acceptodds
Withdrawn from ICLR 2027

MRI-CARE: COMPOSE, ASSESS, AND REFINE WITH EXPERTS FOR 3D MRI RESTORATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mixed-artifact MRI restoration requires effective correction without compromising anatomy. Specialized experts offer useful priors, but incomplete correction leaves residual artifacts and mismatched inputs for subsequent experts. We propose MRI-CARE, Compose, Assess, and Refine with Experts, for multi-artifact 3D MRI restoration. An agent combines adaptive expert selection and dose control with verification and learned stopping. Its policy is trained through supervised initialization and trajectory-level reinforcement learning. We further introduce Ruler-guided Residual Refinement (R³): a 3D Ruler compares the endpoint with acceptable MRI exemplars, and a class-agnostic residual map guides a single refinement when needed. HCP experiments show improved restoration fidelity and cortical reconstruction, with R³ providing up to 1.09 dB over expert composition. External evaluation on IXI and ABIDE examines transfer to simulated and real artifacts. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MRI artifact correction, Agentic image restoration, Reinforcement learning, Image quality assessment, Residual-guided refinement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WjFxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Brain MRI scans often have multiple problems. The authors use 3 experts, one per problem, and a controller that decides which expert to use, when to stop and when to rollback. The controller is trained w/ SFT & RL. After the controller, a final R³ generic restoration further refine results. Results show that MRI-CARE is competitive on the authors' benchmark.

Critical strengths
  1. A real problem and a reasonable way to solve it. The paper writing is generally clear and easy to follow.

  2. The designed evaluation show that MRI-CARE is competitive. The paper also checks whether restored scans give better Cortical reconstruction.

Critical weaknesses
  1. The comparison with baselines is hard to trust, because the baselines and the MRI-CARE are poorly documented and not set up on equal terms. This contains:
  • Baselines are 2D methods applied to a 3D problem. PromptIR, InstructIR, TAT-MRI and Restormer are 2D-like. Motion and many other MRI artifacts are 3D problem, and each 2D slice cannot see its neighbors. MRI-CARE works on the whole 3D with pretrained 3D MRI models. A slice-by-slice baseline is therefore at a structural disadvantage. No 3D method is included, and classical methods are also missing.
  • Baseline training is also confused. The paper also does not say whether they started from photo-pretrained weights or from scratch, or what lr, batch size, loss were used. Their training differ a lot (PromptIR 6,400 steps, InstructIR 23,296 without validation-based selection, DFPIR and TAT-MRI 36,096, this is very clear in their F.2), so part of the gap may simply be under-training. The authors also admit "These checkpoints have different pretraining and optimization budgets; their comparison does not establish performance under equal training compute." in F.2.
  • The proposed method is poorly documented. Only some hyper-parameters are given for the BME-X retrain. Which LLM is used? No code is available at review time. The authors should clarify whether the baseline codes will be released, since they re-train very differently.
  • The appendix promises analyses that never appear. Eg in Appendix C, a two-reader study on IXI, but there seems no results at all. I also see no ABIDE cortical reconstruction, but Appendix C " ABIDE cortical reconstruction uses the same software and configuration for all methods". The pattern in appendix makes me unsure, and that weakens trust in the whole paper.
  1. The gains cannot be credited to the controller. They seem to come from the final R3, which is a strong model the baselines do not have. Without the final R3, the controller alone does not beat baselines on mixed problems (Groups B & C). R³ itself is a all-in-one fixer. It starts from BME-X and is fine-tuned on paired HCP scans. The paper doesn't show results of BME-X alone. The paper doesn't say how often the Ruler asks for refinement. The Ruler's agreement with human is described in the appendix but never reported (it seems they compute AUROC?).
Questions for authors

Questions:

  1. If time available, what is the results when (a) the R³ refiner applied directly (b) the R³ refiner applied to TAT-MRI's output?
  2. For baselines, please show full details in appendix. Will the baseline training code be released together with your code?
  3. Can you add a 3D baseline, or show that the 2D baselines are not disadvantaged on motion?
  4. Which appendix analyses were actually run? (two-reader study on IXI, ABIDE cortical reconstruction)
Reviewer_cUizOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MRI-CARE, an agent-based framework for correcting multiple MRI artifacts. It combines specialized restoration models, reinforcement learning, and a refinement module (R3) to improve image quality. The method is evaluated using simulated artifacts, external MRI datasets, ablation studies, and cortical reconstruction analysis.

The problem is meaningful, and the framework is technically interesting. However, the experiments do not clearly show which components are responsible for the improvements. There are also concerns about baseline fairness and generalization to real MRI data.

Critical strengths

Multiple artifacts, such as motion, noise, and intensity inhomogeneity, often occur together in MRI scans. Correcting these artifacts with a single model or a fixed sequence of models can be challenging, since correcting one artifact may affect subsequent correction steps. The proposed framework addresses a practical problem in MRI restoration.

MRI-CARE selects specialized models, adjusts correction strength, evaluates intermediate results, and decides when to stop. This adaptive strategy is an interesting alternative to conventional fixed-order restoration pipelines.

ach specialized model is trained to remove one type of artifact while preserving the others. This design helps the models focus on specific restoration tasks and supports their combination within a unified framework.

The paper includes experiments on multiple datasets, ablation studies, and downstream cortical reconstruction analysis. These experiments provide useful evidence beyond standard image quality metrics and help demonstrate the potential of the proposed framework.

