Explainable Multimodal Framework for Brain Imaging with 3D MRI and Age Data

Advanced volumetric neuroimaging modalities, particularly structural Magnetic Resonance Imaging (sMRI), are widely utilized in the clinical assessment of neurodegenerative conditions such as Alzheimer’s Disease (AD). Neuroradiologists examine three-dimensional sMRI datasets to detect and evaluate morphological alterations in the brain linked to AD, encompassing both widespread and region-specific atrophy, along with changes in the morphology of certain brain structures. Recent studies have progressively concentrated on creating diagnostic frameworks that employ Deep Learning (DL) techniques to analyze sMRI data for the identification of AD. However, it remains essential to interpret the anatomical characteristics obtained from sMRI within the context of the patient’s age to distinguish between AD-associated patterns and those resulting from typical aging processes. In this context, part-prototype neural networks present a viable strategy by merging the computational capabilities of DL with a transparent architecture, showing promise in medical imaging applications. In this study, we present ADNet, a novel interpretable multimodal system that combines three-dimensional imaging information with demographic data for the binary classification of AD, utilizing both 3D sMRI and patient age. Although the integration of age-informed prototypes does not significantly enhance the predictive accuracy beyond that of unimodal models, this research lays the groundwork for the future development of model architectures and multimodal prototype training approaches in forthcoming investigations.