method is an accurate,
reliable, and objective innovative tool to determine chestnut internal quality, and
would be applicable to an automated noninvasive in-line CT sorting system.

148

CHAPTER 5. DISCUSSION AND FINAL REMARKS
In response to the need for postharvest internal quality sorting equipment
for fresh chestnuts, and the economical benefits of appropriate produce quality
assessment, the long-term goal of this research is to aid in the design of
commercial-scale fast in-line CT sorting equipment, thus enabling the chestnut
industry to quickly sort produce, based on its internal quality attributes (i.e.
internal decay). Nonetheless, before developing commercially available in-line CT
equipment to sort chestnuts and other commodities, some concerns need to be
addressed, including the potential sorting speed of the CT equipment, cost, and
client acceptability (reaction to the potential of food irradiation).
The first CT scanner developed by Godfrey Hounsfield in 1971, took
several hours to acquire the projection images and took several days to
reconstruct them into a single 2D slice. After the CT system's 42 years of applied
history, current CT systems can scan a whole adult human body and create a 3D
reconstruction within seconds (Bushberg et al., 2002). In non-medical fields (e.g.
agriculture), scanning can be made several times faster by sacrificing image
resolution and quality, reducing the number of projection images to reconstruct
2D slices, decreasing number of slices acquired per sample, changing the energy
and current required for scanning, implementing new technical and hardware
scanning concepts, among several other parameters. In the agriculture and food
industries, the sorting throughput requirements vary among commodities, but
range between one to forty items per second (Chen and Sun, 1991), with
chestnuts falling within this range. The intention would be to eventually develop a

149

sorting machine that can fulfill these throughput needs. In this study, image
acquisition and reconstruction is equal to approximately 0.6 s per chestnut.
Feature extraction, re-slicing and classification all together total 1.8 s per
chestnut. This would yield a classification throughput rate of approximately 1
chestnut every 2 s. This time is only ideal for the lower end of the agricultural
application speed sorting range, resulting in equipment that could sort around
500 kg of chestnuts on an 8 h shift day. I am confident that future research,
including equipment parameter and hardware modifications for in-line sorting,
could significantly increase this throughput. Development of such sorting
equipment to fulfill these high throughput requirements, using computed
tomography, is a challenge. Yet, based on observed changes over the life span
of CT, and especially recent advances that have significantly increased image
acquisition rate, reconstruction speed, image quality, resolution, and reduced Xray exposure, I think that soon it will be possible to use this technique for sorting
purposes.
Such technology is already being developed and tested for a fast in-line
CT quality inspection system for mass production in the car industry (Stuke and
Brunke, 2010). Other scientists, such as Hampel et al. (2005) and Bierberle et al.
(2009) are evaluating the use of an ultra-fast CT X-ray system to study gas-liquid
two-phase pipe flow. As another example, with the purpose of preliminary
evaluation of a method that can significantly increase traditional CT scanning
speed and can be developed further to in-line sorting purposes, a small set of
chestnut samples (n = 400, year = 2011) were scanned using the ultrafast
ROFEX-scanner (Fraunhofer Institute of Electron Beam and Plasma Technology,
150

Dresden, Germany – Fig. 26a). This scanner was operated at its fixed settings
equal to 150 keV maximum X-ray energy with a maximum electron beam power
of 10 kW. The system can provide a temporal resolution of up to 7000 frames s -1
at a spatial resolution of roughly 1 mm, depending on attenuation behavior of the
object. To achieve similar results in objects with higher attenuation behavior than
water, temporal resolution has to be reduced, therefore increasing scanning time
as was required in chestnuts. Chestnuts were labeled and packed in a thin plastic
hose. The hose was pulled through the scanning plane using a stepper motor, at
a constant speed of 1 m s-1, as seen in Fig. 26b. From this, sets of cross
sectional 2D CT images for every chestnut were acquired. The frame rate was
chosen to 2000 frames s-1 (10-20 times faster than the traditional CT imaging
systems), which is a good compromise between image quality and temporal
resolution. Fig. 27 offers preliminary visual results of what can be inferred about
fresh quality using ultrafast ROFEX CT-images. It can be seen that differences in
the gray scale intensity values (HU-values) between different tissues types are
better observed when distinctive tissue types coexist within the same chestnut,
such as in the images in Fig. 27b. It can also be observed from these CT images
it is difficult to visually distinguish between extremely decayed (Fig. 27c) and
healthy chestnuts (Fig. 27a). On the other hand, when decay tissue is embedded
between healthy tissue (Fig. 27b), a slight visual difference in grayscale values
can be visually observed. Nonetheless, the presence of void spaces and pellicle
were easily discerned. Preliminary results indicated that the scanning conditions using the ROFEX scanner were not optimal to detect decay because the
151

maximum X-ray energy of 150 keV was too high, resulting in diminishing the
ability to differentiate tissue differences. Therefore, future studies will focus in
optimizing image quality using different scanning parameters (e.g 60 keV X-ray
energy); with the objective of rapidly and effectively detecting properties in
different fresh agricultural produce, for example chestnut decay. Technology,
such as ROFEX CT technology, if appropriately applied with its fast scanning
capabilities, could easily be potentially installed in-line to automatically and
rapidly sort agricultural commodities, including chestnuts.

Figure 26. (a) Ultrafast Rossendorf electron beam X-ray tomograph (ROFEX)
scanner working principle. (b) Experimental setup of the ROFEX-scanner.
152

Figure 27. Visual preliminary results of CT images (lower row) obtained
using the ROFEX-scanner with its corresponding color raw image slices
(upper row). (a) Healthy, (b) partially decayed (rotten) and (c) completely
decayed chestnuts. (d) 3D reconstruction of two chestnuts, showing a
rotten section in one of the chestnuts (white arrow).
Currently, an accurate estimation of the cost and cost effectiveness of
using CT sorting systems is not available and is difficult to predict or even
speculate because the equipment is not commercially available. More studies are

153

required to accurately measure these. Prices depend on the type of machine, but
in general, a new CT scanning machine costs between 75,000 US$ to 1,000,000
US$, without considering operating and maintenance costs (Abrams and McNeil,
1978; Cenegage and Krapp, 2002).This is considerably less expensive than a
homologous MRI device, but more expensive than 2D X-ray equipment (Alanen
et al., 2004). Prices are especially high and significantly inflated because
currently CT is only used in the medical industry and the economics of scale and
Moore’s law have not applied (Goetz, 2010). Nevertheless, it can be expected
that in the near future, changes in the technology, efforts to develop new
concepts, and different applications, should significantly reduce the price of CT
equipment, enabling the technology to be used in agriculture, food, and other inline sorting applications.
Differently to X-ray for food quality diagnostic purposes, as described in
this dissertation, food treatment using X-ray radiation to improve microbial safety
has been extensively studied. The safety of consumption of irradiated food has
also been comprehensively studied. Several international specialist groups in
collaboration with the World Health Organization (WHO), the FAO, the
International Atomic Energy Agency (IAEA), and the Scientific Committee on
Food (SCF) of the European Commission concluded that irradiated foods, with
appropriate technologies, are safe and nutritionally suitable. Specific applications
of food irradiation are approved in over 55 countries, including the United States
of America. In general, food irradiation is described as a process where food is
exposed to ionized energy, utilizing X-rays of 5 MeV maximum kinetic energy,

