List of Papers Browse by Subject Areas Author List
Abstract
All-in-one medical image restoration (MedIR) aims to recover high-quality images from diverse degradations. However, existing methods predominantly rely on visual features and adopt a one-step mapping from low-quality inputs to high-quality outputs, which limits their ability to identify and adapt to heterogeneous degradations. To tackle these limitations, we propose TAIR, a novel Text-guided AdaptIve prompt Refinement framework for coarse-to-fine all-in-one MedIR. TAIR leverages dual-level textual prompts to guide restoration, where the task-level prompt specifies the restoration objective, indicating what to restore; and the instruction-level prompt directs how to restore via a coarse-to-fine strategy, with a coarse instruction recovering global structures and a fine instruction refining textural details. Based on this design, TAIR further constructs two composite prompts, i.e., the coarse composite prompt (CCP), which integrates degradation-related visual cues with task-level prompt and coarse instruction to recover global structures in the coarse phase; and the composite fine prompt (CFP), which combines task-level prompt with fine instruction and the coarsely restored image for detailed refinement in the fine phase. During restoration, an adaptive multi-prompt interaction (AMPI) module dynamically weights the contributions of each component in CCP and CFP for effective guidance. In addition, a text-driven quality-aware loss aligns restored images with high-quality textual descriptions while suppressing low-quality associations, enhancing overall fidelity and seman-tic consistency. Experiments across MRI, CT, and PET demonstrate that TAIR achieves competitive performance in single-task and all-in-one MedIR. Code is available at https://github.com/gluucose/TAIR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1045_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/gluucose/TAIR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CuiJia_TAIR_MICCAI2026,
author = { Cui, Jiaqi AND Xiong, Haocheng AND Han, Jize AND Liu, Bo AND Wu, Xi AND Wang, Yan},
title = { { TAIR: Text-Guided Adaptive Prompt Refinement for Coarse-to-Fine all-in-One Medical Image Restoration } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a text-guided framework for all-in-one medical image restoration, covering MRI super-resolution, CT denoising, and PET synthesis. The method combines task-level and instruction-level textual prompts with a coarse-to-fine restoration pipeline, and further introduces an adaptive prompt interaction module and a text-driven quality-aware loss. The topic is interesting, and the idea of introducing textual guidance into medical image restoration is potentially meaningful. However, the current experiments do not fully support the main claims of the paper, especially regarding the actual contribution of text guidance.
- 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 paper studies a practically relevant problem, since a unified model for multiple restoration tasks is more appealing than maintaining separate models for each modality or degradation type. 2.The overall framework is logically organized, and the combination of textual prompts, coarse-to-fine refinement, and adaptive prompt weighting is reasonably motivated. 3.The experiments cover several tasks and include both single-task and all-in-one settings, which gives the paper a relatively broad empirical scope.
- 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 paper emphasizes text guidance as its core novelty, but the experiments do not convincingly verify whether the textual information itself is truly helpful. Although there are ablations removing textual prompts, there is no deeper analysis showing whether semantically meaningful text is necessary or whether the model is actually using the text in a meaningful way. 2.The claim of state-of-the-art performance seems somewhat too strong. The comparisons are mainly against a limited set of recent methods, and the performance gains over the strongest baselines are often small. It would be more appropriate to describe the method as competitive rather than definitively state of the art. 3.The ablation study is not fully convincing. It mainly reports performance drops after removing components, but does not provide enough insight into why the proposed modules are necessary or whether simpler alternatives could achieve similar effects. This makes the support for the network design less solid.
- 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 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
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.
(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?
See weaknesses
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After carefully considering the author’s feedback and other reviews, I think this paper is well written and deserves an acceptance.
Review #2
- Please describe the contribution of the paper
The paper addresses the problem of all-in-one medical image restoration (MedIR). Specifically, it proposes a text-guided adaptive prompt refinement framework (TAIR) for MedIR. The core innovation of the method is to leverage dual-level text prompting strategy from coarse to fine, where the coarse stage focuses on overall restoration task text and restoration guidance while the fine stage offers structural and textural guidance to improve details. To usefully incorporate prompt-guidance, the paper proposes the adaptive multi-prompt interaction module (AMPI) which uses cross-attention mechanisms. Finally, a text-driven quality alignment loss uses a contrastive objective to ensure restored images are aligned with high-quality text instructions while suppressing alignment with degradation containing text. The paper claims state-of-the-art performance and generalization.
