Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction

Researchers introduced a discrete autoregressive MRI reconstruction method that improves image quality under extreme undersampling.

The model moves reconstruction into a multi-scale discrete latent space and predicts next acceleration scales with codebook tokens.

A student is trained with on-policy privileged-information distillation from a teacher that can access fully sampled data.

On fastMRI, the approach reports stronger reconstruction results across sampling patterns and better preservation of fine anatomy.