154

gamma photons emitted by 60Co and 137CS radioisotopes, and 10 MeV
accelerated electrons. None of these sources have induced radioactivity in the
food or its packaging. In general, the maximum allowed absorbed dose
(measured in grays - Gy) in food applications is 60 kGy, which occurs with meat,
poultry and fish. To extend shelf-life of fresh fruits/vegetables and dry spices
dosages ranging between 0.1-3.0 kGy are recommended (Farkas and MohácsiFarkas, 2011). In the case of the CT scanner used in this study, applying the
settings optimized in Chapter 2, the CT equipment emits up to 0.1 MeV kinetic
energy, and a maximum of 8 x 10-9 kGy per cm3 scanned. These values are
considerably lower than the amount permitted in food irradiation. Because of this,
I am confident that CT sorting of agricultural products should not cause any
substantial safety concerns. Consumer acceptance is a matter of education, and
good communication leads to lessening the unfair image that irradiated food is
malicious (Teisi et al., 2009). Misrepresentation of food irradiation prevents the
utilization of a safe and beneficial process. The advantage is that marketing trials
have shown that an increasing amount of consumers are willing to buy irradiated
foods, if properly informed about the technology (Eustice and Bruhn, 2006).
Up to now, work related to CT for internal quality attributes in food and
agriculture has been limited to visualization. In addition to the optimum
visualization of internal attributes of chestnuts, which was described in this study,
this research also provides a novel technique to automatically classify chestnuts
using CT images. The research conducted to develop this dissertation provides
important tools and scientific information to appropriately develop, and hopefully
155

move forward in the incorporation of CT equipment technology into a specific
agro-food industry application (chestnuts quality inference), and possibly others.
Information obtained from this dissertation, even though specific to chestnuts,
offers a blue print to be applied to other commodities (see appendices C), it is
likely to be adaptable to measure and classify a wide range of internal quality
attributes in other commodities. In addition, if practically applied, the sorting
equipment will enable the chestnut industry to: Efficiently use by-products; lessen
the negative environmental and health effects of field agro-chemical usage
because of better postharvest monitoring; offer safer products (e.g. early
detection of mycotoxins producing microorganism); avoid loss of produce
quantity; detect postharvest chestnut pathogens, with the primary objective of
taking immediate actions to avoid negative effects of microorganisms colonization
(e.g. reduction of nutritional quality), thus increasing industry revenue and
sustainability.
This study demonstrated that CT is a reliable technique for chestnut
internal quality detection and classification, capable of providing unique quality
information that is not obtainable by any other commercially available equipment,
and is appropriate to apply as an in-line sorting system. It indicated that CT
technology has the ability to effectively detect common and economically
important internal quality attributes from fresh chestnuts, with better and higher
spatial and contrast resolution than commercially available 2D X-ray sorters (Fig.
28a). As a demonstration and evaluation, a preliminary study with whole fresh
intact chestnuts, conducted in the years 2009 and 2010 with a commercially
available 2D X-ray sorter, yielded slight, almost undetectable, differences
156

between chestnut quality levels (Fig. 28b). On the other hand, as demonstrated in
this study and dissertation, when chestnuts were imaged using a medical grade
CT scanner, excellent visual chestnut tissue characterization,

accurate

quantification of internal defects, and high non-invasive classification rates were
observed.

Figure 28. (a) Commercially available in-line traditional X-ray sorter, and (b)
its corresponding 2D X-ray images.

157

Additional to the previously mentioned benefits, the possible integration of
this postharvest technology sorting system into the traditional packinghouse
setup, which includes chestnuts, as sketched in Fig. 29, will lead to reducing the
industry reliance on hand labor, decreasing steps during the packaging process,
possibly minimizing waste, reducing water usage and, lessen product losses.

158

Figure 29. Typical unit operations in mechanized packinghouse (Thompson et al., 2002). Vision of future
integration of in-line Computed Tomography (CT) for sorting fresh agricultural products*. *: Flowchart in red are
the proposed sections that will be potentially improved.

159

APPENDICES

160

APPENDIX A
Estimation of SNR, Volume accuracy, High contrast spatial resolution
(HCSR), Low Contrast Detectability (LCD), and digital quality assessment
(DQA)

161

A. 1: Estimating SNR. The mean value of 10 repeated ROI measurements from
different 16-bit CT images (Fig. 7a), containing healthy chestnut tissue was used
as the final SNR value within each run.

∑ 𝑥𝑝𝑥=1 𝐻𝑈 𝑝𝑥
(
)
𝑥

∑#𝑅𝑂𝐼
𝑖=1

(A. 1)

2

√∑ 𝑥𝑝𝑥=1(𝐻𝑈 𝑝𝑥 − ̅̅̅̅)
𝐻𝑈
𝑥−1

𝑆𝑁𝑅 =

((

#𝑅𝑂𝐼

))

where,
HUpx = Hounsfield unit value of each pixel (px).

HU = Mean intensity of all pixels in each ROI.
x = Number of pixels for each ROIs.
A. 2: Estimating Volume Accuracy. Teflon® reference cylinder estimated volume
was digitally estimated by calculating the area of each XY-plane 2D CT image as
seen in Fig. 7b, using corresponding binary images (after applying a simple
global threshold of 135) to count the number of pixels in the segmented region,
multiplying this value by the slice thickness and then summing the total number of
acquired images per cylinder. Therefore, total volume accuracy is defined in Eq.
(A. 2).

162

#𝑅𝐶

𝑉𝐴 = ∑ (
𝑗=1

1

)
#𝑅𝐶 ∗ |𝑇𝑉𝑗 − 𝐸𝑉𝑗 |

(A. 2)

where,
TV = Teflon® reference cylinders true volume (~ 14,743 mm3).
EV = Digitally estimated Teflon® reference volume
RC = Number of reference cylinders (3).
A. 3 and A. 6: Estimating HCSR and LCD. HCSR was calculated using a
replication of three 8-bit CT images (e.g. Fig. 7c), without the aid of a human
observer and adapted from the specified protocol in General Electric Company
(2007), by deriving and applying Eq. (A. 3).

(A. 3)

where,
im is the number of analyzed images (3).
N each box sized to fit into each pattern (lp) as observed in Fig. 7c,
representing a quantitative assessment of changes in image resolution.

163

Values of N range from 1 to 5, where N equal to 1, 2, 3, 4, 5 are boxes in
the patterns with a lp equivalent to 1.6 mm, 1.3 mm, 1.0 mm, 0.6 mm, and
0.5 mm, respectively.
Ipx is the intensity value of each px.

I

is the mean intensity of all px in each box (N).

x is the total number of pixels in each box (N).
c = Constat equal to 40 I, indicating the optimum SD intensity for an lp of
1.6 mm (baseline), as specified by General Electric Company (2007).
CC = Number of Connected Components (objects or blobs) in the binary
image (Shapiro and Stockman, 2001) after applying a simple global
threshold of 134 to its corresponding XY-plane 2D CT image (Fig. 7d).
TCC = Total Connected Components (CC) in one image (30).
The LCR was consistently calculated for each run; using a replication of
three CT images as exemplified in Fig. 7d. Combined sets of criteria (Eq.
(A. 4) and (A. 5)) were used to develop Eq. (A. 6), which characterizes
LCD.

(A. 4)

where,
MC = Mean Circularity (Shapiro and Stockman, 2001) from three binary
image repetitions after applying a simple global threshold of 94 to its
164

corresponding XY-plane 2D CT image (e.g. Fig. 7d) for each set of holes
(H) - k (total of 5). Values of k range from 1 to 5, where k equal to 1, 2, 3,
4, 5 represents the circular holes in the membrane with a diameter
equivalent to 10.0 mm, 7.5 mm, 5.0 mm, 3.0 mm, and 1.0 mm,
respectively.
im = Number of analyzed images (3).
A = Area of each k.
P = Perimeter of each k
In theory, MC increases as each k becomes more circular (Shapiro and
Stockman, 2001).

(A. 5)

where,
MAE = Mean Area Error from a triplicate of binary images after applying a
simple global threshold of 94 to its corresponding XY-plane 2D CT image
(e.g. Fig. 7d) for each hole (H).
im = Number of analyzed images (3).
TA = True Area (known from each H true diameter).
EA = Estimated Area (counting number of pixels for each H in binary
image).
MAE decreases, as each hole (H) is closer to its true area (TA).

165

(A. 6)

where,
MC = Mean Circularity using Eq. (A. 4)
MAE = Mean Area Error using Eq. (A. 5).
SDH = Mean area value of the Smallest Detected hole (H).
A. 7: Estimating Digital Quality Assessment (DQA). To perform the DQA, each
16-bit CT gray scale image (e.g. Fig. 8b) corresponding to the subjectively
evaluated fresh-raw color image (e.g. Fig. 8a), was first segmented into a binary
image containing the area (total number of pixels) of whole chestnut tissue per
slice, by applying a simple global threshold of 400 HU (e.g. Fig. 8c). Second,
transition points, Shell, and Pellicle area was determined after applying a Sobel
filtering method to detect edges and tissue transition points (Shapiro and
Stockman, 2001), as seen in Fig. 8d. Third, healthy tissue was determined by
calculating the binary image area after segmentation, using a simple global
threshold of 1050 HU (Fig. 8e). A simple global threshold can effectively be used
to segment any of the CT images because images are not sensitive to changes in
lighting conditions, scanned object color, and potential seasonal changes, as it is
the case when processing real color images (Blasco et al., 2007). Ultimately, CT
digital DQA was expressed as seen in Eq. (A. 7).

166

(A. 7)

where,
HT = Area of healthy chestnut tissue.
WT = Area of whole chestnut tissue.
SP = Area of shell, pellicle and transitions points.

167

APPENDIX B
CT imaging in other fresh agricultural commodities

168

Figure 30. Color raw images of (a) chestnuts, (b) pineapples, (c) tart cherries and (d) pickling cucumbers showing
the regions of interest (ROIs) for each fresh commodity.

169

Figure 31. (a) Color raw image slices, cross-sectional 2D CT images acquired using the GE BrightSpeed™ RT 16
Elite CT scanner, and 3D reconstruction of pineapples.

170

Figure 31 (cont’d)

(b) Black dots showing the HU-values for the mean of 100 data points per each pineapple ROI (n = 100). Values
followed by the same letter are not significantly different between each other at P = 0.05 (ANOVA) (Tukey multiple
comparison of means). Vertical bars represent the standard deviation (SD) of each ROI.

171

Figure 32. (a) Color raw image slices, cross-sectional 2D CT images acquired using the GE BrightSpeed™ RT 16
Elite CT scanner, and 3D reconstruction of tart cherries.

172

Figure 32 (cont’d)

(b) Black dots showing the HU-values for the mean of 100 data points per each tart cherry ROI (n = 100). Values
followed by the same letter are not significantly different between each other at P = 0.05 (ANOVA) (Tukey multiple
comparison of means). Vertical bars represent the SD of each ROI.

173

Figure 33. (a) Color raw image slices, cross-sectional 2D CT images acquired using the GE BrightSpeed™ RT 16
Elite CT scanner, and 3D reconstruction of pickling cucumbers.

174

Figure 33 (cont’d)

(b) Black dots showing the HU-values for the mean of 100 data points per each pickling cucumber ROI (n = 100).
Values followed by the same letter are not significantly different between each other at P = 0.05 (ANOVA) (Tukey
multiple comparison of means). Vertical bars represent the SD of each ROI.

