Wavelet-Fusion Diffusion Model for Multimodal Brain MRI Synthesis With Modality and Metadata Conditioning

Multimodal MRI provides complementary information for neuroimaging analysis, with different modalities capturing distinct anatomical, tissue, and pathological features relevant to downstream artificial intelligence (AI) applications. Although large-scale structural MRI resources are increasingly available, modality coverage remains uneven across public and pooled neuroimaging datasets. Large pooled datasets span heterogeneous sites, scanners, and acquisition protocols and often contain incomplete or inconsistent demographic and clinical metadata. Synthetic MRI can supplement missing or underrepresented modalities and support metadata-conditioned synthetic cohorts. However, many existing approaches use limited modality combinations or relatively homogeneous datasets, limiting their applicability to heterogeneous pooled neuroimaging data. Diffusion models can generate high-fidelity, diverse MRI, but direct 3D voxel-space sampling is computationally expensive. Latent diffusion reduces this cost by generating in a compressed space, although information loss during compression can limit synthesis fidelity. We develop the Wavelet-Fusion Diffusion Model (WFDM), a 3D latent diffusion framework for multimodal brain MRI synthesis with target-modality and metadata conditioning. WFDM combines a Wavelet-Fusion variational autoencoder (WF-VAE) compressor with a conditional 3D U-Net diffusion model trained in the latent space. Training and evaluation use 45,880 preprocessed T1w, T2w, T1Gd, and FLAIR scans from 157 BrainScape datasets. Compared with the BrainScape-trained MAISI (Medical AI for Synthetic Imaging) baseline, WFDM reduced pooled MedicalNet feature-space Fréchet distance (FID) from 0.00550 to 0.00404 and pooled Maximum Mean Discrepancy (MMD) from 0.1918 to 0.1471 while maintaining comparable sample diversity. WFDM achieved lower FID and MMD for T1w, T2w, and FLAIR, whereas MAISI achieved lower values for T1Gd. Rectified Flow sampling with 50 steps was 12.4–13.4 times faster than 1000-step DDPM sampling while achieving the lowest pooled FID and MMD in the sampler ablation. In T1w MRI, WFDM also more closely preserved age- and sex-associated morphometric patterns, including age-group morphometric separation, regional age–volume associations, and male–female differences. We release BrainMint for brain MRI synthesis workflows and medmetric for standardized evaluation of 3D medical image synthesis models.