Abstract

Training accurate medical image segmentation models requires large amounts of densely annotated data, which is costly and time-consuming to obtain. Semi-supervised learning (SSL) alleviates this by learning from both abundant unlabeled data and limited labeled data. However, most modern SSL methods rely on pseudolabels for unlabeled data, and typically assess their reliability through model confidence or uncertainty, measures that are self-referential and lack explicit grounding in segmentation quality. Instead, we propose a quality-guided SSL framework that trains a dedicated network to estimate segmentation quality from image-mask pairs. The predictor is trained on variable-quality masks generated through synthetic corruptions augmented with imperfect outputs from partially trained segmentation models, capturing realistic error patterns encountered during training. We integrate the quality predictor into SSL through two complementary mechanisms: a quality-aware regularization loss and a quality-based pseudolabel sample reweighting scheme. We show that our method serves as a drop-in enhancement to existing SSL frameworks. Extensive experiments across five datasets and multiple architectures demonstrate consistent improvements over competing SSL methods, advancing the state-of-the-art in semi-supervised medical image segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4232_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/sfu-mial/QG-SSL

Link to the Dataset(s)

PH2 dataset: https://www.fc.up.pt/addi/ph2%20database.html Skin Cancer Detection dataset: https://vip.uwaterloo.ca/skin-cancer-detection/ DermoFit dataset: https://licensing.edinburgh-innovations.ed.ac.uk/product/dermofit-image-library CVC-ColonDB dataset: https://pages.cvc.uab.es/CVC-Colon/index.php/databases/ CVC-ClinicDB dataset: https://pages.cvc.uab.es/CVC-Colon/index.php/databases/

BibTex

@InProceedings{AbhKum_QualityGuided_MICCAI2026,
        author = { Abhishek, Kumar AND Hamarneh, Ghassan},
        title = { { Quality-Guided Semi-supervised Learning for Medical Image Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

The authors bridge two research directions, i. e. , segmentation quality prediction and semi-supervised learning, by training a quality predictor that estimates segmentation performance from image–mask pairs and leveraging it to provide supervisory signals for unlabeled data. This idea is sufficiently novel for presentation at MICCAI. In addition, several technical components, including a differentiable quality regularizer, a pseudo-label reweighting scheme, and a mask corruption strategy, further strengthen the proposed framework.
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 is well organized and clearly written. The proposed method is technically sound and demonstrates a reasonable level of novelty. Its effectiveness is supported by comprehensive ablation studies and comparative experiments.
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.

I do not identify any significant weaknesses in the paper.
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
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?

My overall assessment is primarily driven by the novelty and technical soundness of the proposed approach. The idea of bridging segmentation quality prediction with semi-supervised learning is both meaningful and sufficiently original, and the proposed framework is well-motivated and coherently designed.
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 introduces a quality-guided SSL framework for medical image segmentation. It trains an independent quality predictor to estimate segmentation quality and integrates it into existing SSL pipelines through two ways: Quality-Aware Regularization (QAR) and Quality-Weighted Pseudolabels (PL-QW). Various experiments are conducted across five datasets, different SSL paradigms, and several architectures.
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.It is highly useful and interesting to explore such plug-and-play components to enhance current SSL methods by using an external, image-grounded quality predictor to guide unlabeled data learning. 2.The framework is designed to be model-agnostic and serves as a drop-in enhancement, with very detailed experiments covering diverse modalities and multiple backbone architectures.
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.Hyperparameter sensitivity is mentioned but not thoroughly analyzed, and cross-domain generalization of the quality predictor is not evaluated. 2.The novelty is primarily in the integration strategy rather than the quality prediction module itself. And the approach builds upon existing quality assessment literature. 3.Lack of visualization of the proposed method (e.g. result of quality predictor)
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
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?

This paper practical quality-guided semi-supervised framework for medical image segmentation by training an independent quality predictor and integrating it through two model-agnostic mechanisms with experimental breadth. It demonstrates consistent gains across multiple SSL paradigms, and several architectures. Despite the limitations discussed before, I believe it is a meaningful work .
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 #3

