Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction

TL;DR

This paper introduces a novel discrete autoregressive approach for MRI reconstruction that improves image quality from highly undersampled data by leveraging multi-scale latent space modeling and privileged information distillation.

Contribution

It proposes a discrete autoregressive model for MRI reconstruction and a privileged information distillation technique, enhancing reconstruction quality under extreme undersampling conditions.

Findings

Achieves sharper MRI reconstructions from sparse measurements.

Outperforms existing methods on the fastMRI benchmark.

Demonstrates robustness across various sampling patterns.

Abstract

MRI reconstruction is an inherently ill-posed inverse problem, since incomplete measurements admit many plausible solutions. This ambiguity becomes more severe under high acceleration, where pixel-domain continuous predictors tend to average over feasible reconstructions and suppress high-frequency anatomy. We address this limitation by moving reconstruction to discrete multi-scale latent space and posing it as autoregressive next-acceleration-scale prediction. Leveraging discrete priors proven effective in visual autoregressive modeling, our method restricts the solution to compact sequences of codebook tokens, enabling sharp reconstructions even from extremely sparse measurements. This discrete autoregressive formulation also aligns naturally with modern large language model post-training techniques. Building on this observation, we introduce on-policy privileged information distillation for visual autoregressive modeling, where a teacher is provided training only privileged context that is unavailable at inference, in our case fully sampled acquisitions, and supervises a student trained on its own rollouts, leading to consistent reconstruction gains. Through extensive experiments on the fastMRI benchmark, we show that our approach delivers improved reconstruction performance across diverse sampling patterns under extreme undersampling. Project website is \href{https://yilmazkorkmaz1.github.io/discrete-mri-reconstruction-opd/}{here}.