Multimodal MRI and Tabular Data Synthesis via Diffusion Model

A team of researchers has introduced a multimodal latent diffusion model that integrates volumetric magnetic resonance imaging (MRI) alongside tabular clinical data in a unified latent space through cross-attention mechanisms. This model employs a variational autoencoder to merge the MRI and tabular data prior to diffusion-based synthesis, featuring distinct decoders for each type of data. The model was tested using information from the German National Cohort (NAKO Gesundheitsstudie), which includes over 10,000 participants. The resulting MRI volumes demonstrated anatomical accuracy and body composition aligned with the synthesized tabular data, encompassing attributes such as age, sex, body measurements, and ethnicity. Fréchet distance and precision-recall metrics were utilized for quantitative assessment.