175

REFERENCES

176

REFERENCES

Abbott, J.A., 1999. Quality measurement of fruits and vegetables. Postharvest
Biol. and Technol. 15, 207-225.
Abrams, H.L., McNeil, B.J., 1978. Computed Tomography: Cost and Efficacy
Implications. Am J Roentgenol 131, 81-97.
Abrams, H.L., McNeil, B.J., 1978. Computed Tomography: Cost and Efficacy
Implications. Am J Roentgenol 131, 81-97.
Adams, M.R., Moss, M., 2000. Food Microbiology. Royal Society of Chemestry,
Guildford, UK.
Ahonen, T., Matas, J., He, C., Pietikäinen, M., 2009. Rotation invariant image
description with local binary pattern histogram fourier features. In: Salberg,
A.-B., Hardeberg, J.Y., Jenssen, R. (Eds.), In Proceedings of SCIA. 61-70.
Alanen, J., Keski-Nisula, L., Blanco-Sequeiros, R., Tervonen, O., 2004. Cost
comparison analysis of low-field (0.23 T) MRI- and CT-guided bone
biopsies. European Radiology 14, 123-128.
Anagnostakis, S.L., Gordon, P., Hebard, F.V., 1998. Identification of chestnut
trees. Annu. Rpt. Northern Nut Grower Assn. 89, 1-4.
Anagnostakis, S.L., Devin, P., 1999. Nutrients in chestnuts. Annu. Rpt. Northern
Nut Grower Assn. 89, 1-4.
Anjos, A.D., Shahbazkia, H., 2008. Bi-level image thresholding - A Fast Method,
In Proceedings of BIOSIGNALS. 70-76.
Ariana, D.P., Lu, R., 2010. Evaluation of internal defect and surface color of
whole pikles using hyperspectral imaging. J of Food Eng 96, 583-590.
Arnold, B.A., Xiang, P., Mao, S.S., Budoff, M.J., 2010. Peak SNR in automated
coronary calcium scoring: Selecting CT scan parameters and statistically
defined scoring thresholds. Medical Physics 37, 3621.
Badr, I., Thomas, S.M., Cotterill, A.D., Oduko, J.M., Fitzgerald, M., Adam, E.J.,
1997. X-ray pelvimetry - Which is the best technique? Clinical Radiology
52, 136-141.
Barcelon, E.G., Tojo, S., Watanabe, K., 1999a. X-ray computer tomography for
internal quality evaluation of peaches. J. Agric. Eng. Res. 73, 323-330.
177

Barcelon, E.G., Tojo, S., Watanabe, K., 1999b. Relating X-ray absorption and
some quality characteristics of mango fruit (Mangifera indica L.). J. Agric.
Food Chem. 47, 3822-3825.
Bierberle, M., Fischer, F., Schleicher, E., Koch, D., Menz, H.-J., Mayer, H.-G.,
Hampel, U., 2009. Experimental tho-phase flow measurement using ultra
fast limited-angle-type electron beam X-ray computed tomography. Exp.
Fluids 47.
Biomhoff, R., Carisen, M.H., Andersen, L.F., Jacobs, D.R., 2006. Health benefits
of nuts: potential role of antioxidants. British J. Nutr. 96, S52-S60.
Bishop, C., 2007. Pattern Recognition and Machine Learning (Information
Science and Statistics). Springer.
Blackwell, R., 2006. Chestnut Farms, Inc. Business plan. Chestnut Farms, Inc.
Allendale, Michigan: CGI.
Blasco, J., Aleixos, N., Molto, E., 2007. Computer vision detection of peel defects
in citrus by means of a region oriented segmentation algorithm. Journal of
Food Engineering 81, 535-543.
Bolliger, S.A., Oesterhelweg, L., Spendlove, D., Ross, S., Thali, M.J., 2009. Is
differentiation of frequently encountered foreign bodies in corpses possible
by hounsfield density measurement. J. Forensic Sci. 54, 1119-1122.
Boone, J.M., Geraghty, E.M., Siebert, J.A., Wooton-Gorges, S.L., 2003. Dose
reduction in pediatric CT: A rational approach. Radiology 228, 352-360.
Borges, O., Goncalves, B., Soeiro de Carvalho, J.L., Correia, O., Silva, A.P.,
2008. Nutritional quality of chestnut (Castanea sativa Mill.) cultivars from
Portugal. Food Chem. 106, 976-984.
Braga, F.R., Araujo, J.V., Soares, F.E., Araujo, J.M., Genier, H.L., Silva, A.R.,
Carvalho, R.O., Queiroz, J.H., Ferreira, S.R., 2010. Optimizing protease
production from an isolate of the nematophagous fungus Duddingtonia
flagrans using response surface methodology and its larvicidal activity on
horse cyathostomins. J Helminthol, 1-7.
Brosnan, T., Sun, D.-W., 2004. Improving quality inspection of food products by
computer vision––a review. Journal of Food Engineering 61, 3-16.
Bullerman, L.B., Bianchini, A., 2007. Stability of mycotoxins during food
processing. Int. J. of Food Microb. 119, 140-146.

178

Bushberg, J.T., Seibert, J.A., Leidholdt, E.M., Boone, J.M., 2002. The essential
physics of medical imaging, 2. Lippincott Williams & Wilkins, Philadelphia,
Pa.
Butz, P., Hofmann, C., tauscher, B., 2005. Recent developments in noninvasive
techniques for fresh fruit and vegetable internal quality analysis. J. of Food
Sci. 70, R131-R141.
Cenegage, G., Krapp, K., 2002. CT Imagin Equipment. In: Cenegage, G., Krapp,
K. (Eds.), eNotes.com. 2010. Encyclopedia of Nursing & Allied Health,
http://www.enotes.com/nursing-encyclopedia/ct-imaging-equipment.
Chai, Z., Mendez-Vazquez, H., He, R., Sun, Z., Tan, T., 2013. Semantic pixel
sets based local binary patterns for face recognition. Lecture Notes in
Computer Science 7725, 639-651.
Chen, P., Sun, Z., 1991. A review of non-destructive methods for quality
evaluation and sorting of agricultural products. J of Agr. Eng. Res. 49, 8598.
Cubero, S., Aleixos, N., Moltó, E., Gómez-Sanchis, J., Blasco, J., 2010.
Advances in Machine Vision Applications for Automatic Inspection and
Quality Evaluation of Fruits and Vegetables. Food and Bioprocess
Technology 4, 487-504.
De la Montaña Miguelez, J., Miguez Bernández, M., García Queijero, J.M., 2004.
Composition of varieties of chestnuts from Galicia (Spain). Food Chem.
84, 401-404.
Deglint, H.J., Rangayyan, R.M., Ayres, F.J., Boag, G.S., Zuffo, M.K., 2007.
Three-dimensional segmentation of the tumor in computer tomographic
images of neuroblastoma. J. of Digital Imaging 20, 72-87.
Donis-Gonzalez, I.R., Ryser, E.T., Guyer, D., Fulbright, D.W., 2010. Shell mold
and kernel decay of fresh chestnuts in Michigan. Acta Hort. 866, 353-358.
Donis-Gonzalez, I.R., Guyer, D.E., Pease, A., Fulbright, D.W., 2012. Relation of
computerized tomography Hounsfield unit measurements and internal
components of fresh chestnuts (Castanea spp.). Postharvest Biology and
Technology 64, 74-82.
Donis-González, I.R. 2008. Management of microbial decay of fresh and peeled
chestnuts in Michigan. Michigan State University East Lansing, MI, USA.
Donis-González, I.R., Mandujano, M., Medina-Mora, C., Fulbright, D.W., 2010a.
Presence of mycotoxins after 90 days of storage in fresh chestnuts. 4th
179

