Abstract

Medical image segmentation commonly relies on U-shaped encoder–decoder architectures such as U-Net, where skip connections preserve fine spatial detail by injecting high-resolution encoder features into the decoder. However, these skip pathways also propagate low-level textures, background clutter, and acquisition noise, allowing irrelevant information to bypass deeper semantic filtering—an issue that is particularly detrimental in low-contrast clinical imaging. Although attention gates have been introduced to address this limitation, they typically produce dense sigmoid masks that softly reweight features rather than explicitly removing irrelevant activations. We propose ProSMA-UNet (Proximal-Sparse Multi-Scale Attention U-Net), which reformulates skip gating as a decoder-conditioned sparse feature selection problem. ProSMA constructs a multi-scale compatibility field using lightweight depthwise dilated convolutions to capture relevance across local and contextual scales, then enforces explicit sparsity via an ℓ1 proximal operator with learnable per-channel thresholds, yielding a closed-form soft-thresholding gate that can remove noisy responses. To further suppress semantically irrelevant channels, ProSMA incorporates decoder-conditioned channel gating driven by global decoder context. Extensive experiments on challenging 2D and 3D benchmarks demonstrate state-of-the-art performance, with particularly large gains (≈20%) on difficult 3D segmentation tasks. Project page:https://math-ml-x.github.io/ProSMA-UNet/

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5245_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{CheChu_ProSMAUNet_MICCAI2026,
        author = { Cheng, Chun-Wun AND Cheng, Yanqi AND Jing, Peiyuan AND Yang, Guang AND Montoya-Zegarra, Javier A. AND Schönlieb, Carola-Bibiane AND Aviles-Rivero, Angelica I.},
        title = { { ProSMA-UNet: Decoder Conditioning for Proximal-Sparse Skip Feature Selection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a U-Net variant that treats skip connections as a decoder-conditioned sparse feature selection problem rather than a dense reweighting problem. The key idea is to build a multi-scale decoder-encoder compatibility field and then apply an l1 proximal soft-thresholding operator with learnable per-channel thresholds so that irrelevant skip activations can be set to exact zero. This is complemented by decoder-conditioned channel gating. The paper argues that this yields more explicit suppression of noisy or semantically irrelevant skip information than standard attention-gated skip connections, and reports improvements on several 2D and 3D medical segmentation benchmarks.

  • 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 addresses a real and important weakness of U-shaped segmentation networks: skip connections are useful for localisation, but they can also leak low-level clutter and bypass semantic filtering. That motivation is sensible, especially in low-contrast medical images, and the paper stays focused on that specific design problem rather than trying to overcomplicate the architecture. 2) I like the technical clarity of the proposed mechanism. The method is easy to follow: construct a compatibility field from encoder and decoder features, enforce sparsity with a proximal l1 step, and then use spatial and channel gating together. Figure 1 communicates the idea well, especially the contrast between standard soft attention and explicit sparse selection with exact zeros. 3) The use of a proximal operator for skip gating is also an interesting design choice. Even if the underlying soft-thresholding idea is mathematically standard, its use here as a skip-filtering mechanism is reasonably novel in this context. The paper’s framing of skip fusion as feature selection rather than merely reweighting is conceptually sound. 4) The empirical results are broadly positive. On the reported 2D datasets, the method is consistently strongest, with gains over prior baselines on BUSI, GlaS, and Kvasir-SEG. On 3D, the improvement on Colon is particularly large, with F1 rising from 53.05 to 63.14 over the strongest listed baseline, while Spleen also improves modestly. The method therefore appears effective across both 2D and 3D settings. 5) The ablation study is simple but useful. It suggests that sparse spatial selection contributes the larger gain, while channel gating adds complementary value. That is at least consistent with the method’s intended design.

  • 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 not clear to me if the experimental comparison is fully balanced. On the 2D side, some baseline results are taken directly from prior work rather than reproduced here under a unified training protocol, and UKAN2.0 is missing on GlaS due to memory limits. That weakens the SOTA claim somewhat, because not all methods are being compared on equal footing. The paper also explicitly states that for BUSI and GlaS some baselines are taken from U-KAN under the same protocol, rather than rerun here. 2) The strongest gains occur on the 3D Colon task, but the paper does not provide enough analysis of why the gain is so large there. A nearly 10-point F1 jump over the best baseline is substantial. That could be a genuine strength, but it also demands more scrutiny: was Colon especially sensitive to noisy skip features, were there optimisation differences, or is the baseline set simply too narrow? At the moment the paper reports the number, but does not really explain it. 3) Related to that, the 3D baseline set feels limited. The comparison is against 3D U-Net, U-KAN 3D, and UKAN2.0 3D, but not against stronger contemporary 3D segmentation baselines such as nnU-Net-style pipelines or other established 3D U-Net variants like UNETR or Swin UNETR. Given that the method is a skip-connection modification rather than a wholesale architectural replacement, it would have been more convincing to compare against stronger and more standard 3D medical segmentation baselines. 4) The narrative around ‘explicit noise removal’ is plausible but not yet fully evidenced. Figure 1 is intuitive and the qualitative examples are encouraging, but the paper does not directly measure sparsity levels, feature suppression patterns, or whether harmful skip activations are actually reduced in a meaningful way. In other words, the mechanism is interpretable in theory, but the empirical validation of that interpretation is still limited. 5) Since ProSMA adds multiple components at each skip connection, a complexity analysis in terms of parameter count, FLOPs, memory footprint, or inference time is needed. Otherwise, it remains unclear whether the gains reflect better skip selection or are partly explained by increased parameter count and computational budget.