Please describe the contribution of the paper

This paper introduces a quality‑guided semi‑supervised learning (SSL) framework for medical image segmentation that explicitly models segmentation quality and uses it as a training signal, rather than relying on model confidence or uncertainty proxies. The core idea is that segmentation quality—how well a predicted mask aligns with image evidence—can be learned independently and then leveraged to guide SSL more effectively across datasets, architectures, and SSL paradigms. The framework operates in two phases. In Phase 1, the authors train a segmentation quality predictor that estimates the Dice similarity coefficient (DSC) of an image–mask pair without access to ground truth. To construct training data for the quality predictor, they generate a synthetic dataset from a small labeled set by applying stochastic degradations to ground‑truth masks. These include (i) classical morphological and geometric corruptions and (ii) predictions from weakly trained segmentation models captured at early epochs, which introduce realistic error patterns seen during SSL. Each corrupted mask is paired with its true Dice score relative to the ground truth. The quality predictor is implemented as a ResNet‑18 regression model trained with Smooth L1 loss. Extensive ablations show that incorporating weak‑model corruptions, using a moderate number of corrupted masks per image, and choosing an appropriately sized backbone are all important for accurate quality prediction. Across five datasets, the model achieves low mean absolute error (0.043–0.088) and high Pearson correlation (ρ > 0.92), and control experiments confirm that it relies on image context rather than mask statistics alone. In Phase 2, the frozen quality predictor is used to guide semi‑supervised segmentation training on the labeled data and a much larger unlabeled set, drawn from different sources to reflect realistic domain shifts. Two complementary mechanisms are proposed. Quality‑Aware Regularization (QAR) feeds the model’s soft predictions on unlabeled images into the quality predictor and backpropagates gradients from the predicted quality score, encouraging the segmentation network to produce masks that the quality predictor judges as high quality—without generating explicit pseudo‑labels. Quality‑Weighted Pseudo‑Labels (PL‑QW) instead uses the quality predictor from phase 1 to weight the loss associated with any pseudo‑labels generated by an SSL method, down weighting unreliable pseudo‑labels and upweighting high‑quality ones. Importantly, PL‑QW is framework‑agnostic and can be added to existing SSL methods without architectural changes. The method is evaluated on five medical image segmentation datasets spanning dermatology and colonoscopy, using three diverse segmentation architectures (U‑Net++, Attention U‑Net, and Swin‑Unet). The results show that predicted segmentation quality is a more effective training signal than model confidence or uncertainty, and that learning quality independently enables both strong performance and broad applicability. Overall, the paper presents a general, modular, and empirically validated approach to incorporating learned segmentation quality into semi‑supervised medical image segmentation, achieving state‑of‑the‑art performance while remaining compatible with a wide range of existing SSL frameworks.
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 introduces a context‑aware, independently trained quality predictor that estimates true segmentation quality (Dice) from image–mask pairs and uses it to guide semi‑supervised learning. This moves beyond confidence or uncertainty heuristics and provides a more direct and semantically meaningful supervision signal. 2.Quality‑Aware Regularization (QAR) and Quality‑Weighted Pseudo‑Labels (PL‑QW) demonstrate strong generality across datasets, architectures, and SSL paradigms. Across five datasets spanning two medical imaging modalities (dermatology and colonoscopy), QAR consistently achieves the best or tied‑best performance among all compared methods. For example, with Swin‑Unet on PH2, QAR improves DSC from 93.02 ± 0.55 (supervised‑only) to 95.58 ± 0.42 (+2.56), and on SCD from 91.54 ± 0.58 to 93.72 ± 0.49 (+2.18). Similar gains are observed for polyp segmentation: on CVC‑ColonDB, QAR reaches 92.70–93.36 DSC across architectures, outperforming Mean Teacher, ICT, contrastive learning, and cross‑pseudo supervision despite using a single segmentation network. Importantly, these improvements hold across three diverse architectures—U‑Net++ (26.1M parameters), Attention U‑Net (24.7M), and Swin‑Unet (34.3M)—indicating that the gains are not architecture‑specific. In parallel, PL‑QW acts as a robust drop‑in enhancement: quality‑weighted variants outperform their original SSL counterparts in nearly all settings (e.g., MT‑QW, ICT‑QW, CPS‑QW), often yielding +1–2 DSC improvements over the unweighted baselines. The consistency of gains across datasets, models, and learning paradigms demonstrates that leveraging predicted segmentation quality provides a broadly applicable training signal rather than one tuned to a specific dataset or SSL method. 3.The paper systematically studies the impact of backbone choice, corruption strategy, number of degraded masks, loss weights, and unlabeled data size, showing that the method is robust and not overly sensitive to hyperparameter tuning. Across five quality‑predictor backbones spanning a 6× range in model size (4.3M–28.0M parameters), performance varies modestly, with ResNet‑18 achieving the best MAE/ρ (0.043 / 0.972), while even suboptimal choices (e.g., EfficientNet‑B0) remain competitive (MAE 0.054 / ρ 0.961). Incorporating weak‑model corruptions consistently improves quality prediction. Increasing the number of degraded masks per sample improves performance up to a clear saturation point: MAE drops from 0.061 (K=10) to 0.043 (K=50), with only marginal changes beyond that (MAE 0.047 at K=100, 0.050 at K=200). The results presented in the paper show that the method’s gains are not driven by fragile hyperparameter choices but persist across a wide and practically relevant operating range.

