Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction

2026-05-19

AI Summary

This work addresses the challenges of solution space ambiguity and high-frequency detail loss in highly accelerated magnetic resonance imaging (MRI) caused by extreme undersampling. To this end, the authors propose an autoregressive reconstruction method operating in a discrete multiscale latent space. The approach formulates image reconstruction as an autoregressive prediction task across successive acceleration scales, leveraging discrete codebook sequences to constrain the solution space. Notably, it introduces, for the first time, a policy-based privileged information distillation mechanism to enhance reconstruction fidelity. Evaluated on the fastMRI benchmark, the method consistently outperforms existing approaches under various extreme undersampling patterns, effectively preserving high-frequency anatomical details and achieving stable, high-quality reconstructions.

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 \hyperlink{https://github.com/yilmazkorkmaz1/discrete-mri-reconstruction-opd}{here}.

Problem

• MRI reconstruction
• ill-posed inverse problem
• high acceleration
• extreme undersampling
• high-frequency anatomy

Innovation

• discrete autoregressive modeling
• multi-scale latent space
• next-acceleration-scale prediction
• privileged information distillation
• MRI reconstruction