International Chestnut Symposium, Beijing, China. Acta Horticulturae 844,
69-74.
Donis-González, I.R., Ryser, E.T., Guyer, D., Fulbright, D.W., 2010b. Efficacy of
postharvest treatments for reduction of molds and decay in fresh Michigan
chestnuts. First European Chestnut Conference. Acta Horticulturae 866,
563-569.
Donis-González, I.R., Guyer, D.E., Leiva-Valenzuela, G.A., Burns, J., 2013.
Assessment of chestnut (Castanea spp.) slice quality using color images.
Journal of Food Engineering 115, 407-414.
Du, L.Y., Umoh, J., Nikolov, H.N., Pollmann, S.I., Lee, T.Y., Holdsworth, D.W.,
2007. A quality assurance phantom for the performance evaluation of
volumetric micro-CT systems. Phys Med Biol 52, 7087-7108.
Duda, R.O., Hart, P.E., Stork, D.G., 2000. Pattern Classification. WileyInterscience.
Ellis, M.B., Ellis, P.J., 1985. Microfungi on land plants, an identification handbook,
2. Macmillan Publishing company, New York, NY, USA.
Ertan, E., 2007. Variabiligy in leaf and fruit morphology and in fruit composition of
chestnuts (Castanea sativa Mill.) in the Nazilli region of Turkey. Genet.
Resour. Crop. Evol. 54, 691-699.
Eustice, R.E., Bruhn, C.M., 2006. Consumer acceptance and marketing of
irradiated foods. Oxford: Blackwell Publ., Ltd.
Farkas, J., Mohácsi-Farkas, C., 2011. History and future of food irradiation. Food
Science & technology 22, 121-126.
Food and Agriculture Organization Statistics Division (FAOSTAT), 2013.
Cheastnut harvested area, yield and production quantity. Jun. 20, 2013.
http://faostat.fao.org/
Ford, N.L., Thornton, M.M., Holdsworth, D.W., 2003. Fundamental image quality
limits for microcomputed tomography in small animals. Medical Physics
30, 2869.
Fromm, J.H., Sautter, I., Matthies, D., Kremer, J., Schumacher, P., Ganter, C.,
2001. Xylem water content and wood density in spruce and oak trees
detected by high-resolution computed tomography. Plant Physiology 127,
416-425.
Fulbright, D.W., Mandujano, M., 2000. Establishing chestnut orchards in
Michigan. Annu. Rpt. Northern Nut Grower Assn. 91, 1-5.
180

Furnols, M.F.I., Teran, M.F., Gispert, M., 2009. Estimation of lean meat content in
pig carcasses using C-ray Computed Tomography and PLS regression.
Chemometrics and Intelligent Lab. Syst. 98, 31-37.
Gao, Z.-j., Lin, H., Xiao, H., 2008. Air-impingement de-shelling of chestnuts (C.
mollisima): Process parameter optimization. . Internatl. J. Food
Engineering 4, 1-15.
Garvin, D., 1984. What does product quality really mean. Sloan Management
Review 26, 25-45.
General Electric Company, 2007. BrightSpeed Series: Technical Reference
Manual. Tokyo, Japan. 192.
Goetz, T., 2010. From success of the Fab Four, a key driver of health-care costs
arose. The Washington Post,
http://www.washingtonpost.com/wpdyn/content/article/2010/02/19/AR2010021905963.html.
Gold, M., Cernusca, M.M., Godsey, L., 2004. Consumer preferences for
chestnuts, eastern black walnuts, and pecans. Hort. Tech. 14, 583-589.
Goldman, L.W., 2007. Principles of CT: Radiation dose and image quality. J of
Nuclear Medicine Technology 35, 213-225.
Goldman, L.W., 2008. Principles of CT: Multislice CT*. J Nucl Med Technol 36,
57-68.
Gomes, J.F.S., Leta, F.R., 2012. Applications of computer vision techniques in
the agriculture and food industry: a review. European Food Research and
Technology 235, 989-1000.
Grau, P., France, A., 1999. Chestnut production in Chile: Some steps toward its
improvement. Acta Hort. 494, 37-42.
Gravatt, G.F., Marshall, R.P., 1926. Chestnut blight in the Southern
Appalachians. U.S.D.A. circular 370, Washington, D.C.
Guyer, D., Fulbright, D.W., Mandujano, M., 2003. A perspective on chestnut
handling and peeling. Annu. Rpt. Northern Nut Grower Assn.
94.
Guyer, D., Fulbright, D., Mandujano, M., 2005. A perspective on Chestnut
handling and peeling. Chestnut Growers of America 7, 1-5.
Hampel, U., Speck, M., Koch, D., Menz, H.-J., Mayer, H.-G., Fietz, J., Hoppe, D.,
Schlieicher, E., Zippe, C., Prasser, H.-M., 2005. Experimental ultra fast X181

ray computed tomography with a linearly scanned electron beam source.
Flow Measur. & Instr., 65-72.
Hanke, R., Fuchs, T., Uhlmann, N., 2008. X-ray based methods for nondestructive testing and material characterization. Nuclear Instr. and Meth.
in Physics Research A 591, 14-18.
Haralick, R.M., 1979. Statistical and structural approaches to texture. In
Proceedings of the IEEE 67, 786-804.
Harte, J.B., Dolan, K., Camaron, A., Fulbright, D.W., 2003. Effects of curing and
storage on fresh Chinese chestnuts quality. Annu. Rpt. Northern Nut
Grower Assn. 94, 22-31.
Holmström, L., Koistinen, P., 2010. Pattern recognition. Wiley Interdisciplinary
Reviews: Computational Statistics 9999, n/a.
Huda, W., Scalzetti, E.M., Levin, G., 2000. Techniques factors and image quality
as functions of patient weight at abdomina CT. Radiology 217, 430-435.
Jacobs, D.F., 2007. Toward development of silvical strategies for forest
restoration of American chestnut (Castanea dentata) using blight-resistant
hybrids. Biolog. conservation 137, 497-506.
Jagannath, H.S., Virmani, J., Kumar, V., 2012. Morphological enhancment of
microcalcifications in digital mammograms. J of The Intitution of Enginners
(India): Series B 93, 163-172.
Jain, A.K., Duin, R.P.W., Jianchang, M., 2000. Statistical pattern recognition: a
review. Pattern Analysis and Machine Intelligence, IEEE Transactions on
22, 4-37.
Jerimini, M., Conedera, M., Sieber, M., T.N., S., Scharer, A., Scharer, H., Jelmini,
G., 2006. Influence of fruit treatments on perishability during cold storage
of sweet chestnuts. J. Sci. of Food and Agri. 86, 877-885.
Jha, S.N., Narsaiah, K., Sharma, A.D., Singh, M., Bansal, S., Kumar, R., 2010.
Quality parameters of mango and potential of non-destructive techniques
for their measurement - a Review. J. of Food Sci. and Techn. 47, 1-14.
Kader, A.A., 2002. Posharvest technology of horticultural crops. Regents of the
University of California Division of Agriculture and Natural Resources. UCDavis, California, USA.
Kamm, K.-F., 1998. Grundlagen der Röntgenabbildung. In: Ewen, K. (Ed.),
Moderne Bildgebung: Physik, Gerätetechnik, Bildbearbeitung und 182