- 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.Leveraging textual instructions for MedIR is a novel and intuitively strong idea. 2.The idea of iteratively improving the restored result based on coarse and fine text prompts is interesting. 3.The paper is well written and easy to understand
- 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.Closeness to prior work: [1] proposes a very similar dual-prompting strategy for all-in-one image restoration but is neither discussed nor compared with. Specifically, TAIR’s proposed AMPI module looks very similar to Prompt-to-Prompt and Prompt-to-Feature interaction modules of [1]. The authors need to discuss the differences with [1].
[1] Li, Zilong, et al. “Prompt-in-prompt learning for universal image restoration.” Available at SSRN 6115113 (2023).
2.While the idea of using coarse to fine prompts is interesting, the paper does not investigate the sensitivity to the text prompts provided. For instance, are the prompt used during training diverse enough to accommodate varying text prompts in inference?
3.The paper claims it is the state-of-the-art, but the improvements are very marginal.
4.For out-of-distribution testing, what were the text prompts provided? Are those text instructions very different from the training tasks? If not, can the method work for a task which requires different text instructions than those seen during training?
5.For out-of-distribution testing, the other all-in-one models (AdaIr, AMIR, etc.) were excluded. Is there any particular reason for this?
6.Can the authors provide an additional metric such as LPIPS instead of RMSE? The reviewer feels PSNR and RMSE convey very similar information.
- 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
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 — 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?
As stated in the Major Weaknesses, the paper requires further clarification on its similarity to prior work, experiments and comparison setup.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addressed most of my concerns. My remaining concern is that the performance gains are not very significant. However, the authors mentioned that the writing in the final version would soften the claims of state-of-the-art.
Review #3
- Please describe the contribution of the paper
This paper introduces a text-guided prompt refinement framework for all-in-one medical image restoration, where hierarchical prompts are used in a coarse-to-fine manner. The integration of language guidance into unified restoration is interesting, but its necessity remains unclear.
- 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 coarse-to-fine prompting strategy is well motivated.
The framework is flexible and applicable to multiple degradation types.
- 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.It is unclear why natural language is needed instead of standard task embeddings (e.g., one-hot or learnable tokens). The benefit of semantic language guidance is not demonstrated.
2.Missing key baselines: random text, shuffled prompts, or learnable embeddings should be included to verify the actual contribution of text.
3.The coarse-to-fine design may introduce error propagation. Incorrect structures in the coarse stage may be amplified in the refinement stage.
4.The framework may aggravate over-smoothing, since the fine stage depends on already smoothed intermediate outputs.
5.The experimental validation is incomplete. Variants without text guidance or with frozen/random encoders are not provided. No out-of-distribution evaluation is conducted, which is important given the variability in medical imaging data.
- 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.
- 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.
(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?
See weakness
- 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 thank all the reviewers (R1, R2, R3, Mata-R). Q1: Actual necessity of text. (R1, R3, Meta-R) A1: We have trained TAIR completely without text (33.92), with learnable embeddings (34.22), random text (34.04), and shuffled prompts (34.01), all inferior to TAIR (34.44). (1) Why natural language over embeddings. First, instruction-level prompts provide guidance that simple learnable tokens cannot represent (Fig. 2, variant C). Second, coarse-to-fine framework requires distinct guidance per phase, while a shared task embedding would collapse this differentiation. Empirical gains over learnable embeddings further confirm. (2) Whether the model uses text meaningfully. Drops with random/shuffled prompts, where text exists but meaningless or mismatched, show the model relies on prompt semantics, not generic conditioning signal. We will strengthen these in final paper. Q2: Missing baselines and variants. (R3, Meta-R) A2: (1) The variants requested (without text, random text, shuffled prompts, learnable embeddings) are in A1, showing model learns meaningful prompt semantics to guide restoration. Variants (A) and (D) in Table 4 show removing textual prompts from either CCP or CFP degrades performance, verifying textual guidance at both stages. (2) Replacing