Beyond Diffusion: Consistency Models for One‑Step, High‑Fidelity MRI Reconstruction

Mary-Brenda Akoda, Chen Qin

Keywords: MRI reconstruction, diffusion models, consistency models, inverse problems, cardiac MRI

Abstract: Magnetic resonance imaging (MRI) provides excellent soft-tissue contrast, but suffers from long acquisition times, limiting throughput and increasing patient discomfort. Diffusion-based generative models have recently achieved state-of-the-art reconstruction quality for accelerated MRI, but typically require hundreds to thousands of neural function evaluations (NFEs), which severely limits their practicality in time-sensitive clinical settings. We introduce C-MORE (Consistency-Model-based One-step REconstruction for MRI), to our knowledge, the first one-step consistency model framework for accelerated MRI reconstruction. C-MORE investigates an unconditional one-step prior and solves the inverse problem in one NFE by leveraging measurement-guided encoding and tunable physics-based refinement, thus eliminating multi-NFE diffusion sampling, while retaining a controllable quality-speed trade-off. On the MICCAI CMR$\times$Recon dataset spanning multiple cardiac contrasts and both single- and multi-coil acquisitions, C-MORE outperforms state-of-the-art diffusion‑based samplers and strong non-diffusion unrolled methods across accelerations in just 1 NFE, while reconstructing images in $0.18-0.52$ s ($\approx$$22-193$$\times$ faster than diffusion-based methods requiring hundreds of NFEs). Remarkably, without any retraining or finetuning, C-MORE also demonstrates cross-anatomy generalisation to the unseen fastMRI knee dataset from NYU Langone Health and Facebook AI Research, again surpassing state‑of‑the‑art methods across accelerations. These results establish C-MORE as a practical blueprint for real-time, high-fidelity MRI reconstruction across diverse contrasts, acquisition settings, anatomies, and accelerations.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 162