List of Papers Browse by Subject Areas Author List
Abstract
All-in-one medical image restoration (MedIR) recoveries diverse degraded multimodal medical images with a single model, improving efficiency and deployment. However, its performance is often hindered by cross-modality task interference and optimization imbalance across tasks. To address these challenges, we propose a Dual-Scale Conditioned Evolving Network (DSCE-Net) with balanced optimization for task-adaptive restoration across multiple modalities and degradations. which dynamically evolves its parameters and feature representations during training to accommodate the varying complexities of different tasks. Specifically, we introduce a frequency-prompt-conditioned dynamic convolution that leverages a learnable prototype kernel bank to generate sample- and spatiallly-adaptive kernels, enabling effective modeling of both globally consistent and locally heterogeneous degradations across tasks and therefore reducing task-specific representation interference. In addition, we develop a progressive reweighting strategy that jointly considers update strength and convergence pace to alleviate optimization imbalance. Extensive experiments show that DSCE-Net consistently outperforms mainstream all-in-one models on multi-modal, multi-task MedIR benchmarks, demonstrating its effectiveness and superiority.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4298_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{HuXul_DSCENet_MICCAI2026,
author = { Hu, Xulin AND Chen, Yi AND Gu, Xianming AND Cao, Ying AND Wang, Lihui},
title = { { DSCE-Net: Dual-Scale Conditioned Evolving Network with Balanced Optimization for 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
The paper proposes DSCE-Net, an all-in-one medical image restoration model that addresses cross-task interference and optimization imbalance. The key idea is to use frequency prompt extraction (FPE) to encode modality-specific degradation and condition a prototype-bank dynamic convolution (PBDC) for adaptive global and local feature modeling (AGFA + ALTE). In addition, a progressive reweighting strategy balances tasks using gradient norms and convergence pace, improving multi-task training stability.
- 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 paper tackles an important problem of all-in-one MedIR and clearly identifies task interference and imbalance as key issues. The use of frequency-domain prompts (FPE) to guide dynamic convolution is well motivated and effectively models modality-specific degradation. The PBDC design is flexible and allows sample- and channel-adaptive filtering. The progressive optimization strategy is practical and improves training balance. Experiments show consistent gains across CT denoising, MRI super-resolution, and PET synthesis, and ablations validate each component.
- 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 overall novelty is moderate, as the method mainly extends existing ideas such as prompt-based conditioning and dynamic convolution. 2.The contribution of each component (FPE, PBDC, AGFA/ALTE) is not deeply analyzed beyond ablations, and their individual impact is somewhat incremental. 3.The progressive reweighting strategy relies on hyperparameters (e. g. , α, update interval T), but sensitivity and stability are not fully discussed. 4.The optimization formulation based on gradient norms and loss ratios lacks theoretical justification. 5.The method is evaluated on limited datasets (IXI, LDCT, PET), with no cross-dataset generalization study. 6.The model complexity (multiple AGLFE blocks with dynamic convolution) may impact efficiency, but runtime or memory cost is not reported. 7.Comparison is limited to all-in-one methods, and it is unclear how it performs relative to strong task-specific models.
- 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
If the authors can clarify the impact of design choices such as FPE and PBDC, as well as the sensitivity of the optimization strategy, it would strengthen the paper.
- 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 paper presents a solid and well-motivated approach for all-in-one medical image restoration, with consistent improvements across multiple tasks. The integration of frequency prompts and dynamic convolution is effective, and the optimization strategy addresses an important issue in multi-task learning. However, the overall novelty is somewhat incremental, and some design choices and generalization aspects are not fully analyzed. Overall, the paper is above the acceptance threshold.
- 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 authors have adequately clarified the previously raised concerns and provided sufficient technical details during the rebuttal process. The responses addressed the major questions regarding the methodology, implementation, and experimental design. The clarifications improved the overall clarity and strengthened the contribution of the work. Based on the post-rebuttal discussion, the paper should be considered for acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes a Dual-Scale Conditioned Evolving Network (DSCE-Net) with balanced optimization for all-in-one medical image restoration. Experimental results on multi-modal multi-taks MedIR benchmarks demonstrate the superior performance of DSCE-Net over mainstream all-in-one methods.
- 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 effectively addresses two critical limitations in all-in-one medical image restoration, i.e., cross-modality task interference and task-wise optimization imbalance. The research goal is technically meaningful. 2.Extensive evaluations are conducted on three representative tasks, including CT denoising, MRI super-resolution, and PET synthesis. The method outperforms six state-of-the-art all-in-one models in PSNR, SSIM, and RMSE. Ablation studies, visual comparisons, and t-SNE analysis consistently validate the effectiveness of each key component.
- 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.Frequency prompts, dynamic convolutions, and multi-task loss weighting have all been well explored in prior work.The contribution of this paper is mainly systematic integration rather than introducing a new paradigm. Moreover, the underlying motivations, i.e., cross-modality task interference and task-wise optimization imbalance, though meaningful, have already been discussed in previous studies (e.g., AMIR, TAT), which limits the novelty. 2.The FPE module only uses the log‑magnitude of FFT and completely ignores the phase component, which carries critical structural information in medical images. This design choice is not justified and may limit representation ability. 3.The paper does not provide visualization or quantitative analysis of which frequency bands are actually used by the frequency prompt. It is unclear whether the frequency domain really contributes to performance improvement or is just an empirical design. 4.Model size, FLOPs, and inference speed are not reported, leaving the computational cost and practicality unclear. 5.There is no comparison with single-task specialist models, making it difficult to quantify the performance trade-off introduced by the all-in-one design. 6.In Fig. 2, each CT image should be rotated 90° clockwise to ensure correct orientation and improve visual clarity. 7.The word “Evolving” appears in the title, but the paper does not clearly define or verify what is “evolving” during training or inference. This term is vague and misleading.