kommunikation, Strahlenschutz, Qualitätskontrolle. Georg Thieme Verlag,
Stuttgart & New York. pp. 45-62.
Kavdir, I., Guyer, D.E., 2007. Evaluation of different pattern recognition
techniques for apple sorting. Postharvest Biol. and Technol. 99, 211-219.
Kelley, K.M., Behe, B.K., 2002. Chefs' perceptions and uses of 'Colossal'
chestnuts. Hort. Tech. 12, 172.
Kelly, D.M., Hasegawa, I., Borders, R., Hatabu, H., Boiselle, P.M., 2003. HighResolution CT using MDCT: Comparison of Degree of Motion Artifacts
Between Columetric an Axial Methods. American J. of Roentgenology 182,
757-759.
Kim, S., Schatzki, T.F., 2000. Apple watercore sorting system using X-ray
imagery: I. Algorithm development. Trans. of the ASABE 43, 1695-1702.
Kincaid, R., Krol, A., Fourmaux, S., Kieffer, J.C., Serbanescu, C., Servol, M.,
Vogelsang, L., Wilkins, S., Stevenson, A., Nesterets, Y., Lipson, E., Ye, H.,
Pogany, A., 2008. Development of ultrafast laser-based X-ray in-vivo
phase-contrast micro-CT beamline for biomedical applications at
Advanced Laser Light Source (ALLS). Proc Soc Photo Opt Instrum Eng
7078, 707818 707811-707818 707812.
Klinac, D., Seelye, R., Nguyen, H., 1999. Chestnut production and research in
New Zealand. Acta Hort. 494, 43-48.
Kotwaliwale, N., Weckler, P.R., Brusewitz, G.H., 2006. X-ray attenuation
coefficient using polychromatic X-ray Imaging of pecan components.
Biosystems Eng. 94, 199-206.
Kumar, A., Pang, G.K.H., 2002. Defect detection in textured materials using
Gabor filters. Industry Applications, IEEE Transactions on 38, 425-440.
Lammertyn, J., Dresselaers, T., Hecke, P.V., Jacsók, P., Wever, M., Nicolaï, B.,
2003. MRI and X-ray CT study of spatial distribution of core breakdown in
'Conference' pears. Reson. Imaging 21, 805-815.
Ledenius, K., Gustavsson, M., Johansson, S., Stalhammar, F., Wiklund, L.M.,
Thilander-Klang, A., 2009. Effect of tube current on diagnostic image
quality in paediatric cerebral multidetector CT images. Br J Radiol 82, 313320.
Leiva, G., Mondragón, G., Mery, D., Aguilera, J., 2011. The automatic sorting
using image processing improves postharvest blueberries storage quality.
In Proceedings of 11th International Congress on Engineering and Food.
183

Lillard, H.Z., Hanlin, R.T., Lillard, D.A., 1970. Aflatoxigenic isolates of Aspergillus
flavus from pecans. Applied Microb. 19, 128-130.
Long, S., 2012. Internal kernel breakdown threatens the Michigan chestnut
industry - Interview with Dr. Dennis Fulbright. Midwest Nut Producers
Council Journal 1, 3.
Lorente, D., Aleixos, N., Gómez-Sanchis, J., Cubero, S., García-Navarrete, O.L.,
Blasco, J., 2011. Recent Advances and Applications of Hyperspectral
Imaging for Fruit and Vegetable Quality Assessment. Food and
Bioprocess Technology 5, 1121-1142.
Lu, R., Peng, Y., 2006. Hyperspectral scattering for assessing peach fruit
firmness. Biosystems Eng. 93, 161-171.
Mandujano, M., Nash, B., Fulbright, D.W., 1998. Ethephon treatment of chestnuts
for harvest management. Annu. Rpt. Northern Nut Grower Assn. 89, 3641.
Mateo, R., Medina, A., Mateo, E.M., Mateo, F., Jiménez, M., 2007. An overview
of Ochratoxin A in beer and wine Int. J. of Food Microb. 119, 471-475.
Mayo, J.R., Hartman, T.E., Lee, K.S., Primack, S.L., Vedal, S., Müller, N.L., 1994.
CT of the Chest: Minimal tube current required for good image quality with
the least radiation dose. AJR 164, 603-607.
McCollough, C.H., Brueswitz, M.R., Kofler, J.M., 2006. CT Dose reduction and
dose management tools: Overviews of available options. RadioGraphics
26, 503-512.
Mencarelli, F., 2001. Postharvest handling and storage of chestnuts (English).
FAO-AG-TCP/CPR/8925, Job No: AC645. Agricultural Support Systems
Div. 2003. Project: Integrated Pest Management and Storage of Chestnuts
in Xinxian County, Henan Province, China, CPR/8925, Rome, Italy.
Mendoza, F., Verboven, P., Mebatsion, H.K., Kerckhofs, G., Wevers, M., Nicolai,
B., 2007. Three-dimensional pore space quantification of apple tissue
using X-ray computed microtomography. Planta 226, 559-570.
Merkel, H.W., 1905. A deadly fungus on the American chestnut. N.Y. Zoological
Society 10, 97-103.
Mery, D. 2001. Automated Flaw Detection in Castings from Digital Radioscopic
Image Sequences - Thesis in German. Technical University of Berlin,
Berlin.