Critical weaknesses

Several baseline methods appear to operate on 2D slices, while MRI-CARE uses 3D models. The methods also differ in model architectures, pretrained resources, and computational budgets. These differences make it difficult to determine whether the improvements come from the proposed framework or simply from stronger models and additional resources. A fair comparison with strong 3D baselines is better.

The paper does not convincingly demonstrate that RL performs better than simpler methods for selecting the correction order. Since there are only three specialized models, the search space is relatively small. Greedy search, exhaustive search, or simpler decision models may achieve similar performance without the complexity of RL. The authors should justify why RL is necessary.

Unclear contribution of the R3 module. R3 improves PSNR by 0.63–1.09 dB, which is substantially larger than the improvement from GRPO This raises the possibility that R3, rather than the agent-based framework or RL, is responsible for much of the reported improvement. Applying R3 to other baseline methods would help clarify its contribution.

Although the paper includes external datasets such as IXI and ABIDE, the experiments do not convincingly demonstrate robust generalization across scanners, hospitals, vendors, and acquisition protocols. MRI image characteristics can vary substantially across these settings. Generalizability is essential for MRI restoration methods, and stronger cross-dataset validation might be useful, particularly for structural brain MRI.

Most quantitative experiments use simulated motion, noise, and intensity artifacts. These simulations may not fully represent the complexity and variability of real MRI artifacts

The ABIDE experiment mainly uses no-reference image quality metrics developed for natural images. Higher scores on these metrics do not necessarily indicate that MRI artifacts have been correctly removed or that anatomical structures have been preserved. Expert reader evaluation and quantitative anatomical analysis would provide more convincing evidence.

The proposed refinement module may remove real anatomical details or introduce artificial structures, especially in unusual or pathological cases. The paper does not sufficiently evaluate these risks or verify whether the refined images preserve anatomical accuracy.

he cortical reconstruction experiment includes only nine inputs. Although this analysis is useful, the sample size might be too small to support strong conclusions about anatomical preservation and the reliability of the proposed method.

Questions for authors

Is the MRI-CARE a fair comparision with baseline?

Why is RL necessary for selecting among only three specialized models?

Whether the improvement mainly comes from R3 or the agent-based framework?

How about cross-scanner or cross-dataset validation on real MRI artifacts?

Possible to do expert assessment of anatomical preservation?

Reviewer_NyWTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MRI-CARE restores 3D brain MRI corrupted by unknown mixtures of bias field, motion and noise. In the first stage, a learned planner picks among three frozen artifact-specific experts, blends each output at three "doses", and a pairwise verifier accepts or rolls back. The policy is trained with SFT and then GRPO. In the second stage, an LMM-based "3D Ruler" scores the result against a small exemplar bank, and if it falls below a threshold, a residual detector guides one refinement pass with BME-X. Experts are trained on counterfactually paired simulated artifacts from HCP. The paper reports the best PSNR/SSIM on 126 HCP validation volumes, better FreeSurfer surfaces on a 9-volume subset, consistent ablation gains, and external results on simulated IXI, real ABIDE (no-reference IQA) and a zero-shot ghosting test.

Critical strengths
  1. Sequential experts see inputs that do not match their training distribution, and the paper frames this clearly. The counterfactual pairing (remove only the target artifact, replay the others with identical realisations) is a sensible, reusable idea. This is the main contribution I would expect others to build on.
  2. It evaluates with downstream anatomy, not only image metrics. Reporting FreeSurfer surface holes and cortical thickness error alongside PSNR/SSIM is more meaningful, and the gains there are large (SHN 26.0 vs. 68.7). Together with a systematic, consistent component ablation, this supports the design, although the cortical analysis uses only 9 volumes.
Critical weaknesses
  1. The paper presents MRI-CARE as a safe, clinically useful restorer for real scans, but almost everything is measured on simulated artifacts. The thresholds and weights were tuned on that same data, there are no confidence intervals or repeated seeds, and several ablation gains are only 0.3–0.6 dB. The evaluations that would actually support the clinical claims are described in the appendix but never reported: the two-reader blinded study, the Ruler's accuracy against human labels, and the comparison of gated vs. always-on refinement. The real-data evidence is limited to ABIDE scored with natural-image quality metrics (MUSIQ 38.82 vs. 38.47), and the cortical analysis uses 9 volumes. With a generative refiner in the loop and almost no pathology in any dataset, the risk of altering anatomy is also never tested. As it stands, the paper shows a system that does well on simulated, tuned-on data, not that it is reliable on real scans.

  2. It is unclear that the agentic design is what helps. MRI-CARE sits on two strong pretrained 3D MRI models (a frozen foundation backbone and BME-X), while the main baselines are 2D slice-wise models with different pretraining. The large margins could therefore come from the prior rather than from planning, verification and gated refinement. The comparisons that would separate the two are missing: BME-X alone, a single all-in-one head on the same backbone and data, and the experts applied once in a fixed order without the verifier. This matters because the novelty claim rests on the agentic pipeline, and learned tool selection with assessment and rollback is already well studied (RL-Restore, RestoreAgent, AgenticIR, MAIR). Without these controls, the paper doesn't show which of its many components are needed or what it adds beyond prior agentic restoration.

Questions for authors
  1. Can you report the main results on a held-out, subject-disjoint test split that was not used for any tuning, with subject-level confidence intervals?
  2. Where are the reader-study results and the Ruler's accuracy against human labels? How often does the Ruler reject, and does gating beat simply always or never refining?
  3. How do BME-X alone, a same-backbone all-in-one head, and fixed-order experts without the verifier perform?
  4. What evidence do you have that the pipeline preserves anatomy on real artifacts and on scans with lesions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.