Diagnosis of Neurodegenerative Diseases using Deep Learning

Automated disease classification systems can assist radiologists by reducing workload while initiating therapy to slow disease progression and improve patients’ quality of life. With significant advances in machine learning (ML) and medical scanning over the last decade, medical image analysis has experienced a paradigm change. Deep learning (DL) employing magnetic resonance imaging (MRI) has become a prominent method for computer-assisted systems because of its ability to extract high-level features via local connection, weight sharing, and spatial invariance. Nonetheless, there are several important research challenges when advancing toward clinical application, and these problems inspire the contributions presented throughout this thesis. This research develops a framework for the classification of neurodegenerative diseases using DL techniques and MRI. The presented thesis involves three evolution stages. The first stage is the development of a robust and reproducible 2D classification system with high generalisation performance for Alzheimer’s disease (AD), mild cognitive impairment (MCI), and Parkinson’s disease (PD) using deep convolutional neural networks (CNN). The next phase of the first stage extends this framework and demonstrates its use on different datasets while quantifying the effect of a highly observed phenomenon called data leakage in the literature. Key contributions of the thesis presented in this stage are a thorough analysis of the literature, a discussion on the potential flaws of the selected studies, and the development of an open-source evaluation system for neurodegenerative disease classification using structural MRI. The second stage aims to overcome the problems stem from investigating 3D data with 2D models. With this goal, a 3D CNN-based diagnostic framework is developed for classifying AD and PD patients from healthy controls using T1-weighted brain MRI data. The last stage includes two phases with a focus on AD and MCI diagnosis. The first phase proposes a new autoencoder-based deep neural network structure by integrating supervised prediction and unsupervised representation. The second phase introduces the final contribution of the thesis which is a novel ensemble approach that may also be used to predict diseases other than neurodegenerative ones (e.g., tuberculosis (TB)) using a modality apart from MRI.