Overall, the paper is well written and practically grounded, evaluating labeled and unlabeled data from different sources to reflect real‑world domain shift while clearly separating quality estimation from segmentation for stability and reuse. It also demonstrates efficient use of unlabeled data, where accuracy saturates early but convergence accelerates significantly as unlabeled samples increase, highlighting strong practical training benefits.
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.While the method performs strongly overall, the paper does not deeply analyze cases where gains are marginal or absent (e.g., the DMF + U‑Net++ setting for CL‑QW). Add a short failure‑case analysis section highlighting scenarios where quality guidance helps less. Maybe visualize representative examples (good vs. bad cases) to illustrate when quality prediction may be less reliable. 2.For PL‑QW, the exact mechanism used to generate pseudo‑labels (hard vs. soft, thresholding strategy, teacher vs. student predictions) is not fully specified, which affects reproducibility and interpretation. Existing semi-supervised segmentation pipelines use a wide range of pseudo-label generation strategies—from simple argmax to Confidence-Aware Cross Pseudo Supervision (CACPS), where the KL-divergence between predictions under different conditions serves as a confidence measure, and low-quality labels receive less influence through exponential weighting. The authors should specify their pseudo-labeling procedure precisely so readers can understand how PL-QW interacts with the base method. Add a short appendix summarizing pseudo‑label pipelines for all methods. 3.Although the paper argues that QAR is more efficient than multi‑network approaches like CPS, no quantitative runtime or memory comparisons are provided. The two-phase protocol requires: (1) training multiple weak U-Net checkpoints, (2) generating K corrupted masks per labeled image, (3) training gφ on the resulting DQ, and (4) evaluating gφ for every unlabeled sample at every training iteration of Phase 2.For QAR specifically, this evaluation must be differentiable (requiring backpropagation through gφ). The paper should report the computational cost relative to baseline methods. Is gφ lightweight relative to fθ? How does K affect gφ training time and quality? Existing uncertainty-based approaches (Monte Carlo Dropout, Ensembles methods) already face criticism for requiring 10 to 100 network evaluations during inference, leading to significant computational overhead. 4.While the quality predictor is validated numerically, the paper does not explore how its predictions relate to clinically meaningful segmentation errors. The authors should add qualitative visualizations as appendix (or to main paper) showing masks with high vs. low predicted quality alongside ground truth. This can also help in analyzing correlation between predicted quality and specific error types (boundary leakage, missing lesions, false positives).

Overall, the paper flows well with good scientific backing but there are some minor spacing, formatting and grammatical issues that can be improved. Such as use of numeral consistently (5 vs five), terminology standardization (segmentation network vs segmentation model, plurality (missing eqn vs eqns ), long sentences that can be split.
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
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?

This paper presents a novel two-phase semi-supervised framework for medical image segmentation that first learns an independent, context-aware quality predictor to estimate segmentation accuracy from image–mask pairs, then uses it to guide segmentation training on large unlabeled datasets without requiring ground truth. It should be accepted because it introduces a simple but new idea, is carefully tested on many datasets and models, and shows reliable improvements without heavy tuning.
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

N/A

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.

The reviewers provide consistently strong support for this paper (scores: 5, 5, 6), highlighting the clear novelty, technical soundness, and broad practical impact of the proposed quality-guided semi-supervised framework . In particular, the idea of explicitly modeling segmentation quality and integrating it into SSL through model-agnostic mechanisms is regarded as both original and highly effective, with comprehensive experiments demonstrating consistent improvements across datasets, architectures, and SSL paradigms. While a few minor concerns are raised regarding additional analysis and clarity, these do not detract from the overall strength of the contribution. Based on the unanimous positive feedback and my own assessment, I recommend acceptance.

back to top

Quality-Guided Semi-supervised Learning for Medical Image Segmentation

Author(s):