Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Cross-modality MRI synthesis addresses a pervasive clinical bottleneck—brain MRI acquisitions are frequently incomplete due to contraindications, motion artifacts, or time constraints, yet neuro-oncology workflows require the full multi-modal stack. Current models prioritize textural realism over protocol compliance, generating images that pass pixel-level metrics while violating American College of Radiology (ACR) imaging standards (e.g., CSF-suppression failures). We introduce MedRAG-SCA, an agentic framework enforcing clinical rigor via a \textit{Retrieve–Verify–Refine} loop: a Protocol-Aware Retrieval-Augmented Generation (RAG) Consultant converts qualitative ACR guidelines into verifiable pixel-space criteria; a Tri-Critic Committee (Anatomist, Radiologist, Diagnostician) localizes violations for targeted inpainting; and a Geometric Fail-safe preserves pathological geometry in complex edema cases. On the Brain Tumor Segmentation (BraTS) 2024 benchmark, the Clinical Acceptability Rate (CAR)—rated by blinded board-certified neuroradiologists—reaches 92\%, 11 percentage points above the strongest diffusion baseline ($p{<}0.001$, McNemar’s test); the Protocol Adherence Score (PAS) improves by 19\% ($0.94$ vs.\ $0.79$, $p{<}0.001$), with downstream segmentation Dice maintained at $\geq$0.80. These results demonstrate that explicit protocol grounding bridges the gap between visual plausibility and clinical validity in MRI synthesis. Code is available at \url{https://github.com/exsinger-hub/MedRAD-SCA}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3774_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YuaFen_MedRAGSCA_MICCAI2026,
        author = { Yuan, Feng AND Gao, Yifan AND Li, Haoyue AND Gao, Xin},
        title = { { MedRAG-SCA: A Retrieval-Augmented Self-correcting Agent for Clinically Compliant Cross-Modality MRI Synthesis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

This paper addresses an important limitation of current cross-modality MRI synthesis methods: they often optimize for pixel-level realism without ensuring compliance with clinical imaging protocols. The main contribution is MedRAG-SCA, an agentic retrieve-verify-refine framework for clinically compliant brain MRI synthesis. The method introduces three tightly coupled components: (1) a Protocol-Aware Guideline Consultant that retrieves expert-reviewed ACR/NIH/MedlinePlus documents and converts qualitative imaging recommendations into quantitative pixel-space constraints; (2) a Source-Anchored Tri-Critic Committee (Anatomist, Radiologist, Diagnostician) that verifies geometry, modality fidelity, and pathology preservation without requiring target-domain ground truth during verification; and (3) a Targeted Refinement Controller that performs local inpainting of protocol-violating regions while protecting tumor boundaries and edema. On BraTS 2024, the framework improves blinded radiologist-rated Clinical Acceptability Rate (CAR) from 81% to 92% over the strongest diffusion baseline, while also improving protocol adherence and preserving downstream segmentation performance.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

1.Strong and clinically meaningful problem formulation. A major strength of the paper is that it explicitly distinguishes visual plausibility from clinical validity. This is an important point for medical image synthesis, and the authors motivate it convincingly with concrete examples such as CSF suppression failures that may not be captured by standard similarity metrics.

2.Novel protocol-grounded framing of MRI synthesis. This paper introduces protocol-aware synthesis by grounding generation and refinement in retrieved imaging guidelines. The idea of translating qualitative clinical protocol language into quantitative pixel-space constraints is interesting and clearly differentiated from standard GAN/diffusion synthesis papers.

3.Well-designed multi-component framework. The Retrieve–Verify–Refine loop is well structured and easy to follow. The Protocol-Aware RAG Consultant, the Source-Anchored Tri-Critic Committee, and the Targeted Refinement Controller have distinct roles that fit together coherently. The Geometry–Intensity Decoupling principle is also well articulated and clinically intuitive.

