Generative Brain Priors-Driven Self-Regulated and Clinically-Guided Multimodal Fusion for Early Alzheimer's Disease Diagnosis.

Alzheimer's disease (AD) and mild cognitive impairment (MCI) require accurate early diagnosis to support timely clinical intervention and disease management. Although multimodal neuroimaging with structural MRI, FDG-PET, and AV45-PET has shown promise, existing deep learning methods still face three major challenges: insufficient modeling of pathology-related features across local brain regions and whole-brain connectivity, heterogeneous feature distributions among modalities, and limited use of clinical information for dynamic representation learning. To address these challenges, we propose GBP-SCMF, a multimodal diagnostic framework for AD spectrum classification. The framework first extracts multi-scale features from each imaging modality using parallel modality-specific branches. It then integrates three complementary mechanisms: generative brain-prior enhancement to inject disease-related pathological knowledge into imaging representations, self-regulated multimodal alignment to reduce modality bias and feature redundancy, and clinical-imaging collaborative fusion to embed cognitive measurements as semantic tokens for gated interaction with imaging features. These components are jointly optimized to improve both cross-modal consistency and disease-discriminative representation learning. We evaluated GBP-SCMF on the ADNI dataset for three-class classification among cognitively normal controls, MCI, and AD. The proposed method achieved an overall classification accuracy of 0.9021 and a macro-average area under the receiver operating characteristic curve (AUC) of 0.9514. Ablation studies and attention visualizations further demonstrated the complementary contribution of each component and the biological plausibility of the learned representations. These results suggest that GBP-SCMF provides an effective and interpretable multimodal strategy for computer-aided AD diagnosis.