Deep learning-based Alzheimer’s disease diagnostics using medical images

Abstract

In the ever-evolving landscape of medical innovation, the pursuit of accurate and efficient diagnostic tools for neurodegenerative diseases, notably Alzheimer's, has become increasingly urgent. The relentless progression of medical imaging technology, particularly magnetic resonance imaging (MRI), has sparked a wave of research aimed at harnessing advanced computational techniques, such as deep learning, to enhance diagnostic accuracy and efficiency. This thesis delves into the exploration of innovative deep-learning methodologies tailored specifically for Alzheimer's disease diagnosis based on MRI data, with a focus on three distinct sections of the brain: Sagittal, Axial, and Coronal. The research journey begins with the development of a 2D slice-based convolutional neural network (CNN) model, meticulously designed to analyze MRI scans in each section individually. Notably, the Sagittal section achieves an accuracy of 44%, the Axial section achieves 47%, and the Coronal section achieves 48%. As the study progresses, attention turns towards the utilization of 3D CNNs, a more holistic approach capable of processing MRI volumes as three-dimensional entities. This model, accommodating the depth of MRI data, presents a novel perspective on Alzheimer's disease diagnosis, capturing intricate spatial relationships within the brain. However, despite its promise, the 3D CNN model exhibits varying degrees of accuracy across different sections, with an overall accuracy of 60%. Furthermore, the research delves into the exploration of a patch-based classification model, aiming to enhance diagnostic accuracy through a localized analysis of MRI data. By dividing MRI volumes into smaller patches and classifying each patch individually, this approach offers a fine-grained understanding of pathological changes associated with Alzheimer's disease. However, despite its potential, the patch-based model demonstrates challenges in achieving comparable accuracy to its slice-based counterparts, with an overall accuracy of 52%. Moreover, the evaluation of these models reveals intriguing insights into the strengths and limitations of different approaches. The study explores both slice-based 2D CNN and 3D CNN approaches for MRI analysis. In the 3D CNN approach, the MRI volumes are resized to a uniform shape, accommodating variations in MRI depth across different datasets. This diverse dataset facilitates the development of models capable of accurately distinguishing between different disease stages, thereby aiding in early diagnosis and personalized treatment planning. In conclusion, this thesis contributes to the burgeoning field of medical image analysis by presenting innovative deep learning approaches tailored for Alzheimer's disease diagnostics. Through the integration of advanced algorithms and comprehensive evaluation, this research paves the way for more reliable and efficient diagnostic tools, ultimately leading to improved patient outcomes and healthcare management in neurodegenerative disease.