184

Mery, D., Filbert, D., 2002. Classification of Potential Defects in Automated
Inspection of Aluminium Castings Using Statistical Pattern Recognition, In
Proceedings of 8th European Conference on Non-Destructive Testing
(ECNDT 2002), Barcelona, Spain.
Mery, D., Pedreschi, F., 2005. Segmentation of colour food images using a
robust algorithm. Journal of Food Engineering 66, 353-360.
Mery, D., Soto, A., 2008. Features: the more the better, In Proceedings of the 8th
conference on Signal processing, computational geometry and artificial
vision. World Scientific and Engineering Academy and Society (WSEAS),
Rhodes, Greece.
Mery, D., Chanona-Perez, J.J., Soto, A., Aguilera, J.M., Cipriano, A., VelezRivera, N., Arzate-Vazquez, I., Gutierrez-Lopez, G.F., 2010. Quality
classification of corn tortillaz using computer vision. Journal of Food
Engingeering 101, 357-364.
Mery, D., Lillo, I., Loebel, H., Riffo, V., Soto, A., Cipriano, A., Aguilera, J.M., 2011.
Automated fish bone detection using X-ray imaging. Journal of Food
Engineering 105, 485-492.
Metcalf, H., Collins, J.F., 1909. The present status of the chestnut bark disease.
Volume 1. G.P.O., Washington.
Milczarek, R.R., Saltveit, M.E., Garvey, T.C., McCarthy, M.J., 2009. Assessment
of tomato pericarp mechanical damage using multivariate analysis of
magnetic resonance images. Postharvest Biology and Technology 52,
189-195.
Miller, G., 2003. Chestnuts. In: D.W., F. (Ed.), A Guide to Nut Tree Culture in
North America1. Northern Nut Growers Association, Inc., East Lansing,
Michigan. pp. 167-181.
Monarca, D., Cecchini, M., Antonelli, D., Mordacchini Alfani, M.L., Salcini, M.C.,
Massantini, R., 2005. Mechanical harvesting and quality of 'Marroni'
chestnut. Acta Hort. 682, 1193-1198.
Montealegre, J., Gonzalez, S., 1986. Hongos causantes de pudriciones en fruto
de Castanea sativa (Mill). Simiente 56, 166-169.
Moreda, G.P., Muñoz, M.A., Ruiz-Altisent, M., Perdigones, A., 2012. Shape
determination of horticultural produce using two-dimensional computer
vision – A review. Journal of Food Engineering 108, 245-261.
Myers, R.H., Montgomery, D.C., Anderson-Cook, C.M., 2008. Response surface
methodology: Process and product optimization using designed
experiments. John Wiley & Sons, Inc. , Hoboken, New Jersey.
185

Ng, C.B.R., Guojun, L., Dengsheng, Z., 2005a. Performance Study of Gabor
Filters and Rotation Invariant Gabor Filters, In Proceedings of the 11th
International Multimedia Modelling Conference. 158-162.
Ng, C.B.R., Guojun, L., Dengsheng, Z., 2005b. Performance Study of Gabor
Filters and Rotation Invariant Gabor Filters, Multimedia Modelling
Conference, 2005. MMM 2005. Proceedings of the 11th International. 158162.
Nixon, M.S., Aguado, A.S., 2008. Feature extraction and image processing.
Elselvier London, UK.
Ojala, T., Pietikainen, M., Maenpaa, T., 2002. Multiresolution gray-scale and
rotation invariant texture classification with local binary patterns. Pattern
Analysis and Machine Intelligence, IEEE Transactions on 24, 971-987.
Overy, D.P., Seifert, K.A., Savard, M.E., Frisvad, J.C., 2003. Spoilage fungi and
their mycotoxins in commercially marketed chestnuts. Int. J. of Food
Microb. 88, 69-77.
Paglietta, R., Bonous, G., 1979. Il castagno da frutto. Frutticoltura Moderna,
Collana di Manuali Tecnici Dell'Edagricole, 2. Italy
Patil, M.S., Mathew, J., Rajendrakumar, P.K., Karade, S., 2010. Experimental
studies using response surface methodology for condition monitoring of
ball bearings. J. of Tribology 132, 044505.
Peng, Y., Lu, R., 2008. Analysis of spatially resolved hyperspectral scattering
images for assessing apple fruit firmness and soluble solids content.
Postharves Biology and Technology 48, 52-56.
Perkins-Veasie, P., 2004. Commodity summary, raspberries, the commercial
storage of fruits, vegetables and florist nursery stock. In: Gross, K.C.,
Wang, C.Y., Saltveit, M. (Eds.), United States: US Department of
Agriculture Handbook. Volume: 66.
Pietikäinen, M., Ojala, T., Xu, Z., 2000. Rotation-invariant texture classification
using feature distributions. Pattern Recognition 33, 43-52.
Pratx, G., Xing, L., 2011. GPU computing in medical physics: A review. Medical
Physics 38, 2685.
Prokop, M., 2003. General principles of MDCT. European Journal of Radiology
45, S4-S10.

186

Qin, J., Lu, R., 2005. Detection of pits in tart cherries by hyperspectal
transmission imaging Trans. of the ASAE 48, 1963-1970.
Quanquan, G., Zhenhui, L., Jiawei, H., 2011. Linear Discriminant Dimensionality
Reduction, In Proceedings of the 21st European Conference on Machine
Learning (ECML), Athens, Greece.
Ren, Y., Liu, H., Xue, C., Yao, X., Liu, M., Fan, B., 2006. Classification study of
skin sensitizers based on support vector machine and linear discriminant
analysis. Analytica Chimica Acta 572, 272-282.
Renaudin, C., Magnin, I.E., Picard, C., Trousset, Y., Sergent, M., Amiel, M.,
1993. Image quality study in 3D X-ray angiography: a first approach using
the experimental design strategy. Comput. Med. Imaging Graph 17, 309322.
Ridé, M., Gudin, C., 1960. Sur la biologie de quelques champignons parasites de
la chataigne et plus particulierement de Phoma endogena (Speg.).
Compte Rendu de l'Academie d'Agriculture France 46, 536-543.
Ridley, J.D., 1999. Market development opportunities in the Australian chestnut
industry. Acta Hort. 494, 55-60.
Sande, E.P., Martinsen, A.C., Hole, E.O., Olerud, H.M., 2010. Interphantom and
interscanner variations for Hounsfield units-establishment of reference
values for HU in a commercial QA phantom. Phys Med Biol 55, 51235135.
Schollenberger, M., Muller, H.-M., Rufle, M., Drochner, W., 2008. Natural
occurrence of 16 Fusarium toxins in edible oil marketed in Germany. Food
Control 19, 475-482.
Sellakumar, P., James-Jebaseelan-Samuel, E., Supe, S., 2007. Investigation of
optimal scanning protocol for X-ray computed tomography polymer gel
dosimetry. Nuclear Instruments and Methods in Physics Research Section
B: Beam Interactions with Materials and Atoms 264, 395-399.
Shapiro, L., Stockman, G., 2001. Computer Vision Prentice Hall, Inc. , Upper
Saddle River, NJ, USA.
Sharma, D.S., Sharma, S.D., Sanu, K.K., Saju, S., Deshpande, D.D., Kannan, S.,
2006. Performance evaluation of a dedicated computed tomography
scanner used for virtual simulation using in-house fabricated CT
phantoms. J Med Phys 31, 28-35.
Sieber, T.N., Conedera, M., Jerimini, M., 2007. Effects of the harvest method on
the infestation of chestnuts (Castanea sativa) by Insects and Moulds.
Phytopathology 155, 497-504.
187

