Towards Reliable Alzheimer’s Diagnosis From 3D MRI Scans: A Generalized Approach

This thesis aims to address Alzheimer’s disease detection from 3D MRI scans undera single-domain generalization setting, where a model is expected to generalizeto unseen domains with potentially diverse imaging protocols, patient demographics,and class imbalance levels. Three distinct approaches are investigated. First,a pseudo-morphological augmentation strategy uses learnable modules to produceanatomically coherent, class-specific augmentations, integrated with supervised contrastivelearning to extract robust and discriminative features. Second, the MixStyleframework is extended to incorporate higher-order statistical moments includingskewness and kurtosis alongside traditional mean and variance, enabling enhancedfeature perturbation and focus on disease-specific artifacts. Third, a Mixup-basedaugmentation method leverages distance transforms to spatially decompose MRIscans into layered components and recompose them from multiple samples, preservingstructural integrity while promoting diversity. Extensive experiments acrossthree benchmark datasets, namely NACC, ADNI and AIBL demonstrate that theproposed techniques substantially enhance the generalization capabilities of underlyingmodels, thus providing a strong basis for creating reliable, domain-agnostictools for early Alzheimer’s disease dia