  • 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

    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?

    The paper proposes a clean and well-motivated modification to U-Net skip connections, with a technically coherent sparse gating mechanism and consistently positive results across several 2D and 3D segmentation benchmarks. I find the central idea sensible and practically relevant, and the gains on the more difficult 3D task are particularly encouraging. At the same time, the 3D comparison could be stronger, and the lack of complexity analysis makes the empirical gains harder to interpret fully. The reproducibility is at the current version limited. Overall, this is a solid paper with clear merit, though not one I would score at the very top end.

  • 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

    The authors turn skip connections in a Ushaped segmentation network into a decoder conditioned sparse feature selection step, instead of passing all encoder features directly or only softly reweighting them. The authors combine a multi scale compatibility field, sparse soft threshold selection, channel gating from decoder context, and a simple stability argument. The result is a method that aims to remove noisy skip responses more explicitly, while still keeping useful detail. The experiments across three 2D datasets and two 3D datasets make this contribution concrete, especially on the harder 3D setting.

  • 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 idea is focused and well motivated. Recent medical segmentation papers have improved U-shaped models through stronger configuration rules, transformer context, or redesigned backbones, as seen in nnUNet, SwinUnet, TransUNet, U-KAN, and Implicit U-KAN2.0.The authors take a different and useful angle by asking a simpler question, namely how to stop irrelevant skip features from entering the decoder at all. That makes the contribution easy to understand and practically meaningful. The paper is not only conceptual. The authors support the method with theory, ablation, 2D results, and 3D results. Table 1 to Table 3 show consistent gains, and the visual examples also match the quantitative findings. The authors test on ultrasound, histology, endoscopy, spleen CT, and colon CT, which gives confidence that the idea is not tied to one image type only.

  • 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.

    The major weaknesses are mild, but worth noting. The first is that the novelty is stronger at the level of skip gating design than at the level of opening a completely new segmentation direction. Medical image segmentation has already seen many strong U shaped extensions in recent years, including nnUNet, Swin Unet, TransUNet, U-KAN, and Implicit U-KAN2.0.So, in my view, the paper is best understood as a strong and well designed advance within an active research line, rather than a fully new family of models. The comparison is broad, but the case for state of the art would be even stronger if the authors also included very established references such as nnUNet and strong transformer based baselines such as TransUNet or Swin-Unet. Lastly, the experiments are strong benchmark experiments, but the paper does not yet include external clinical site validation.

  • 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

    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.

    (6) Strong Accept — must be accepted due to excellence

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors present a neat idea, good technical grounding, and convincing evidence. The paper is satisfactory and novel enough for acceptance. The authors explain the problem clearly, the method has a real point of difference, the ablation supports the design choices, and the gains on both 2D and 3D tasks are strong. The clarity of the core idea, the solid experimental section, and the fact that the work is both simple to communicate and useful for difficult medical segmentation settings.

  • 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 The paper proposes a novel proximal-sparse skip gating mechanism that constructs a multi-scale decoder–encoder compatibility field yielding closed-form soft-thresholding and exact removal of irrelevant skip activations. It also introduces a decoder-conditioned channel gating mechanism and provides a theoretical analysis proving that proximal sparse skip gating achieves exact feature selection, guaranteeing robustness to noise in the compatibility field. 2.The module design is supported by certain theories. The model is evaluated on multiple benchmark datasets with several quantitative metrics. Experiments demonstrate that the proposed modules improve performance compared with the baseline network. Extensive experiments on challenging 2D and 3D benchmarks demonstrate state-of-the-art performance, with particularly large gains on difficult 3D segmentation tasks.

  • 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 manuscript is clearly written and well structured, which makes the overall methodology and experimental setup straightforward to understand. There are no too many obvious grammatical or typographical errors. 2.Theoretical contribution: A new definition is proposed. The modulation form of skip regulation is defined as a decoder-conditioned sparse feature selection problem, rather than dense reweighting. 3.This paper analyzes the theoretical properties of the proposed proximal sparse gating operator to provide theoretical foundation for proposed model. 4.This paper addresses a related issue in the field of medical image segmentation, namely that the skip path can propagate low-level textures, background clutter, and acquisition noise. This is a challenging task with practical significance and reasonable motivation. 5.The model is evaluated on multiple benchmark datasets with several quantitative metrics. The method shows competitive performance against recent baselines, with consistent gains on the reported benchmarks. This suggests that the proposed design is effective in practice, even if the margins are not always large. 6.Visualization results indicate improved segmentation accuracy, especially for small lesions.

