Johns Hopkins' autoregressive MRI method sharpens images from sparse data

MRI reconstruction is an ill-posed inverse problem that gets worse under high acceleration—conventional continuous predictors blur high-frequency anatomy. Korkmaz and Patel (Johns Hopkins) address this by reformulating reconstruction as autoregressive next-acceleration-scale prediction in a discrete multi-scale latent space. Instead of averaging over plausible solutions, the model generates compact sequences of codebook tokens, enabling sharp reconstructions from extremely sparse k-space measurements.

To further boost performance, they adapt a technique from large language model post-training: on-policy privileged information distillation. A teacher model is trained with fully sampled acquisitions (privileged context unavailable at inference), then supervises a student model that learns from its own autoregressive rollouts. Experiments on the fastMRI benchmark show consistent gains under extreme undersampling, suggesting this discrete autoregressive framework could significantly reduce scan times without sacrificing image quality.