4.Good attention to pathology preservation. The pathology-preserving shield and the Geometric Fail-safe are thoughtful additions. They address an important failure mode of synthesis/refinement systems, namely improving apparent modality fidelity at the cost of altering lesion boundaries or edema structure. This adds credibility to the clinical framing of the work.

5.Strong clinical evaluation beyond pixel-level metrics. A particularly notable strength is the inclusion of blinded neuroradiologist evaluation with CAR as the primary clinical endpoint. This is much stronger than relying solely on PSNR/SSIM/FID and directly supports the paper’s central claim that protocol-aware synthesis improves clinical acceptability.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

1.Scope is still limited to neuro-oncology MRI synthesis. The results are strong for the stated task, but the broader generality of the method is not yet established. The framework may not transfer directly to other organs or other protocol families without considerable adaptation of the retrieval knowledge base and the constraint extraction machinery.

2.Comparison is strongest on clinical compliance, but pixel-level tradeoffs remain. The paper argues, plausibly, that slight reductions in PSNR/SSIM may reflect protocol-correct synthesis rather than model weakness. Still, this makes the evaluation somewhat harder to interpret, and the reader must largely accept the paper’s clinical framing to fully endorse the tradeoff. More direct downstream task validation (such as segmentation and classification) would make this argument even stronger.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

N/A
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(5) Accept — should be accepted, independent of rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This paper stands out by addressing a clinically important gap in cross-modality MRI synthesis: the fact that images can be visually plausible yet clinically non-compliant. I find the problem formulation compelling, the overall method design coherent, and the integration of protocol retrieval, multi-critic verification, and targeted refinement genuinely interesting. The inclusion of blinded neuroradiologist evaluation is a major strength and provides meaningful evidence that the method improves clinical acceptability rather than merely optimizing image similarity metrics. The ablation study is also well aligned with the proposed method and supports the contribution of each major module.

While PAS is useful mechanistically, it is partly self-referential because the framework directly optimizes protocol adherence. That said, the paper acknowledges this limitation and relies on CAR as the primary endpoint, which is appropriate. On balance, I view this as a clinically meaningful and methodologically interesting paper that is slightly above the acceptance threshold.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #2

Please describe the contribution of the paper

This paper introduces MedRAG-SCA, the first framework to enforce clinical protocol compliance (ACR standards) in cross-modality MRI synthesis by translating qualitative radiology guidelines into quantitative pixel-space constraints via retrieval-augmented generation. The work demonstrates that explicit protocol grounding—through a Protocol-Aware RAG Consultant and Tri-Critic Committee (Anatomist, Radiologist, Diagnostician)—can bridge the gap between visual plausibility and clinical validity. The authors show that their method achieves 92% Clinical Acceptability Rate (CAR) versus 81% for state-of-the-art diffusion baselines, while maintaining downstream segmentation performance (Dice ≥0.80). The core contribution is a geometry-intensity decoupling strategy that enables source-anchored protocol verification without paired ground truth.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

Novel Clinical Grounding: First work to integrate retrieval-augmented generation for clinical protocol adherence in image synthesis, converting text-based ACR guidelines into verifiable pixel-space intensity ratios (e.g., CSF/WM ∈[3.5,4.2]). This addresses a real clinical need often ignored in synthesis literature. Rigorous Clinical Validation: Unlike most synthesis papers relying solely on pixel metrics, this work employs blinded evaluation by three board-certified neuroradiologists (ICC=0.81), establishing CAR as the primary endpoint. This sets a high bar for clinical utility assessment. Technical Innovation: The Tri-Critic Committee architecture elegantly separates concerns—geometry (Anatomist), modality fidelity (Radiologist), and pathology preservation (Diagnostician)—enabling unsupervised, ground-truth-free quality assurance. The Geometric Fail-safe mechanism for complex edema cases shows sophisticated handling of clinical edge cases. Practical Impact: Directly addresses the clinical bottleneck of incomplete MRI acquisitions due to contraindications/motion, with demonstrated protocol transfer to healthy subjects (IXI dataset) without retraining. Comprehensive Evaluation: Multi-dataset validation (BraTS 2024, IXI), ablation studies confirming each component’s contribution (RAG provides largest gain: +21% CAR), and pathological fidelity verification (ET Dice 0.763 vs 0.724 baseline).
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