- 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?
Please see strengths and weaknesses above.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
1.This paper introduces a frequency-prompt-conditioned dynamic convolution that combines a learnable prototype-kernel bank to generate sample- and channeladaptive kernels, enabling expressive modeling of task-specific degradations.
2.This paper develops a progressive reweighting strategy that simultaneously accounts update strength and convergence pace, thereby alleviating optimization imbalance and improving overall restoration 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.
This paper is well motivated and clearly presented. The use of frequency prompting and loss balancing is valuable for all-in-one medical image restoration.
- 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.Please clarify the advantages of the Progressive Optimization Balancing Strategy over the Task-Adaptive Loss Balancing Strategy proposed in TAT.
2.Fig. 2 is too crowded, and the font size is too small. The authors should adjust it for better readability and presentation.
3.Some important references [1][2] should be discussed and cited to improve the completeness of the related works.
[1] Yang Z, Chen H, Qian Z, et al. Region attention transformer for medical image restoration[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 603-613.
[2] Yang Z, Zhou Y, Chen H, et al. UniPET: a universal network for high-quality PET image denoising across varied dose reduction factors[J]. Medical Image Analysis, 2026: 104059.
- 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 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
I encourage the authors to release the code after acceptance.
- 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 is well motivated and organized.
- Reviewer confidence
Very confident (4)
- [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 authors have adequately addressed my concerns. In my view, this is a high-quality paper that merits acceptance.
Author Feedback
@R1-q1 & @R2-q1 novelty: Although frequency prompts, dynamic conv, and loss weighting have been studied individually in prior works, our method is not a naive combination of these components. We design a task-driven coupling between frequency prompts and dynamic convolution, where the prompts explicitly guide the generation of task-adaptive convolutional features, forming an adaptive multi-task feature extraction mechanism. Similarly, our optimization module is fundamentally different from conventional loss weighting. We propose a progressive reweighting strategy that jointly considers both update strength and convergence pace. Unlike existing methods that rely on only one optimization signal, our approach integrates these complementary aspects into a unified weighting mechanism, enabling more reliable multi-task balancing and mitigating task dominance. @R2-q2, q3：Phase and Magnitude in FPE: Our prompt is not designed to capture structural details or specific frequency bands; instead, it distinguishes tasks/modalities and guides dynamic convolution to adaptively extract task-relevant features. Therefore, the prompt should be robust to task-specific variations, making the log-magnitude spectrum more suitable than phase due to its stability and lower sensitivity to noise. For this reason, we do not perform phase or band-wise analysis. Instead, we include prompt separability analysis and t-SNE visualizations to show that the learned prompts are discriminative across tasks/modalities, supporting the intended role of FPE. @R1-q2, q3 Ablation: We agree that the current ablation mainly provides quantitative evidence, while the motivations of PBDC and AGFA/ALTE could be further supported by qualitative visualization. In the revised manuscript, we will add visualizations showing how PBDC improves cross-task feature adaptation and how AGFA/ALTE promotes more balanced multi-task optimization. In addition, we have conducted sensitivity analyses on α and the update interval T, and the results show that our method is robust within a reasonable range of these settings. We will include these results in the revised manuscript. @R1-q4 & @R4-q1 Optimization formulation: We agree that the current version does not provide a formal theoretical proof, and we will clarify this in the revised appendix. Our progressive optimization balancing strategy is motivated by two complementary signals: instantaneous update strength and relative convergence progress. The gradient norm reflects how strongly a task affects the shared parameters at the current iteration, while the loss ratio indicates whether the task is learning faster or slower relative to its initial state. By combining them, our method characterizes task optimization status more comprehensively than TAT, which mainly relies on task-adaptive loss signals. This helps suppress task dominance and balance optimization more effectively. @R1-q5 Generalization: We will include cross-dataset experiments in the revised appendix to show generalization ability. @R1-q6 & @R2-q4 Complexity: We acknowledge that the current architecture introduces additional computational cost. We will report the parameter count, FLOPs, and inference speed for a clearer efficiency analysis and clarify this limitation in the revised version. @R1-q7 & @R2-q5 task-specific methods comparison: We compared our method with single-task specialist models and found that the all-in-one design introduces a performance trade-off. This suggests that the unified model does not outperform the specialists on every modality but achieves a more balanced optimization across tasks. We will add this comparison in the revised version. @R2-q6,q7, &@R4 q2,q3 format & reference: We will update Fig. 2 by rotating the CT images 90° clockwise and improve the readability of Fig. 1 with larger fonts and a cleaner layout. We will also revise the term “evolving” in the manuscript and add the suggested references [1,2] to the related work section.
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.
Please address whether the method provides a clear methodological advance beyond integrating existing frequency prompts, dynamic convolutions, and task-balancing strategies; whether the frequency and progressive optimization components are actually necessary and interpretable; and whether the all-in-one framework is practically competitive in terms of generalization, efficiency, and performance trade-offs.
- 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.
All reviewers recommended acceptance. After considering the reviewers’ feedback and the authors’ responses, I believe the paper meets the acceptance criteria and 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.
The authors have well addressed the reviewers’ concerns. Hence, my suggestion is Accept.
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.
After the rebuttal, the work consistently received positive scores from all three reviewers. The authors’ responses appear to have addressed the main concerns raised during the review process, and the reviewers’ post-rebuttal assessments indicate a generally favorable consensus. Given this positive outcome and the overall support from the reviewers, I recommend accepting the paper.