A Deep Fusion Network for Multiclass Classification of Alzheimer's Disease Using Multimodal Neuroimaging

Abstract

Alzheimer’s disease (AD), the leading cause of dementia, lacks robust tools for early diagnosis, delaying therapeutic interventions. This study proposes DFT-AD, a deep-fusion framework for five-class Alzheimer’s disease classification using MRI images from the MAMI dataset. The five classes are Very Mild, Mild, Moderate, Severe, and None. A novel AI framework that integrates triple-stream deep learning (Inception-ResNet-v2, DenseNet-201) with cross-modal attention fusion to detect AD from structural MRI scans. Unlike existing methods, DFT-AD dynamically weights feature from each model via a learnable gating mechanism (α=0.78±0.03), validated by gradient-class activation maps (Grad-CAM). Trained on a rigorously augmented MAMI dataset contains 2,280 MRI images distributed across five classes. Our model achieves achieves 99.12% accuracy (95% CI: 98.2–99.8, n=227) in classifying AD across five stages. Statistical significance was confirmed via McNemar's test (p < 0.001 vs. all base models) outperforming standalone CNNs by +2.0% F1-score (0.99 vs. 0.97; 95% CI for difference: 1.2–2.8%). Crucially, DFT-AD identifies prodromal AD (MCI-to-AD conversion) with 93.7% sensitivity (95% CI: 91.2–96.2), surpassing CSF biomarkers (AUC: 0.99 vs. 0.91, *p* < 0.001, DeLong test). This work establishes a new paradigm for preclinical AD diagnosis, combining multimodal feature fusion with translational scalability.