Artificial intelligence based diagnosis of Alzheimer's Disease

Abstract

Accurate and early diagnosis of Alzheimer's Disease (AD) remains challenging due to limitations in current diagnostic methods, including subjectivity in cognitive assessments and restricted accessibility of certain imaging modalities like Fluorodeoxyglucose Positron Emission Tomography (FDG-PET). This thesis addresses these gaps by developing advanced machine learning (ML) frameworks that synthesize FDG-PET images from widely available structural Magnetic Resonance Imaging (sMRI) data and further evaluating the framework on local datasets for generalizability concerns. Apart from image view, the graph view is also explored in this thesis, including both brain connectivity and population-based relationships, which are important for AD diagnosis while remaining underexplored.

A Generative Adversarial Network (GAN) model and a two-way diffusion model are proposed to synthesize FDG-PET from sMRI, achieving high image similarity and improved classification accuracy when combined with sMRI, with a maximum accuracy improvement achieved by the GAN model's synthetic results on public datasets from 82.68% to 83.56% for AD/Mild Cognitive Impairment (MCI)/Normal Control (NC) classification. These synthesis methods are validated on local clinical datasets, demonstrating generalizability across diverse populations, with maximum accuracy improvement achieved by the diffusion model from 87% to 91% and from 85% to 89% on two local datasets accordingly for AD/NC classification.

Additionally, graph-based ML models are developed to leverage brain structural connectivity and population-level information. The proposed Brain Ensemble Network (BrainEnsNet) and Population Ensemble Network (PopEnsNet) explore the AD diagnosis from the view of graphs, with PopEnsNet achieving a superior prediction accuracy of 82.03% for AD/MCI/NC classification. According to experiments, the PopEnsNet surpassed BrainEnsNet in prediction accuracy and generalization ability, while the BrainEnsNet surpassed the model utilizing anatomical features only, which indicates the potential of introducing brain connectivity and population-level relationships in enhancing AD diagnosis.

Date of Award	13 Jul 2025
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Ying Weng (Supervisor), Guokun Zuo (Supervisor), Tom Dening (Supervisor) & Akram A. Hosseini (Supervisor)