Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Image Restoration (IR) aims to enhance degraded images to provide high-quality diagnostic references in Magnetic Resonance Imaging (MRI). Although recent All-in-One IR (AiOIR) methods seek to handle multiple artifacts within a unified network, they still struggle with mixed artifacts, where multiple unknown artifacts occur simultaneously in a single MRI scan. To tackle this challenge, we propose ResMAP, a cascading framework for Restoring MRIs of Mixed Artifacts by Prompt Retrieval. It is trained exclusively on individual artifact types but can effectively handle all their mixed forms in inference, offering a feasible solution instead of requiring exhaustive training on mixed artifacts. Specifically, our ResMAP utilizes a coarse-to-fine correction process for mixed artifacts by cascading retrieval of prompts based on the artifact types. In this process, the retrieval guidance is provided through the perception and classification of fine-grained image features, while the prompts are prepared via LLM-based generation and fine-tuning. Validations on three types of artifacts and their mixed forms demonstrate the superiority of ResMAP over current IR methods. Besides, zero-shot experiments on MRIs from multiple field strengths further confirm the promising generalizability of the proposed framework. Our code is available at https://github.com/Tanishabc/ResMAP.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1894_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Tanishabc/ResMAP

Link to the Dataset(s)

N/A

BibTex

@InProceedings{TanYux_ResMAP_MICCAI2025,
        author = { Tang, Yuxian AND Li, Feng AND Shi, Feng AND Wang, Qian},
        title = { { ResMAP: Restoring MRIs of Mixed Artifacts by Prompt Cascading Retrieval } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {506 -- 515}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes ResMAP, a cascading prompt-based framework for MRI image restoration under mixed artifact conditions. The core contribution lies in using LLM-generated textual prompts guided by an image-based degradation classifier to sequentially remove individual artifacts in a coarse-to-fine manner.

  • 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.

    The strength of the paper lies in its novel cascading framework, ResMAP. It restores mixed MRI artifacts effectively using prompt retrieval.

  • 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.

    Major issues:

    1. Authors didn’t specify the features (Fi) from the encoder (EI) used as input to the two-layer MLP classifier (CI).
    2. The capability of this relatively simple classifier to accurately distinguish between multiple, subtle, MRI artifact types is not sufficiently justified or analyzed. Misclassification at any step could derail the cascade.
    3. The automatic cascading strategy (Section 2.3) is underexplained. There is no principled justification or algorithmic description for determining the correction order.
    4. It is unclear why classifier-based (CT) optimization was chosen over image-text alignment methods for enhancing restoration guidance, and the training details are insufficient.
    5. The description mentions training and validation slices derived from the 120 HCP subjects. There is no explicit mention of a separate, unseen test set from the same HCP source.
    6. The analysis of adaptive correction order effectiveness (Fig 5) is limited to one artifact pair (bias+motion) with varied bias intensity. The paper doesn’t properly show the proposed methods generalizability to other artifact combinations or variations. Moreover, the “reverse order” comparison lacks clear definition.
    7. The current model assumes one artifact per iteration, which might not generalize well to heavily degraded inputs.
    8. Stronger evidence is required to support the central claim of handling mixed artifacts via single-artifact training. The cascade effectiveness analysis (Sec 3.4) is currently limited to a single artifact pair. Minor issues:
    9. First sentence in the second page seems incomplete.
    10. Conclusion, “clean sdandard” should be “clean standard”.
    11. Fig. 3 Caption, correction diviation” should be “correction deviation”.
    12. The label for the accuracy row (“IterRes (ACC)”) could be more explicit, “ResMAP Classifier Accuracy (ACC)”.
    13. The heading “2.3 Inferance” should be corrected to “2.3 Inference”.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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.

    (3) Weak Reject — could be rejected, 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 novel idea with strong potential. It uses LLM-generated prompts and a cascading framework to address mixed MRI artifacts. However, several key components are underexplained. The classifier design and correction order strategy lack clarity. The analysis of generalizability is limited. These issues reduce confidence in the method’s robustness and scalability.

  • 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 #2

  • Please describe the contribution of the paper

    This paper proposed a new method for mixed-artifact removal in MRI through a cascading architecture using dynamically retrieved prompts.

  • 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. The novelty is that the method utilizes LLM-generated prompts to guide iterative artifact correction, effectively improving restoration quality.
    2. It demonstrates superior performance compared to state-of-the-art methods, with extensive validation including mixed artifacts and zero-shot experiments across multiple field strengths.
  • 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 method relies heavily on accurate artifact classification to select prompts, potentially leading to error propagation if misclassification occurs.
    2. The method is primarily validated on synthesized artifact types (motion, noise, bias), which may not fully capture complexities encountered in clinical scenarios.
  • 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.

  • 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/2025/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?

    a novel method with sufficient validations on mulitple datasets.

  • 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 #3

  • Please describe the contribution of the paper

    1. Cascading Restoration for Mixed MRI Artifacts: The paper introduces ResMAP, a novel framework that performs iterative artifact removal via prompt-guided cascading steps, effectively addressing the limitation of current AiOIR methods in handling mixed degradations without requiring joint training on all artifact combinations.

    2. Cross-Modal Prompt Retrieval Mechanism: A degradation-aware prompt retrieval system is proposed, leveraging large language model (LLM)-generated prompts for artifact-specific restoration. These prompts are encoded and fused with image features to dynamically adapt the network to different degradation types.

    3. Zero-Shot Generalization on Unseen Field Strengths and Artifacts: The framework demonstrates robust generalization to unseen MRI data, including external datasets from various field strengths (0.064T–5T), showing strong performance in zero-shot settings with real-world mixed artifacts.