Silva, R.R., Siqueira, M.H.S., Calôba, L.P., 2002. Contribution to the
Development of a Radiographic Inspection Automated System, In
Proceedings of 8th European Conference on Non-Destructive Testing
(ECNDT 2002), Barcelona, Spain.
Smyth, A.B., Cameron, A.C., 1998. Modified-atmosphere packaging of cut
iceberg lettuce: Effect of temperature and oxygen partial pressure on
respiration and quality. J. of Agriculture and Food Chem. 46, 4556-4592.
Sornsrivichai, J., Yantarasri, T., Kalayanamitra, K., 2000. Nondestructive
techniques for quality evaluation of pineapple fruits. Acta Hort. 529, 337341.
Sreedhar, K., Panlal, B., 2012. Enhancement of Images Using Morphological
Transformations. International Journal of Computer Science and
Information Technology 4, 33-50.
Stadler, A., Schima, W., Prager, G., Homolka, P., Heinz, G., Saini, S.,
Eisenhuber, E., Niederle, B., 2004. CT density measurement for
characterization of adrenasal tumors ex vivo. AJR 182, 671-675.
Stuke, I., Brunke, O., 2010. Inline-CT: A new concept for the fast CT inspection in
mass production. In: Brunke, O., Neuber, D. (Eds.), High-Resolution X-ray
CT Symposium. GE Sensing & Inspection Technologies, Dresden,
Germany. 23.
Talasila, P.C., Cameron, A.C., Joles, D.W., 1994. Distribution of steady-state
Oxygen partial pressures in modified-atmosphere packages of cut
broccoli. J. of American Society Hort. and Scie. 119, 556-562.
Tanaka, K., Sago, Y., Zheng, Y., Nakagawa, H., Kushiro, M., 2007. Mycotoxins in
rice. Int. J. of Food Microb. 119, 59-66.
Teisi, M.F., Fein, S.B., Levy, A.S., 2009. Information effects on consumer
attitudes toward three food technolgies: organic production, biotechnology
and irradiation. Food Quality and Preference 20, 586-596.
Thompson, J.F., Mitchan, E.J., Mitchell, F.G., 2002. Preparation for fresh market.
In: Kader, A.A. (Ed.), Postharvest technology of horticultural crops.
University of California Agriculture and Natural Resources, Oakland, CA.
pp. 67-79.
Tsukagoshi, S., Ota, T., Fujii, M., Kazama, M., Okumura, M., Johkoh, T., 2007.
Improvement of spatial resolution in the longitudinal direction for isotropic
imaging in helical CT. Phys. Med. Biol. 52, 791-801.

188

Upchurch, B.L., Affeldt, H.A., Aneshansley, D.J., Birth, G.S., Cavalieri, R.P.,
Chen, P., W.M. Miller, Sagig, Y., Schmilovitch, Z., Throop, J.A., Tollner,
W.W., 1993. Detection of Internal Disorders, Proceedings of the
International Workshop funded by the United States-Israel Binational
Agricultural Research and Development Fund (BARD). American Society
of Agricultural Engineers, Sponkane, Washington 5, 80-85.
van-Daatselaar, A.N., van-der-Stelt, P.F., Weenen, J., 2004. Effect of number of
projections on image quality of local CT. Dentomaxillofacial Radiology 33,
361-369.
Varga, L., Koncz, Z., Kocsube, S., Matrai, T., Teren, J., Ostry, V., 2007.
Mycobiota of grapes collected in Hungarian and Czech vineyards in 2004.
Acta Alimentaria 36, 329-341.
Verboven, P., Kerckhofs, G., Mebatsion, H.K., Ho, Q.T., Temst, K., Wevers, M.,
Cloetens, P., Nicolai, B.M., 2008. Three-dimensional gas exchange
pathways in pome fruit characterized by synchrotron X-ray computed
tomography. Plant Physiology 147, 518-527.
Vettraino, A.M., Paolacci, A., Vannini, A., 2005. Endophytism of Sclerotinia
pseudotuberosa: PCR assay for specific detection in chestnut tissue.
Mycol. Res. 109, 96-102.
Vossen, P., 2000. Chestnut culture in California. UCDAVIS-Publication 8010, 16
California.
Wang, D.C.C., Vagnucci, A.H., Li, C.C., 1983. Digital image enhancement: A
Survey Computer vision, graphics, and image processing 24, 363-381.
Wang, T., Liang, H., Yuan, Q., 2011. Optimization of ultrasonic-stimulated solvent
extraction of sinigrin from Indian mustard seed (Brassica Juncea L.) using
response surface methodology. Phytochem Anal 22, 205-213.
Wells, J.M., 1980. Mycoflora and Market quality of chestnuts treated with hot
water to control the chestnut weevil. Plant Dis. Rep. 64, 999-1001.
Willis, R., McGlasson, B., Joyce, D., 2007. Postharvest. An introduction to the
physiology and handling of fruit, vegetagle and ornamentals. CABI, South
Wales, Australia.
Wirth,

M., Fraschini, M., Lyon, J., 2004. Contrast enhancement of
microcalcifications in mammograms using morphological enhancement

189

and non-flat structuring elements In Proceedings of 17th IEEE Symposium on
Computer-Based Medical Systems (CBMS'04). 134-139.
Woodford, C., Yartsev, S., Van Dyk, J., 2007. Optimization of megavoltage CT
scan registration settings for brain cancer treatments on tomotherapy.
Phys Med Biol 52, N185-193.
Wulfshohn, D., Sarig, Y., Algazi, R.V., 1993. Defect sorting of dry dates by
image-analysis. Canadian Agr. Eng. 35, 133-139.
Zhang, J., 2002. Invariant Texture Segmentation Via Circular Gabor Filters. In:
Tieniu, T., Li, M. (Eds.). 20901-20901.
Zhu, B., Jiang, L., Luo, Y., Tao, Y., 2007. Gabor feature-based apple quality
inspection using kernel principal component analysis. Journal of Food
Engineering 81, 741-749.
Zimmerman, J.B., Pizer, S.M., Staab, E.V., Perry, J.R., Mccartney, W., Brenton,
B.C., 1998. An evaluation of the effectineness of adaptive histogram
equalization for contrast enhancment IEEE Trans of Medical Imaging 7,
304-312.

190