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

Medical imaging is vital to non-invasive diagnosis and prognosis of medical abnormalities.Medical image acquisition has greatly advanced in terms of acquisition speed and the ability to resolve fine objects over the last decades.However, advances in technology have increased the size of the image and number of images to be interpreted by radiologists.Imaging studies of a single patient may now consist of hundreds of images, reformatted in multiple imaging planes with three-dimensional (3D) reconstruction.Detection of abnormalities, such as cancer or scar tissue, is an important part of disease diagnosis based on medical images.Abnormalities in tissue may manifest as differences in image intensity, contrast, and texture to the normal tissue in medical images.Currently, the medical images are interpreted manually by clinical experts, which is a tedious task and subject to large inter-and intra-operator variability due to observer limitations (e.g., constrained human visual perception, fatigue, or distraction) and the complexity of the clinical cases themselves (e.g., overlapping structures).Therefore, automated analysis of medical images is highly desirable.Parallel to the developments in imaging hardware, machine learning technologies, including deep learning (DL) methods, have evolved over the last decade and are providing exciting solutions in identification, classification, and quantification of abnormalities in medical images.In this dissertation, with the availability of the unique datasets of kidney computed tomography (CT) scans, prostate and cardiac magnetic resonance images (MRI) through a clinical collaboration with the Ottawa Hospital and the Libin Cardiovascular Institute of Alberta at the University of Calgary, I have focused on the computer-aided detection of kidney and prostate tumors and cardiac scar tissue as medical abnormalities.The goal of this dissertation is to describe the development of novel DL-based methodologies for the detection and segmentation of abnormalities in 3D CT scans and MR images for three high-impact clinical v applications.These applications are computer-aided detection of kidney tumors (renal masses) on CT scans, quantification of scar tissue in the heart in 3D cardiac MRI, and prostate tumor localization in multi-slice MRI.Our research has novelty in both methodology and clinical applications.For the application of detecting kidney cancer, I developed a decision fusion of convolutional neural network (CNN)-based method for renal masses classification into cyst versus solid and then categorized solid renal masses into benign and malignant.For the application of detecting scar tissue in the heart, 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 delineate LV scar fully automatically.For the application of detecting prostate cancer (PCa), 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.Furthermore, I presented an ensemble learning system for fully automated localization of peripheral zone PCa from the ADC map that has not been described previously.In this dissertation the method developed for different applications progressed from a CNN to the cascaded multi-planar U-Net and ensemble learning system.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.