Integrating anatomical priors and clinical semantics for MRI-based diagnosis and care support in Alzheimer's disease.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder, and magnetic resonance imaging (MRI) has become an important tool for its auxiliary diagnosis because it can reveal key structural abnormalities, including hippocampal and parahippocampal atrophy, ventricular enlargement, temporal cortical degeneration, and gray matter loss. However, reliable stage-aware classification remains challenging because current methods are still limited in handling weak anatomical boundaries, low-contrast lesions, subtle inter-stage differences, and the semantic gap between neuroimaging features and clinical descriptions. To address these issues, we propose AFCG-Net, a diagnosis-guided and frequency-aware cross-modal network for joint MRI-text diagnosis. The framework consists of three stages: anatomy-guided visual encoding, cross-modal semantic alignment, and gated fusion with anatomical priors. Specifically, ASFG enhances visual representations through multi-scale modeling and low frequency-guided refinement, CASF strengthens the consistency between clinical semantics and imaging features, and AGDF performs anatomy-guided deep fusion to improve both discriminability and interpretability. Experiments on a combined cohort of 4,197 original MRI-text samples, including 2,106 self-collected samples and 2,091 ADNI samples, show that AFCG-Net achieves Precision, Recall, and F-score values of 96.3%, 96.5%, and 95.8%, respectively. The proposed method achieves the best results in the Mild dementia and Moderate dementia categories, while also showing stronger performance in the more challenging Non-dementia and Very Mild dementia categories. These results suggest that AFCG-Net provides an effective and interpretable multimodal solution for AD-assisted diagnosis.