Generative and Explainable Deep Learning for Alzheimer’s Disease Analysis

Alzheimer's disease (AD) is a growing clinical challenge that demands earlier and more reliable computational biomarkers. Although recent deep learning systems have reported promising results in AD analysis, their clinical translation remains constrained by three recurring limitations: insufficient interpretability, limited generalizability across cohorts, and the scarcity of labelled medical data. This thesis investigates these issues across two complementary diagnostic modalities, handwriting and neuroimaging, with the goal of developing methods that are clinically aligned, methodologically rigorous, and transferable. The first modality considered is handwriting, explored as a preliminary and non-invasive source of diagnostic information for AD. A unified multi-task framework is developed to learn shared patterns across 25 heterogeneous writing tasks acquired from scanned paper sheets in a dataset specifically designed for diagnosis. The results show that a single model can support automated AD discrimination with promising accuracy and that transfer learning from large-scale natural image datasets remains effective even in this underexplored setting. At the same time, this contribution should be interpreted as a first step: future studies will require richer datasets that better control relevant confounders, such as arthritis and other age-related motor impairments, and longitudinal data that enable investigation of disease onset and progression rather than diagnosis alone. The second modality is structural MRI, where the thesis addresses the core challenges more directly. First, the AXIAL framework tackles interpretability within a supervised classification setting through multi-plane attention fusion and both quantitative and qualitative explainability validation. Rather than relying on post-hoc visualisation methods, AXIAL embeds attention into the diagnostic pipeline itself and produces fold-consistent 3D attention maps that align with established AD-related neuroanatomical markers, including hippocampal atrophy, medial temporal lobe degeneration, and ventricular enlargement. AXIAL shows that interpretability and diagnostic performance are not mutually exclusive, and that parameter-efficient 2D/2.5D designs can recover clinically meaningful volumetric information under limited-data conditions. However, AXIAL still depends on labelled cohorts and task-specific supervision. To address the remaining limitations of annotation scarcity and cross-dataset transfer, the thesis then moves from supervised classification to unsupervised representation learning. This progression culminates in the Latent Diffusion Autoencoder (LDAE), a generative foundation model that performs diffusion in a compressed latent space. By learning from unlabeled MRI data, LDAE decouples representation learning from downstream supervision and enables zero-shot transfer to unseen datasets. A single pre-training stage yields a reusable representation that supports multiple downstream tasks, including diagnosis, age prediction, temporal interpolation, and counterfactual generation, while remaining substantially more computationally efficient than voxel-space diffusion autoencoders. Taken together, these contributions support three main claims. First, non-invasive modalities such as handwriting deserve further investigation as potentially useful sources of AD-related information, but stronger evidence will require better-controlled and longitudinal cohorts. Second, interpretability and methodological rigor are prerequisites for clinical trust, often more important than marginal improvements in benchmark performance. Third, diffusion-based unsupervised learning provides a practical route toward richer and more transferable representations when labelled data are limited, while also enabling generative capabilities beyond purely discriminative models. Overall, the thesis argues that AI systems for Alzheimer's disease should prioritize clinical alignment, transparency, and representation quality over isolated performance gains.