Integrating Structural and Metabolic Neuroimaging Biomarkers for Alzheimer's Disease Diagnosis and Cognitive Score...

Structural atrophy and metabolic dysfunction provide complementary biomarkers for Alzheimer's disease (AD), and their joint modeling may support diagnostic assessment and cognitive score estimation. However, many multimodal methods rely on global fusion and insufficiently enhance cross-modal consistency before interaction, limiting the discriminative quality and clinical relevance of learned representations. We propose CGMF-Net, a cross-modal gated learning framework for joint AD classification and clinical score estimation using paired structural MRI (sMRI) and fluorodeoxyglucose PET (FDG-PET) data. CGMF-Net extracts multi-scale representations from both modalities, introduces a Cross-Modal Similarity Gate to strengthen consistent structural-metabolic responses before fusion, and employs bi-directional cross-attention to capture complementary interactions. The shared representation is optimized with classification supervision, MMSE-based auxiliary regression, and HSIC regularization to improve discriminability and reduce redundant coupling between directional representations. Experiments on ADNI demonstrate that CGMF-Net achieves the best overall classification performance among the compared methods, with 94.22% ACC and 97.74% AUC for AD vs. CN, and 86.67% ACC and 94.84% AUC for AD vs. MCI, while also showing favorable ADNI-2 to ADNI-1 generalization and competitive estimation of ADAS13, CDRSB, and MMSE. These results suggest that cross-modal gated learning provides clinically relevant multimodal representations for AD diagnosis and cognitive score estimation.