Scope Limitation: Evaluation restricted to neuro-oncology (glioma). Generalization to other anatomies (spine, abdominal) or pathologies (stroke, MS) is unverified, limiting immediate clinical translatability beyond brain tumors. Confirmation Bias in PAS: The Protocol Adherence Score (PAS) is directly optimized by the framework (via gradient guidance), making it susceptible to confirmation bias. While CAR (primary endpoint) is independent, the tight correlation between PAS and CAR (ρ=0.72) suggests potential circularity in the optimization target. Performance Degradation in Risk Zones: The Geometric Fail-safe triggers in 12.7% of cases (extensive edema), causing PAS to drop to 0.68 and requiring reduced-confidence flagging. This reveals fragility when source-target anatomy diverges significantly. System Complexity: The multi-component pipeline (RAG retrieval, MedSAM segmentation, three critics, iterative refinement) introduces multiple failure points and computational overhead that may hinder clinical deployment, despite the <50ms retrieval latency claim. Retrieval Dependency: Performance relies heavily on the curated 1,170-document knowledge base. Rare protocols or emerging guidelines may not be covered, and the parse accuracy (>90%) leaves edge cases dependent on conservative wide-range priors. Marginal PSNR Trade-off: While explained as reference non-compliance, the PSNR reduction (27.85 vs 28.30 dB) compared to SynDiff may concern practitioners prioritizing pixel fidelity over protocol adherence.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

Consider suggesting the authors clarify the generalizability to non-glioma pathologies and discuss computational requirements for real-time clinical use. The distinction between “visual plausibility” and “clinical validity” is a valuable conceptual contribution that could be highlighted more prominently in the introduction
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper presents a compelling solution to a clinically significant problem—protocol-compliant MRI synthesis—that has been overlooked by the field’s focus on pixel metrics. The retrieval-augmented approach to enforcing ACR standards is genuinely novel, and the blinded radiologist study provides credible evidence of clinical utility (92% CAR). The geometry-intensity decoupling principle enables practical implementation without paired ground truth. While limited to neuro-oncology and exhibiting complexity that may challenge deployment, the methodological contribution and clinical validation quality warrant acceptance. The work advances the field beyond “toy” synthesis tasks toward clinically deployable solutions
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #3

Please describe the contribution of the paper

MedRAG-SCA is a brain MRI cross-modality synthesis framework that enforces clinical protocol compliance via a retrieve-verify-refine loop. (1) A Protocol-Aware RAG Consultant retrieves 1,170 documents, parses qualitative guidelines with spaCy, and maps them to quantitative intensity-ratio constraints. (2) A Tri-Critic Committee verifies geometry, modality fidelity, and pathology preservation using source-derived segmentation masks and feature-based shields. (3) A Refinement Controller iteratively inpaints violating regions via RePaint, with a Geometric Fail-safe that suspends intensity constraints when edema causes cross-modality mask divergence (Dice < 0.75 for 2 iterations). Evaluated on BraTS 2024 and IXI, the system achieves 92% Clinical Acceptability Rate (CAR) versus 81% for SynDiff, and PAS 0.94 versus 0.79.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

