Title: Wavelet-Fusion Diffusion Model for Multimodal Brain MRI Synthesis with Modality and Metadata Conditioning

Abstract:

Multimodal MRI offers complementary insights crucial for neuroimaging analysis, as distinct modalities highlight specific anatomical, tissue, and pathological characteristics that facilitate the development and assessment of downstream artificial intelligence applications. While large-scale structural MRI repositories are becoming more abundant, their modality coverage remains inconsistent across public and aggregated neuroimaging datasets. This disparity is exacerbated by heterogeneity in sites, scanners, and acquisition protocols, alongside demographic and clinical variables that are frequently sparse, recorded inconsistently, or entirely missing in various studies. Synthetic MRI generation presents a solution to this imbalance by creating target-modality volumes, thereby enabling dataset augmentation and the construction of controlled synthetic cohorts.

However, current MRI synthesis methods are often limited by training on narrow modality sets or homogeneous groups, restricting their utility for large pooled neuroimaging resources where modality availability, acquisition standards, and metadata coverage differ significantly. Although diffusion models are increasingly favored for MRI synthesis due to their high sample fidelity and diversity, direct sampling in 3D voxel space is computationally intensive and slow during inference. Latent diffusion offers a more practical alternative by generating MRIs within a learned 3D latent space, though the quality of the output relies heavily on the autoencoder’s reconstruction accuracy and the nature of the resulting latent distribution.

To address these challenges, our method integrates a Wavelet-Fusion variational autoencoder (WF-VAE) as a latent compressor with a conditional 3D U-Net diffusion model. This model is trained in the learned latent space, utilizing explicit conditioning on both modality and metadata. The proposed Wavelet-Fusion Diffusion Model (WFDM) demonstrated superior distributional alignment compared to other evaluated synthetic MRI generators.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC