Abstract
Alzheimer’s disease (AD) is a neurodegenerative disease that has detrimental impacts on many aspects of modern society (e.g., healthcare systems, communities, families, and elderly individuals). There is a critical need to improve AD diagnoses so that the disease can be detected in an accurate and timely manner. Such diagnoses will result in potential improvements to the treatment, understanding, management, and prevention of AD. There are many cutting-edge approaches to diagnosing AD in the research literature. One important method is the use of functional magnetic resonance imaging (fMRI) and deep learning to create highly accurate AD classification models. fMRI is a neuroimaging technique that can measure functional changes in the brain (e.g., brain activity), while deep learning is an artificial intelligence (AI) method that can perform complex analyses using neural networks. fMRI is important for diagnosing AD as it can non-invasively detect unique biomarkers that are present in the early stages of the disease. Deep learning can be paired with fMRI to create highly accurate and semi-automated classification (i.e., diagnosis) models that could be applied to clinical practices.However, there are some problems with fMRI deep learning models for AD classification. Some notable problems include the lacking understanding of the field (e.g., best methods, approaches, datasets, and architectures), the analytical complexity of fMRI deep learning models, data accessibility, and the lacking clinical viability of classification models. In this thesis, I explore and address these issues in four distinct projects that synthesise the literature and classify AD using novel deep learning techniques. In the first project (Warren & Moustafa, 2023), I conduct a systematic review of the fMRI, AD, and deep learning literature to understand the current state of the field. The results of this systematic review revealed gaps in the field that I address in the remaining three projects.
Specifically, In the second project (Warren, Khan, & Moustafa, 2024), I create a guide to assistive tools for fMRI deep learning pipelines (e.g., methods for automating and streamlining models). This guide provides methods for reducing the analytical complexity and increasing the accessibility of fMRI and deep learning techniques. The guide also includes an example pipeline to display the use of assistive tools and encourage new researchers to enter the field. This example assistive tools pipeline was semi-automated and could classify participants with 71% accuracy. In the third project, I create an fMRI deep learning model to classify AD with a small sample and a single fMRI brain scan per participant. This model addresses the issue of data accessibility by using transfer learning to make a model that could classify a small AD sample (N=64) with 77% accuracy. This transfer learning model also addresses aspects of clinical viability by only using a single fMRI per participant, processing a full 3D fMRI volume at a time (thus reducing the need for additional image processing), and using automated methods. In the fourth and final project, I create a multimodal pipeline that combines fMRI and clinical data in a deep learning model. This model also works on a small sample (N=50), is flexible to multiple data types, and does not require a large training sample (like transfer learning). Thus, the model addresses issues with data accessibility and the need for clinically viable methods. The resulting multimodal model has a high accuracy of 90% and explains which clinical features are important for performing AD classification. Consequently, this thesis makes significant original contributions to the understanding, clinical viability, accessibility, refinement, and complexity of fMRI deep learning models for AD diagnoses.
| Date of Award | 13 Feb 2025 |
|---|---|
| Original language | English |
| Supervisor | Ahmed Moustafa (Supervisor) & Victor Schinazi (Supervisor) |