Diagnostically faithful fMRI-dMRI synthesis using disease-informed diffusion priors

Functional magnetic resonance imaging (fMRI) and diffusion magnetic resonance imaging (dMRI) are vital for studying Alzheimer’s disease but often suffer from missing modalities. Existing generative models fail to capture the fundamental signal mismatch between fMRI and dMRI or incorporate disease-relevant anatomy, yielding diagnostically weak outputs. We propose P attern-aware D iffusion S ynthesis ( PDS ), a pattern-aware diffusion framework that integrates structure-function coupling priors from anatomical atlases into a dual-modal 3D diffusion model, enabling semantically coherent bidirectional synthesis. A tissue-aware projection network and microstructure refinement module further enhance anatomical fidelity. Experiments on OASIS-3, ADNI, and an in-house cohort show PDS achieves state-of-the-art PSNR/SSIM: 29.83 dB/0.908 (fMRI) and 30.00 dB/0.776 (dMRI). Crucially, synthetic scans support robust diagnosis: using hybrid real-synthetic data, PDS yields 67.92% accuracy in normal control (NC)/ mild cognitive impairment (MCI)/Alzheimer’s disease (AD) classification; when trained only on synthetic data and tested on real scans, it achieves 77.35% accuracy-the highest among all methods-demonstrating superior preservation of clinically meaningful information.