frozen CLIP encoder with randomly initialized encoder drops avg PSNR to 34.09, confirming pretrained semantic benefit. Q3: Marginal gains and additional metric. (R1, R2, Meta-R) A3: (1) We will soften “state-of-the-art” to “competitive”. (2) For marginal-gain concern: First, TAIR ranks best across all tasks, with only 24.4M params vs TAT’s 41.69M, offering better accuracy-efficiency trade-off. Second, on OOD (Table 3), TAIR exceeds TAT by 0.77/0.91 PSNR on PET 20\times and sparse-view CT, which we believe are non-trivial and meaningful for clinical application. (3) LPIPS\downarrow: TAIR (0.082/0.061/0.109 on MRI/CT/PET) better than second-best TAT (0.084/0.063/0.123), confirming perceptual gains. Full results will be in final paper. Q4: Novelty of AMPI relative to PIP [1]. (R2, Meta-R) A4: PIP learns degradation via learnable embeddings; TAIR uses pretrained language semantics, driving distinct designs. First, to restore both global structure and fine details, TAIR decouples task identity from restoration strategy via different textual prompts, enabling phase-specific instructions for coarse-to-fine refinement; whereas PIP does single-step without decoupling. Second, to handle heterogeneous text/visual prompts, AMPI adaptively weights prompts per image via gating, in contrast to PIP’s static cross-attention. We will cite [1] and discuss our uniqueness in final paper. Q5: Sensitivity of text prompts. (R2, Meta-R) A5: Several paraphrased instructions show consistent performance with <0.4% PSNR variation, confirming robustness to rewording. OOD experiments (Table 3) use unseen task-level prompts, showing generalization beyond training vocabulary. Q6: Comparison and Setup in OOD. (R2, R3) A6: Table 3 reports OOD results. We reuse raining prompts, replacing only task-level text (e.g., “sparse-view CT reconstruction” for unseen degradation type). TAIR generalizes well to unseen texts. Other methods are excluded only due to space limit, with TAT included as second-best. TAIR surpasses all. Full results will be in final paper. Q7: Error propagation, over-smoothing, and ablation rationale. (R1, R3) A7: (1) Error is mitigated by L1 supervision on intermediate output, residual from degraded input, and AMPI’s integration, which together prevent sole dependency on intermediate output. Variant (C) shows large drop, confirming that coarse-to-fine preserves structures rather than harms. (2) Over-smoothing is prevented via degradation-free prompt which reinject guidance regarding detail restoration into fine phase. Removing this (variants E) degrades performance. (3) Simpler alternatives (variant F with plain cross-attention; variants in A1) yields inferior results. Analysis will be in final paper.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
While the framework is interesting, reviewers share significant concerns regarding the actual necessity of text guidance over simple embeddings and the technical novelty relative to prior work. The experimental evaluation is currently insufficient, with reviewers noting marginal gains, missing baselines (e.g., random/shuffled prompts), and sensitivity analysis. The authors are encouraged to address these critical gaps in their rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal has addressed the major concerns raised by reviewers. Finally, the paper receives 2 accept and 1 weak reject. Although R3 has not dealt with the rebuttal, I think the rebuttal has solved his/her major concerns. Thus, I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper introduces TAIR, the first text-guided all-in-one medical image restoration (MedIR) framework that simultaneously handles multiple modalities and degradations (MRI super-resolution, CT denoising, and PET synthesis) within a single model. By introducing dual-level textual prompts (task-level and instruction-level) combined with an Adaptive Multi-Prompt Interaction (AMPI) module and a text-driven quality-aware contrastive loss, the architecture provides explicit semantic guidance over a sequential coarse-to-fine trajectory. The paper received split post-rebuttal reviews (two Accepts, one Weak Reject), with reviewers strongly commending its conceptual novelty, clear presentation, and broad cross-modality empirical validation. Minor concerns remain regarding marginal quantitative gains and under-discussed baseline comparisons, but these are effectively cleared by the authors’ commitment to soften their state-of-the-art claims. Ultimately, the framework’s value is heavily justified by its parameter efficiency (24.4M parameters) and its robust zero-shot generalization capabilities on out-of-distribution clinical targets, such as unseen low-dose PET parameters and novel sparse-view CT artifacts. The paper satisfies publication standards and is supported for acceptance as a poster.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
While reviewer#3 did not change the decision after the authors’ response, I think the authors have addressed the comments appropriately. However, the paper should compared to task-specific image restoration DL designs to validate if it is beneficial to do all-in-one.