Towards Improved Medical Image Segmentation Using Deep Learning

Convolutional neural networks have been asserted to be fast and precise frameworks with great potential in image segmentation. Within the medical domain, image segmentation is a pre-cursor to several applications including surgical simulations, treatment planning and patient prognosis. In this thesis, we attempt to solve two major limitations of current segmentation practices: 1) dealing with unbalanced classes and 2) dealing with multiple modalities. In medical imaging, unbalanced classes present as the regions of interest that are typically significantly smaller in volume than the background class or other classes. We propose an improvement to the current gold standard cost function to boost the focus of the network to the smaller classes. Another problem within medical imaging is the variation in both anatomy and pathology across patients. Utilizing multiple imaging modalities provides complementary, segmentation-specific information and is commonly employed by radiologists when contouring data. We propose a image fusion strategy for multi-modal data that uses the variation in modality specific features to guide the task specific learning. Together, our contributions propose a framework to maximize the representational power of the dataset using models with less complexity and higher generalizability. Our contributions outperform baseline models for multi-class segmentation and are modular enough to be scaled up to deeper networks. We demonstrate the effectiveness of the proposed cost function and multimodal framework, both individually and together, on benchmark datasets including the Breast Ultrasound Dataset B (BUS) [1], the International Skin Imaging Collaboration (ISIC 2018) [2], [3] and the Brain Tumor Segmentation Challenge (BraTs 2018) [4]. In all experiments, the proposed methods match or outperform the baseline methods while employing simpler networks