Application of Deep Learning and Cortical Flat Mapping for the early detection of Alzheimer's Disease

Abstract

This thesis explores the application of deep learning to neuroimaging data for the purpose of detecting Alzheimer's disease. Initially designed for small object de- tection, our deep learning and visualisation methodologies were extended to multi- dimensional Cortical Flat Maps extracted from neuroimaging data. The initial phase involved developing deep learning models and visualisation tech- niques for small object detection, validated across diverse datasets ranging from spiders to road signs. Through extensive experimentation, we identified VGG16 as the most accurate model for this data type, achieving an 84% test accuracy com- pared to Faster RCNN with 75% test accuracy. Additionally, our novel visualisation method Feature Activation Mapping (FAM) surpassed the performance of the stan- dard CAM technique. These experiments informed the selection of suitable deep learning and visualisation methods for neuroimaging data. In our neuroimaging experiments, we introduced a 2D mapping approach to capture multi dimensional cortical information, including thickness, surface area, and cur- vature, a first application of deep learning to neuroimaging data of this kind. The first neuroimaging experiment focused on applying previously developed deep networks to a one-dimensional Cortical Flat Map representing cortical thickness, achieving a 75% accuracy in classifying sex (male/female) using a CNN with an incorporated Support Vector Machine (SVM) layer. The second neuroimaging experiment extended the deep learning methods to a three-dimensional cortical flat map, including curvature and surface area. Applied to neuroimaging data for disease classification, specifically the detection of early Alzheimer's Disease (Mild Cognitive Impairment), this approach demonstrated in- creased accuracy compared to two-dimensional brain slice data. In the final neuroimaging experiment, a novel method for identifying Alzheimer's disease related brain regions was developed, combining deep learning methods, the novel flat mapping approach, and a masking method. We identified brain regions that may be related to the early detection of Alzheimer's disease. To summarise, this thesis introduces three novel methods tailored for the application of deep learning techniques in neuroimaging data for disease classification. Prelimi- nary results are promising, suggesting the potential clinical utility of these domain- specific deep learning methodologies with further refinement and exploration. Show more