Abstract

Automated burn segmentation from color images remains challenging due to tissue heterogeneity, ambiguous burn boundaries, and limited pixel-level annotations. We propose UADA-MSPCNet, a semi-supervised segmentation framework that improves annotation efficiency and boundary delineation. The method integrates a dual-path cascaded architecture that combines high-resolution spatial encoding with hierarchical atrous contextual aggregation for joint structural and semantic modeling. To leverage unlabeled data, we employ an uncertainty-aware teacher-student module in which predictive uncertainty modulates consistency supervision and cross-scale pseudo-label learning. In addition, diffusion-generated burn images are introduced exclusively through the teacher pathway, enriching feature exposure while avoiding synthetic bias in student optimization. This teacher-side generative exposure improves pseudo-label calibration under limited supervision. Experiments on the EBIS dataset demonstrate consistent improvements over state-of-the-art supervised and semi-supervised methods across multiple labeling settings, achieving superior Dice scores and boundary accuracy. These results highlight the effectiveness of integrating cascaded contextual modeling, uncertainty-aware consistency, and controlled generative augmentation for reliable burn segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5539_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)

https://github.com/VEDAs-Lab/EBIS

BibTex

@InProceedings{ChaJoo_Learning_MICCAI2026,
        author = { Chauhan, Joohi AND Singh, Prabal Pratap},
        title = { { Learning to Segment Burns with Limited Labels: A Dual-Path Cascaded Network with Uncertainty-Guided Teacher Diffusion Exposure } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper addresses semi-supervised burn segmentation. It uses a dual-path encoder that combines a shallow spatial stream with a ResNet backbone and cascaded atrous convolutions. This is placed inside an uncertainty-aware mean teacher setup with cross-scale pseudo-supervision between the two pathways. The main design choice is to pass diffusion-generated synthetic burn images only through the teacher in a forward-only manner, so synthetic data shapes pseudo-labels through EMA and BatchNorm updates without directly affecting student training.

  • 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.

    Teacher only diffusion sounds interesting and seems novel, which can further be extended to other, more general problems.

    Experimental results support efficacy of the proposed approach.

    The application targets an important clinical problem, as burn annotation is expensive and expert-dependent, so a label-efficient method with preserved boundary fidelity is genuinely useful rather than benchmark-driven.

  • 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.

    Since the proposed method relies on diffusion, it would have been more informative to include one or two generative-model-augmented SSL baselines in the comparison.

    Some technical details are missing - e. g. , the definition of the uncertainty term in the loss or the exact form of the spatial and contextual predictions in CSPS.

    Evaluation is restricted to a single dataset, and the clinical generalization claim would be substantially strengthened by including an external cohort.

    The comparative SSL methods appear to have been re-implemented by the authors, and UA-MT underperforming MT in their table is unusual, raising the possibility that baselines are not tuned to their typical strength.

  • 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 provide sufficient information for 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?

    Although technical novelty seems to be decent, I am not fully convinced that the efficacy of the proposed method is well demonstrated. Only one small dataset was used, and reproducibility is questionable, especially for the baseline methods. See weaknesses for details.

  • 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

    Burn injury image segmentation faces challenges such as morphological heterogeneity and requirements for high precision edge detection. Balancing between contextual feature representation and boundary detection remain challenge in deep learning model design for burn image segmentation especially when labeled data is scarce. In this paper, a dual-path encoder is proposed that consists of two complementary streams: a high-resolution spatial stream that preserves boundary fidelity through three convolutional blocks, and a dense cascade atrous contextual stream that progressively captures multi-scale contextual information via a ResNet-101 backbone. This architecture enables the model to simultaneously exploit local spatial cues and global semantic dependencies.

  • 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.

    A semi-supervised teacher–student framework is designed based on the proposed dual-path backbone to enhance diffusion-generated samples through the teacher pathway by jointly optimizing representation learning and pseudo label calibration.

  • 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.

    Experiments are conducted on the dataset of burn images of 512 *512 resolution including 392 training and 208 test images. There is still a lack of experimental analysis on multicenter 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 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.

    (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?

    Burn injury image segmentation is critical for clinical treatment planning and assessment. However, burn image processing remains challenging due to the high variability in skin texture and precise edge detection requirement. This paper provides a clear and thorough analysis of the challenges in burn image processing, particularly the high variability in texture and the stringent requirements for edge detection accuracy. In response to the specific demands of burn image analysis, the authors design an algorithm that effectively balances texture recognition and boundary delineation performance. The research demonstrates a notable innovation for burn image analysis.

