MACHINE LEARNED DATA AUGMENTATION TECHNIQUES FOR IMPROVING PATHOLOGY OBJECT DETECTION

Artificial intelligence (AI) has evolved immensely in recent years, with AI achieving human levels of performance on a wide variety of tasks. However, AI has had limited adoption in clinical settings despite its promising prediction, classification and pathology detection applications. For a machine learned (ML) model to train effectively, the observed data must be a diverse, accurate representation of the true distribution. Therefore, to properly estimate the true distribution, extremely large datasets become necessary. In healthcare scenarios, datasets of sufficient size may be rare or absent, thus hindering the training of ML models. One of the ways to mitigate this problem is through data augmentation, where we supplement our datasets with slightly modified copies of already existing data or newly created synthetic data. Recently, sophisticated data augmentation methods are based on a class of neural networks (NNs) called Generative Adversarial Networks (GANs), which can generate new images of high perceptual quality. This dissertation describes the design and development of a new type of GAN, named near-pair patch cycleGAN (NPP-cycleGAN), which generates realistic pathology-present images. Here, we train and test this network using pediatric chest radiographs. We demonstrate that the proposed GAN can generate high quality fracture-present pediatric chest radiographs. With the addition of these synthetic images to an object detector’s training dataset, we are able to improve the fracture detection performance. These results suggest that our proposed method can be applied to other pathology detection tasks and could potentially enable improved object detector performance in multiple clinical scenarios.