ProgFormer predicts future brain MRI with voxel diffusion transformer

Predicting how a brain will look months or years later is crucial for tracking neurodegenerative diseases, but it's inherently difficult. Longitudinal changes are often subtle and localized, while most of the brain structure remains unchanged. Existing approaches that work in compressed latent spaces risk losing fine details, while direct voxel-space methods tend to let the dominant stable anatomy overshadow small but clinically meaningful changes.

ProgFormer tackles this with a hierarchical architecture that splits the task into two pathways: a coarse pathway models the overall brain structure and longitudinal context from 3D patch tokens, and a fine pathway uses those coarse representations as spatio-temporal grounding to refine individual patches at the voxel level. Instead of a separate autoencoder, the two pathways jointly estimate a velocity field directly in voxel space using conditional flow matching, and the future scan is generated from Gaussian noise via Euler integration. This end-to-end approach avoids latent compression leakage while remaining sensitive to disease-related local deformations. Experiments on three widely used benchmarks—ADNI, AIBL, and OASIS—show it outperforms several state-of-the-art methods in both pairwise and multi-step trajectory prediction settings.