Deep Learning Methods for Abnormality Detection and Segmentation in Computed Tomography and Magnetic Resonance Images

Abstract

In this dissertation, I describe development of novel deep learning (DL)-based methodologies for the detection and segmentation of clinical abnormalities including renal lesions, left ventricle (LV) scar, and prostate cancer (PCa) in 3D computed tomography (CT) scans and magnetic resonance (MR) images. In the first phase, I developed a decision fusion of patch-based convolutional neural network (CNN) for renal masses classification into cyst versus solid. The solid renal masses were then categorized into benign and malignant using an image-based CNN. These approaches were selected to capture local and global features of the renal masses including intensity, texture, shape, and size, which are key features used by radiologists to classify renal masses. Results demonstrated that automated assessment of renal mass with moderate-to-high degrees of accuracy is feasible. In the second phase, I designed a novel algorithm that comprehensively learns and integrates inter- and intra-slice features from 3D late gadolinium enhancement (LGE)-MR images and allows to accurately and efficiently delineate LV scar fully automatically. In the proposed method, three U-Nets were trained using LGE-MR images extracted from transversal, coronal, and sagittal directions to learn the description of LV scar from different views and the predicted results were combined through majority voting system for the final segmentation. This algorithm benefited from isotropic property of the voxels in 3D MRI that permits for multiplanar reformation. In the third phase, I described a U-Net-based methodology to segment prostate zones from T2-weighted (T2W) and apparent diffusion coefficient (ADC) map prostate MR images as a fundamental requirement for automated diagnosis of PCa. This work is the first attempt for prostate zonal segmentation using ADC map MR images. Furthermore, I presented an ensemble learning system for fully automated localization of peripheral zone PCa from the ADC map. The ensemble learning model allowed accurate PCa detection, which was not possible using one network due to the complexity of the decision boundary. Our results confirmed that automated PCa detection and segmentation using ADC map MR image are feasible, highly sensitive and can be performed rapidly.