List of Papers Browse by Subject Areas Author List
Abstract
Reliable fundus segmentation is crucial for ophthalmic diagnosis, yet label scarcity and cross-center domain shifts severely degrade the generalization of automated models. Although collaborative frameworks incorporating foundation models show promise for Mixed-Domain Semi-Supervised (MiDSS) learning, they suffer from coarse-grained reliability estimation and representational collapse. Representational collapse forces the conventional model to forfeit its unique inductive biases, preventing effective synergistic inference. To address these challenges, we propose TrustSyn, a synergistic Teacher-Student framework for MiDSS learning, which bridges the gap between label scarcity and domain shifts. To rectify reliability estimation, we introduce a Spatial Reliability Assessment (SRA) coupled with an Uncertainty Weighting (UW) strategy. This mechanism dynamically modulates pixel-level supervision to leverage global context priors for reliable pseudo-label generation selectively. To avert representational collapse, we introduce a Feature Discrepancy (FD) constraint between the latent representations of both models. By enforcing feature divergence, FD ensures output consensus is achieved through diverse inference mechanisms rather than redundant representations. Extensive experiments on the Fundus dataset demonstrate that TrustSyn exhibits robustness against domain shifts, achieving state-of-the-art performance with a Dice score of 89.17% and an ASD of 3.20. Code is available at https://github.com/wwwangxinhui/TrustSyn.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0870_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wwwangxinhui/TrustSyn
Link to the Dataset(s)
Fundus dataset: https://drive.google.com/file/d/1p33nsWQaiZMAgsruDoJLyatoq5XAH-TH/view
BibTex
@InProceedings{QiaBao_TrustSyn_MICCAI2026,
author = { Qiao, Baojun AND Wang, Xinhui AND Zhang, Wanjun AND Turuwhenua, Jason AND Norouzifard, Mohammad},
title = { { TrustSyn: Reliable and Divergent Synergy for Mixed-Domain Fundus Segmentation } },
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 presents a collaborative teacher-student framework for mixed-domain semi-supervised fundus segmentation that combines a foundation model with a conventional model. Its main contribution is a spatial reliability and feature discrepancy design that improves pseudo-label quality while preserving complementary representations under domain shift.
- 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.Mixed-domain semi-supervised fundus segmentation is an important yet challenging scenario, as it combines label scarcity with domain shift across centers and devices. The problem is well motivated from both the methodological and application perspectives. 2.Integrating MedSAM with U-Net is a reasonable and practically relevant design. The paper does not treat the foundation model merely as a stronger teacher, but attempts to combine its robust global priors with the conventional model’s adaptability to domain-specific details, which is an appealing perspective. 3.The proposed SRA mechanism goes beyond image-level confidence by estimating reliability at the local spatial level, which is particularly relevant for fundus structures with ambiguous cup-disc boundaries. This fine-grained treatment is interesting because it allows supervision to adapt to local uncertainty instead of relying on a single global criterion.
- 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.In the current design, the spatial reliability map is constructed from the agreement between the student U-Net and teacher U-Net, as well as between the teacher U-Net and teacher MedSAM. While this is intuitive, local agreement does not always imply local correctness, especially in low-contrast or ambiguous cup-disc boundary regions where multiple branches may share similar biases. Moreover, the reliability estimation already depends in part on the current student state, which may introduce a degree of circularity in pseudo-label selection. 2.The paper motivates FD as a way to encourage feature orthogonality and prevent representational collapse. However, the implemented loss minimizes cosine similarity between projected features, which more directly pushes the two branches toward opposite directions rather than strict orthogonality. The paper argues that the primary segmentation objective balances this effect in practice, but this mechanism would benefit from clearer explanation or stronger empirical validation. 3.The framework involves multiple important settings, including the sliding-window configuration and overlap ratio in SRA, the Hanning weighting, the parameters in uncertainty weighting, the confidence threshold for reliable regions, and the balancing coefficients in the total objective. Since these choices affect pseudo-label generation, supervision strength, and feature regularization simultaneously, more systematic sensitivity analysis would strengthen confidence in the method’s robustness.
- 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 addresses a meaningful and timely problem, and the proposed framework is reasonably coherent and technically well motivated. My score is slightly above the acceptance threshold because the main ideas are interesting and relevant, although some core design assumptions and the robustness of several components would benefit from stronger justification or analysis.
- 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 author has addressed my concerns, and if the author revises the paper as promised, I believe it will meet the acceptance criteria.
Review #2
- Please describe the contribution of the paper
This paper addresses the problems of coarse-grained reliability estimation and representational collapse that arise when conventional models and foundation models are jointly used in mixed-domain semi-supervised learning. It proposes a collaborative teacher-student framework, TrustSyn, which includes SRA, UW, and FD to address these issues. The method is validated on fundus image segmentation.
- 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 proposes a mixed-domain semi-supervised method to address coarse-grained reliability estimation and representational collapse, which offers a certain degree of novelty. 2.The comparison and ablation studies on fundus image segmentation support its effectiveness. 3.The paper is well organized and logically clear.