1.Clinically grounded problem. The decoupling of pixel-level metrics from clinical validity is correctly identified. The SynDiff example—SSIM > 0.90 yet ACR CSF-suppression failure in 29/50 radiologist-reviewed cases—is a compelling motivation that aligns with recent critiques of pixel-wise metrics in medical imaging. 2.Rigorous radiologist study. Three independent board-certified neuroradiologists rated 150 images with ICC = 0.81 [0.76, 0.85]. Pre-specifying CAR as the primary endpoint and powering the study appropriately is commendable. 3.Pathology-preserving refinement. The Diagnostician critic’s expanding feature-based shield (RadImageNet ResNet50 layer-4) prevents progressive overcorrection of lesion boundaries. This is a thoughtful solution to a known failure mode in iterative medical synthesis.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

1.The RAG claim is inflated. The system does not use generative retrieval-augmented reasoning. It retrieves documents with BioBERT + FAISS, but the actual “translation” from text to pixel constraints is performed by spaCy dependency parsing mapped through a 47-entry lookup table built from only 50 radiologist-annotated documents. This is retrieval-augmented templating, not RAG. The title and abstract mislead readers into expecting end-to-end LLM reasoning over retrieved guidelines. 2.The ablation conflates retrieval with any textual guidance. Removing “RAG” (replaced with static prompts) drops CAR from 92% to 71%, but this compares the full system against no protocol text at all. 3.Confirmation bias in PAS. PAS is directly optimized via the hinge protocol loss (Eq. 2). The authors admit this bias, yet they still foreground PAS improvements (0.94 vs. 0.79, including per-constraint breakdowns in Table 2) rather than treating PAS strictly as a training signal. CAR partially mitigates this, but PAS should be de-emphasized in the results. 4.Validation scope is too narrow. Only BraTS 2024 (challenge data) and IXI (healthy subjects). No real clinical PACS data, no multi-vendor scanners, no non-glioma pathology, no cross-ethnic cohorts. For a paper claiming clinical compliance, this gap is too large. Recent contrast-enhancement work has demonstrated validation across multiple external datasets, vendors, and clinical sites; this paper’s footprint is substantially narrower. 5.Parser generalization is questionable. “>90% parse accuracy on held-out validation” with only 50 documents implies a tiny test set (likely ~10 documents under a standard 80/20 split, or ~5 under a 90/10 split).
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

The clinical study is rigorous and the Tri-Critic architecture is clever, but the central RAG claim is overstated.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The clinical study is rigorous and the Tri-Critic architecture is clever, but the central RAG claim is overstated. The mechanism is a small lookup table, not generative retrieval-augmented reasoning. The ablation fails to isolate the value of retrieval, PAS suffers from confirmation bias, and the validation scope is limited to a single challenge dataset plus healthy controls. The Geometric Fail-safe is also hyperparameter-sensitive and poorly validated.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Author Feedback

We sincerely thank the reviewers and meta-reviewer for their constructive comments and positive evaluation of our work. We are grateful that the reviewers recognized the clinical motivation of distinguishing visual plausibility from clinical validity, as well as the value of protocol-aware synthesis, the tri-critic verification design, pathology-preserving refinement, and blinded radiologist evaluation.

We will carefully revise the manuscript for clarity in the camera-ready version, including refining the description of the retrieval-augmented protocol-grounding mechanism, clarifying the role of PAS as a mechanistic measure rather than an independent clinical endpoint, and improving the discussion of validation scope, Risk Zone behavior, and deployment considerations. We also appreciate the suggestions regarding future extensions to broader anatomies, pathologies, vendors, and clinical datasets, and will discuss these limitations and future directions more explicitly in the final version where appropriate.

Meta-Review

Meta-review #1

Your recommendation

Provisional Accept
Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

The paper addresses the important problem of cross-modality synthesis aligned with clinical imaging protocols focusing on the context of euro-oncology. All reviewers highlighted the clinical relevance of the contribution and the novelty of the proposed approach. It appears however that there may be some degree of circularity in the proposed evaluation that should be clarified.

back to top

MedRAG-SCA: A Retrieval-Augmented Self-correcting Agent for Clinically Compliant Cross-Modality MRI Synthesis

Author(s):