Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction
Researchers have developed a novel method for MRI reconstruction that moves the process into a discrete multi-scale latent space, framing it as autoregressive next-acceleration-scale prediction. This approach leverages discrete priors, similar to those used in visual autoregressive modeling, to restrict solutions to compact sequences of codebook tokens, enabling sharper reconstructions even with extremely sparse measurements. The method also incorporates on-policy privileged information distillation, a technique inspired by large language model post-training, to further enhance reconstruction accuracy. Experiments on the fastMRI benchmark demonstrate improved performance across various sampling patterns under extreme undersampling. AI
IMPACT Introduces a novel approach to MRI reconstruction by adapting techniques from visual autoregressive modeling and large language models.