  • 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. Novelty and Relevance of the Problem Setting: The paper addresses an important yet underexplored challenge in MRI restoration—removing multiple, simultaneously occurring, and unseen artifacts—through a cascading framework, which extends current AiOIR paradigms in a non-trivial and clinically meaningful direction.

    2. Innovative Use of LLM-based Text Prompts: The integration of semantically rich, artifact-specific prompts generated by GPT-4o introduces a new dimension to image restoration, enabling cross-modal guidance that significantly improves restoration accuracy and task adaptivity.

    3. Comprehensive and Convincing Evaluation: The authors present strong experimental results across synthetic and real-world scenarios, including ablation studies, external validations, and degradation-aware decision analysis (e.g., Fig. 5), effectively demonstrating the efficacy and interpretability of the proposed method.

  • 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. Limited Analysis on Prompt Quality and Retrieval Strategy: While the use of LLM-generated prompts is a key novelty, the paper lacks a thorough analysis of how prompt variability, quality, or selection mechanism affects restoration performance.

    2. Lack of Computational Efficiency Evaluation: The cascading nature of ResMAP introduces multiple iterations during inference, but there is no discussion on runtime, memory usage, or how the proposed method compares in computational cost to other baselines.

    3. Heuristic Iteration Termination: The stopping criterion for the iterative process is based on a simple ‘clean’ classification, which may be unreliable in ambiguous or low-SNR clinical scenarios. A more principled or learned termination strategy would strengthen the method’s robustness in practice.

  • 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.

  • 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/2025/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 should be accepted due to its originality, solid methodological foundation, and strong empirical evidence. The cascading framework for mixed artifact removal addresses a clinically relevant problem that is often overlooked in existing literature. The use of LLM-based prompts adds a creative and powerful mechanism for cross-modal adaptation, and the thorough evaluation—particularly the adaptive ordering experiments in Fig. 5—demonstrates the real potential of the approach. While there are areas for improvement, such as prompt selection interpretability and computational cost analysis, these are secondary to the paper’s significant contribution to the MICCAI community. The work opens up new directions in AI-assisted medical image restoration and is a promising step toward generalizable, real-world deployable solutions.

  • 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



Author Feedback

We appreciate the reviewers’ recognition of the novelty and contributions of our work. Below, we provide detailed responses to the raised concerns. 1.Artifact Synthesis and Generalization(R1) While primarily trained on synthesized artifacts based on MRI physics, our framework is also validated on real-world datasets (0.064T–5T) with naturally occurring mixed artifacts, showing promising generalization. We acknowledge the limited artifact diversity and plan to extend validation to more complex, clinically relevant degradations. 2.Classification Details(R1,R2) The classifier uses final-layer features from the Restormer encoder. Compared to CLIP-based encoders, our model captures multi-scale contextual information through progressive attention, yielding significantly higher classification accuracy and reducing error propagation from misclassification. Most errors occurred in cases with very subtle artifacts or “clean” images with slight intensity inhomogeneity, suggesting that our large-scale data-driven training provides a more objective assessment of image degradation than potentially inconsistent manual labels. 3.Data Training and Validation Separation (R2,Meta-R) Among the 120 HCP subjects, 100 were used for training. Each subject had four versions of data—clean, noise, bias, and motion—yielding 200 slices per subject by selecting 50 central axial slices from each version. Accordingly, the training set contained 100 × 200 = 20,000 slices. The remaining 20 subjects were used for testing, following the same procedure, resulting in 20 × 200 = 4,000 slices. For mixed artifact evaluation, 10 additional HCP subjects (independent of the previous 120) were used. Each subject was corrupted with four artifact combinations—noise + bias, noise + motion, bias + motion, and noise + bias + motion—with the order of artifact introduction randomized to avoid fixed patterns. From each version, 50 central axial slices were selected, yielding 200 slices per subject and 10 × 200 = 2,000 mixed-artifact slices in total. For zero-shot evaluation across field strengths, we used 10 samples each from the 0.064T LISA and 0.3T M4RAW datasets, and 3 healthy subjects from a 5T in-house dataset. 4.Explanation of the Cascaded Correction(R2,R3) To better illustrate the relationship between artifact severity and correction order, we focus on combinations of two artifact types. In most cases, the model can progressively detect and remove all artifacts, except in rare extreme scenarios. For instance, in Fig. 5a, 71% of the images were initially identified as bias and successfully corrected, while about 2% were detected as bias again in a second iteration. In such cases, applying bias correction twice can indeed be more effective than prematurely switching to another artifact type. The reverse setting refers to intentionally altering the model-predicted prompt order—for instance, forcing motion-first instead of the predicted bias-first—to assess the impact of correction sequence. Currently, iteration stops when the image is judged as clean or a maximum number of steps is reached, preventing over-correction. This strategy can be further improved by integrating a quality assessment module to define more adaptive stopping criteria. 5.Computational Efficiency and Prompts Analysis(R3) All models perform mixed artifact removal iteratively, as prompts are pre-generated and retrieved during inference. Compared to methods without prompts or with learnable ones, our approach shows no notable difference in per-iteration efficiency and is thus not further discussed. The compared methods adopt different prompt design strategies, with the ablation study including a no-prompt baseline. We generated 1,000 prompts per artifact type to ensure diversity and semantic richness, highlighting the benefits of LLM-driven prompting. Future work will further explore the impact of prompt quality and quantity on performance and better leverage cross-modal information.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The reviews agree that this is novel and well-investigated work. However, issues with data training and validation separation (including a withheld testing set) should be addressed in the updated paper.



back to top