  • 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.Although the reported results are competitive, the margins over the strongest baselines are not always large. In some cases, the improvements are quite modest, raising questions about whether the improvements justify the additional architectural complexity, which makes it harder to judge how meaningful the gains are in practice. 2.While the integration is specific, the core components are common in recent medical imaging domain. The paper lacks a detailed analysis of parameter efficiency. There is a lack of numeric metrics to objectively prove how much the edge delineation has improved compared to baseline models. 3.The paper does not fully establish the fairness of the comparisons. It is not completely clear whether all competing methods were reproduced under the same training and preprocessing settings, or whether some results were taken directly from prior papers. More detail is needed here.

  • 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

    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?

    The paper addresses a relevant problem in medical segmentation and presents a technically reasonable model. The overall framework is coherent, and the results are generally competitive. This paper presents theoretical contributions that a new definition modulation form of skip regulation is proposed. The modulation form is defined as a decoder-conditioned sparse feature selection problem, rather than dense reweighting. This paper analyzes the theoretical properties of the proposed proximal sparse gating operator to provide theoretical foundation for proposed model. The model is evaluated on multiple benchmark datasets with several quantitative metrics. The method shows competitive performance against recent baselines, with consistent gains on the reported benchmarks.

  • 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

[Reviewer #1] :: [comparison is not fully balanced ] We clarify that only a subset of the 2D baselines was re-run in our experiments. For all reproduced results, we adopted the same training configurations as those used in the original UKAN papers for the corresponding baselines, thereby maintaining fairness and consistency in the comparison.

Q2.[3D Colon task] Our working explanation is that Colon tumor segmentation is particularly sensitive to noisy high-resolution skip features because of low contrast, irregular tumor shape, and surrounding anatomical clutter. ProSMA’s sparse decoder-conditioned skip filtering performed better in such cases.

Q3.[3D baseline set feels limited] Thank you for this suggestion. Our intention was to compare against architecturally relevant non-self-configuring segmentation frameworks, since our contribution focuses specifically on skip-connection design rather than dataset-specific pipeline optimization. For this reason, we compared against 3D U-Net, U-KAN 3D, and UKAN2.0 3D, where UKAN2.0 is a recent strong KAN-based segmentation model reported at MICCAI 2025.We did not include nnU-Net because it is a highly self-configuring framework involving automated preprocessing, architecture adaptation, training schedules, and post-processing, which would introduce additional pipeline-level factors beyond the architectural contribution studied here.

Q4.The proposed proximal operator is guaranteed by its L1 formulation to produce exact zeros, which is a mathematical property, not a tunable effect. The segmentation gains across all datasets (especially the large improvement on Colon) serve as the functional evidence that harmful skip activations are meaningfully suppressed. While sparsity percentages are not reported, the ablation study isolating spatial sparse selection confirms it as the primary source of improvement, which directly supports the interpretation that explicit zeroing—not just soft reweighting—drives the gains. Thus, the mechanism is both theoretically sound and empirically validated through task performance.

Q5.[complexity analysis] We therefore compared ProSMA-UNet with Attention U-Net on the BUSI configuration.Attention U-Net: 34.90M parameters ProSMA-UNet: 36.59M parameters. The proposed model therefore introduces only 1.69M (approx 4.84%). This demonstrates that ProSMA maintains a comparable model size to Attention U-Net.

[Reviewer #2] :: [More baselines] We thank the reviewer for this valuable comment. In our experiments, we have included comparisons with recent state-of-the-art methods, such as UKAN, published at AAAI 2025, and UKAN 2.0, published at MICCAI 2025, which serve as strong and competitive baselines. For nnU-Net, we note that it is a self-configuring segmentation framework, whereas our method is not designed as a self-configuring framework. To ensure a fair comparison, we therefore primarily compare our approach with non-self-configuring segmentation methods under comparable experimental settings.

[Reviewer #3] Q1.Although some individual margins are modest, the proposed ProSMA-UNet consistently improves over the strongest baseline across all evaluated settings, including 2D ultrasound, histology, endoscopy, and 3D volumetric segmentation. Specifically, ProSMA-UNet improves over the strongest competing method by +2.86 IoU / +2.43 F1 on BUSI, +1.12 IoU / +0.67 F1 on GlaS, +3.48 IoU / +2.07 F1 on Kvasir-SEG, +0.57 F1 on Spleen, and +10.09 F1 on Colon. Therefore, the contribution is not based on a single large gain on one dataset, but on a consistent improvement pattern across heterogeneous medical imaging tasks.

Q2.[parameter efficiency. ] We refer to Reviewer Q5. Q3.[fairness of the comparisons] We indicated that, for BUSI and GlaS, the baseline results are taken from U-KAN. For our reproduced results, we use the same training and preprocessing settings. [Reviewer #1] :: [comparison is not fully balanced ] We clarify that only a subset of




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.

    Reviewers have consistent high rating and merit of this paper. Strong technical sound and complete evaluation.



back to top