Prostate cancer is one of the most common and dreadful type of cancer in men. Due to the unclear symptoms of the disease, the diagnosis of prostate cancer is difficult, and requires multiple procedures. Among these procedures, the most important one is the examination of the prostate tissue biopsy to detect the presence of cancer regions in the tissue, and assign a Gleason score to the tissue (which determines the severity of the cancer). The detection and grading processes are based on the glandular structures as well as the cytological properties of the tissue. In a traditional examination, pathologists have to look at the tissue biopsy under a microscope. With the developments in digital pathology, especially in virtual microscopy, glass tissue slides can be digitized to generate tissue images. These images can be displayed on a monitor, annotated by software tools, and forwarded to experts for examination and diagnosis. However, the large volume of tissue images that are generated poses a challenge for pathologists to efficiently and accurately perform the diagnosis. Hence, there is a need to develop tools for automatic processing of prostate tissue images, which can assist pathologists in decision making and improve the throughput. This thesis deals with design and development of automatic tools for processing and analyzing prostate tissue images.In tissue examination, the grading of tissue slides is a standard procedure to determine the severity of cancer. The most popular grading method is the Gleason grading, which relies on the gland structures in the tissue to assign a Gleason score ranging from 2 to 10 to the tissue image. We utilize the Gleason grading method in automated systems by segmenting glands from the tissue image and extracting features to discriminate them. By thorough analyses and evaluations, we demonstrate that the proposed methods lead to better gland classification accuracies than published methods in the literature.We utilize the proposed gland segmentation and gland feature extraction methods to solve the tissue image classification problem, which receives the most attraction in the literature. By comparing with popular texture-based methods, we show that using the proposed gland features is a better solution for this problem. To further improve the Gleason grade 3 vs Gleason grade 4 classification result, we propose a different approach for gland segmentation and study the properties of the nuclei arrangement in the segmented glands.When a medical laboratory technician or a medical student who is not very experienced with Gleason grading wants to gain additional experience in the grading process, it will be useful if there was an image retrieval engine that could search for tissue regions similar to the region of interest (ROI) in the tissue slides that were annotated by experienced pathologists. The technician (or student) can use the retrieved regions (whose Gleason grades are known) as the references to grade the ROI. To create such an image retrieval search engine, we develop a gland-based method to compute the similarity between two tissue regions. The proposed region retrieval method, which is another application of the gland segmentation and gland feature extraction framework, obtains better performance than the methods presented in medical image retrieval.Besides gland structure information, cytological features of the prostate tissue (which are not used in Gleason grading) also provide useful information for pathologists to detect cancer. Cytological features refer to size, shape, quantity, and arrangement of the basic elements of the tissue such as cell, cytoplasm, nuclei, and nucleoli. One of the most important cytological features is that in cancer glands, nuclei usually contain nucleoli, while nuclei in normal glands do not. To utilize this information, we present a novel method to identify the nuclei with prominent nucleoli (NwN) in the tissue. By applying the result of NwN identification in a prostate cancer detection problem, we see that the use of cytological feature, i.e., the presence of NwN in this case, helps to boost the cancer detection accuracy.With the proposed solutions for different problems in automated prostate cancer grading presented in this thesis, we believe that we are able to provide pathologists with useful tools to assist them with the prostate cancer diagnosis task.
Read
