Analyzing Deep Learning Approaches in Unimodal and Multimodal Frameworks for Alzheimer’s Disease Stage Classification

Abstract

In this paper, we present a detailed study on deep learning approaches for the binary classification of Alzheimer’s disease stages using both unimodal and multimodal frameworks. Alzheimer’s disease is a progressive neurodegenerative brain disorder and the leading cause of dementia, characterized by cognitive decline, memory loss, and behavioral changes due to amyloid plaques and tau tangles. It typically develops in stages, from mild memory lapses to severe cognitive impairment, and while incurable, treatments can temporarily manage symptoms. We effectively utilised PET and MRI scans, as well as CSF biomarkers, demographics and genetic data, widely used in medical settings, for developing both unimodal and multimodal frameworks that can effectively classify MCI and AD stages of Alzheimer’s disease. The traditional diagnostic methods widely use only single modularity for the diagnostic purpose however multimodal effectively utilises all three modality of the data. The unimodal models are implemented using a 3D ResNet-18 architecture, while the multimodal system incorporates modality-specific encoders, cross-modal attention mechanisms, and a gated fusion strategy to combine heterogeneous modality/data sources. We observed that for unimodal setup PET scans have higher AUC of 0.915 than that of MRI 0.84 and performs better for stage classification, It is because PET scans allow for the evaluation of brain metabolism (FDG-PET) and specific neuropathology, which helps distinguish Alzheimer’s from other dementias, enabling earlier diagnosis and treatment. MRI scans is also a widely used diagnostic method for Alzheimer’s disease which particularly detects brain shrinkage (atrophy), particularly in the hippocampus region. The multimodal architecture combines different modality which complements and compensates for missing information and provides a comprehensive view of disease pathology, and performs the best among all the models with the AUC of 0.940. Integrating multiple modalities can also assist us with differential diagnosis, such as distinguishing Alzheimer’s from other forms of dementia or co-existing conditions. Thus, we implemented three different kinds of models, utilising different modalities to provide a comprehensive study for the binary classification of stages of Alzheimer’s.