  • Reviewer confidence

    Somewhat confident (2)

  • [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
    • The paper proposes MSPCNet, a well-justified architecture that effectively decouples high-resolution spatial feature extraction from global contextual reasoning. The use of a regulated cascaded atrous formulation notably improves upon standard parallel pooling as it densifies contextual aggregation for complex tissue morphologies.
  • 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 teacher-side diffusion exposure provides a novel way to leverage generative data without polluting the student network’s training distribution. Standard data augmentation exposes the student network to synthetic artifacts. By restricting these images to the teacher, authors allow the framework to benefit from a broader feature space without polluting the student’s gradient-based learning.
    • The architecture is specifically designed to handle the clinical complexity of burn wounds by simultaneously focusing on fine boundary details and broad tissue patterns.
    • The framework intelligently identifies and ignores uncertain regions in unlabelled data, which prevents the model from reinforcing its own mistakes during training.
    • The method provided effectively addresses a significant bottleneck, the high cost and difficulty of obtaining expert annotations for complex skin injuries.
  • 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 introduction of a diffusion model into the teacher pathway creates a significant computational bottleneck during the training phase. The paper lacks a cost-benefit analysis regarding training time. If the performance gain is marginal compared to the student-teacher model using simpler generative augmentations, the practical utility of a diffusion-based teacher in a clinical pipeline is questionable.
    • In clinical practice, a model trained primarily on lighter skin tones may fail to accurately delineate burn margins on darker skin, where contrast between healthy and injured tissue is reduced. There is no demographic breakdown of performance.
    • If the diffusion model is trained only on the same limited labeled set, it is essentially just a complex interpolator of that small sample size. It’s unclear whether the diffusion model is actually providing new semantic information or simply rearranging the same sparse features. The authors should clarify whether the diffusion model was pretrained on a large external dataset.
  • 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.

    (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 demonstrates substantial improvements over the current state-of-the-art medical segmentation model, nnU-Net, particularly in low-label scenarios.
    • Teacher-side diffusion exposure is a good technical contribution that effectively utilizes diffusion models to enrich feature representation while successfully avoiding the common pitfall of introducing synthetic bias into the student network’s optimization.
    • The MSPCNet dual-path design is well-reasoned for the specific clinical challenge of burn segmentation, effectively balancing the need for fine boundary delineation with global contextual aggregation of heterogeneous tissue.
    • The authors conducted a thorough evaluation using five random splits and multiple metrics, proving that the model excels not just in overlap, but in precise boundary localization as well.
    • The ablation studies clearly validate that each component contributes measurably to the final performance.
  • 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 careful evaluation, constructive feedback, and positive assessment of the paper. We are encouraged that the reviewers and meta-reviewer recognized the clinical relevance of the problem, the effectiveness of the proposed MSPCNet framework, and the novelty of the teacher-only diffusion exposure strategy for low-label burn segmentation.

Regarding the diffusion framework, the diffusion model was trained exclusively on the EBIS training split with strict train/test partition separation and without external datasets or additional annotations. Importantly, synthetic samples were introduced only through the teacher pathway rather than as direct supervision for the student network. Teacher parameters remained governed solely through EMA updates from the student, while synthetic exposure influenced only teacher-side feature statistics and pseudo-label formation. The objective was to improve pseudo-label calibration while avoiding synthetic-distribution bias during student optimization. Consistent with this motivation, the ablation study in Table 3 of the submitted manuscript showed that applying diffusion augmentation to both teacher and student reduced both segmentation performance and calibration quality, whereas restricting synthetic exposure to the teacher improved Dice, HD95, and ECE.

Regarding computational cost, the diffusion component was used only during training and was not involved during inference. Final deployment, therefore, requires only a single forward pass through the student network without additional inference overhead. Teacher-side diffusion exposure introduced approximately 18–20% additional training overhead while consistently improving Dice, HD95, and calibration under limited supervision.

We also appreciate the request for additional methodological clarity. The uncertainty-aware consistency formulation and CSPS supervision mechanism are described in Sec. 2.2–2.3 through uncertainty-modulated consistency weighting, temperature-scaled pseudo-label sharpening, and confidence-guided supervision between spatial and contextual pathways. We will further improve the exposition of these components for clarity and reproducibility in the camera-ready version within the permissible revision limits.

Regarding reproducibility and baseline fairness, all comparative SSL methods were evaluated using identical preprocessing, augmentation policies, train/test splits, optimization schedules, and evaluation protocols. The observed UA-MT behavior remained consistent across five random splits under the same experimental settings.

Finally, we agree that broader demographic and multi-center validation remain important future directions. The current EBIS dataset does not provide demographic subgroup annotations for reliable stratified analysis across skin-tone distributions, and this limitation has already been acknowledged in the manuscript.

We thank the reviewers and meta-reviewer again for their thoughtful comments and valuable suggestions.




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.

    Strong clinical relevance and novel teacher-only diffusion strategy warrant acceptance despite the single-dataset limitation. Authors should clarify diffusion pretraining, add computational cost, and discuss demographic bias.



back to top