Multimodal Deep Learning Methods for Alzheimer's Disease Diagnosis

Abstract

Alzheimer’s disease is the leading cause of dementia worldwide, and its accurate diagnosis is essential for the effectiveness of the therapies currently available. Deep learning methods applied to neuroimaging data have shown promise for this task, but most multimodal approaches assume that all modalities are available for every subject, an assumption that rarely holds in clinical practice or even in research cohorts such as ADNI, where FDG-PET is far less common than MRI. This thesis develops a multimodal framework for the three-class classification of subjects into cognitively normal, mild cognitive impairment and Alzheimer’s disease, which is natively robust to missing modalities. We formulate a Product-of-Experts model in which MRI and FDG- PET each contribute a Gaussian expert to a shared latent representation of the brain state, allowing predictions to be made from either modality alone or from both, always through the same classifier and without imputing the missing modality. Since different modality-availability patterns induce different latent-input regimes for this shared classifier, we further investigate two regularization strategies: a meta-learned regularizer, inspired by domain-generalization methods, and a Kullback-Leibler term that encourages the unimodal and multimodal latent distributions to remain compatible. Experiments on ADNI data show that jointly training the model with a standard supervised objective does not by itself improve over dedicated single-modality baselines, and that the latent representations induced by different modality subsets are clearly separated. Both regularizers improve on this basic model, with the meta-learned regularizer providing the most consistent gains across the three inference pathways. In the multimodal setting, the proposed model reaches a balanced accuracy comparable to that of a recent state-of-the-art transformer-based method, while using roughly two orders of magnitude fewer trainable parameters and remaining usable when one modality is absent at inference time.