NeuroFusion: A Multimodal Transformer Framework for Early and Explainable Detection of Alzheimer’s Disease

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative disorder, and a great deal of pathological changes can be observed in the early stages of the disease before significant cognitive function is clinically evident. In this study, a multimodal Transformer framework named NeuroFusion for early and explainable detection of AD is suggested, which integrates structural magnetic resonance imaging (MRI), positron emission tomography (PET), and clinical information. The framework consists of modality-specific encoders to obtain complementary representations, and cross-modal attention to uncover relationships among anatomical, metabolic, cognitive and demographic features. The architecture proposed aims for the classification of cognitively normal individuals, mild cognitive impairment (MCI), AD patients, and for an easy-to-understand patient-level prediction. An explainable artificial intelligence component is used to determine the influential brain regions and clinical parameters using imaging attribution and feature analysis by SHAP. The illustrative evaluation shows that the unimodal and conventional multimodal configurations are outperformed by NeuroFusion, with a ROC-AUC of 0.961, and ablation analysis reveals the contribution of MRI, PET, clinical information and of cross-modal attention. The framework also includes calibration and explainability assessment which enhance the reliability of the framework. In summary, NeuroFusion offers a holistic solution that integrates multimodal learning, early-stage detection, and interpretability, and has the potential to yield more transparent AI-aided Alzheimer's assessment. The numerical results need to be validated through an empirical study on an actual trained model and an independent set of data.