- 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.In the proposed SRA, what are the differences between Self-Confidence and Mutual-Confidence and the Self-Confidence and Mutual-Confidence introduced in Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation? I do not see a proper citation or a clear discussion of what is adopted and what is different. If this work introduces improvements, the paper should provide evidence that these improvements are effective. Although the current ablation includes an analysis of SPA, it is still unclear whether the performance gain comes from the original Self-Confidence and Mutual-Confidence in SynFoC or from the proposed modifications. 2.What does $\lambda_h$ in Eq.~(5) denote? This symbol is not defined. 3.In Table~1, although TrustSyn achieves better results on Domains 1 and 4, its performance is much worse than that of existing state-of-the-art methods on Domains 2 and 3, especially on Domain 2.An analysis of these failure cases is important for understanding the boundary of the proposed method. 4.In Table~2, some components bring only marginal improvements. It is unclear whether these small gains are statistically meaningful or simply due to randomness. 5.The ablation study lacks an analysis of sensitivity to key hyperparameters.
- 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.
(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?
The paper has some potential, but the current version still falls slightly below the acceptance threshold in terms of novelty justification and experimental support. I may consider raising my score if the authors can adequately address these concerns in the rebuttal.
- 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 addressed my main concerns in the rebuttal regarding the differences from prior work, the experimental analysis, and the novelty of the proposed method. These concerns have been satisfactorily resolved.
Review #3
- Please describe the contribution of the paper
The paper proposes a hybrid-domain semi-supervised fundus image segmentation method based on foundation models and a teacher–student collaborative framework. This method dynamically optimizes pixel-level supervision through a spatial reliability assessment and uncertainty-weighting mechanism, thereby generating high-quality fine-grained pseudo-labels. Meanwhile, by minimizing cosine similarity, it enforces the latent representations of the foundation model and the conventional model to remain orthogonal, effectively preventing representation collapse during joint training and significantly improving robustness to distribution shifts. The authors validate that the proposed method achieves state-of-the-art performance on public datasets.
- 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 problem addressed is novel, and the identification of representation collapse in the collaborative framework between foundation models and conventional models is insightful and inspiring. (2) The model architecture is logically clear and well structured, representing a strong extension and improvement over popular methods in this field. (3) The method achieves strong experimental performance. The ablation studies are comprehensive, and the progressive integration of different modules yields consistent performance gains.
- 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 proposed method introduces two lightweight modules on top of the foundation-model-based teacher–student collaborative framework. While the network design is reasonable, the level of novelty is somewhat limited. Using Shannon entropy to generate uncertainty maps and modulate supervision signals is a common practice in semi-supervised learning, and the FD module largely reuses the cosine-similarity minimization strategy from [25].
(2) The term 𝜆 in Equation 5 is not explained in the paper. (3) Based on the statistical orthogonality observed during cosine-similarity minimization, the paper directly claims that the latent representations are kept orthogonal. This statement is not mathematically rigorous enough.
- 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 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?
This work is of positive significance to the relevant field, but there are still concerns regarding reproducibility. I recommend that the authors provide more detailed implementation information as well as the source code.
- 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
Author Feedback
We thank the AC and Reviewers for the comments. 1.Novelty & Relation to Prior Work (R3, R4) SRA vs. SynFoC (R3): SRA upgrades SynFoC from coarse image-level confidence to fine-grained spatial confidence. 1) Methodological difference: Unlike SynFoC’s scalar confidence, SRA constructs spatial reliability maps via sliding windows and Hanning weighting, capturing boundary heterogeneity. 2) Empirical net gain: In Tab. 2, “Base” is original SynFoC. “Base+SRA” replaces its image-level confidence with our spatial confidence. This replacement yields a consistent performance improvement, representing the net gain of our adaptation. Tab. 1 shows TrustSyn outperforms the full SynFoC baseline on average. We will add proper citations. FD & UW (R4): We uniquely adapt established math to integrate Foundation Models (FMs) with conventional models. 1) Heterogeneous regularization (FD): Unlike [25], which minimizes cosine-similarity to prevent homogeneous CNN collapse globally, TrustSyn addresses representational collapse between highly heterogeneous architectures (MedSAM and U-Net) by selectively applying FD within our dynamically estimated reliable regions. 2) Differentiable soft-gating (UW): Instead of binary thresholding, we leverage the FM’s zero-shot entropy for a differentiable soft-gating mechanism, preventing abrupt supervision loss in ambiguous zones under severe domain shifts. 2.Math Rigor & Terminology (R2, R4) R2 and R4 correctly note our FD loss (minimizing 1+cos) pushes representations toward opposite directions rather than strict orthogonality. We will refine our terminology to “feature divergence”, clarifying its mechanism through two aspects. 1) Theoretical equilibrium: While L_dis pushes features apart, the primary segmentation objective constrains both branches to maintain semantic coherence. Since strict diametric opposition prevents shared decoding, the system reaches a dynamic equilibrium, a phenomenon also observed in [25]. 2) Empirical validation: Consistent with [25], we will add an empirical analysis of training dynamics, confirming L_dis reliably converges around 1.0 (cos near 0), providing the statistical validation requested by R2.3.Domains 2/3 & Failure Analysis (R3) We acknowledge the Domain 2 performance gap compared to reported SOTA and analyze its causes. 1) Unified setup: Reproducing baselines under a unified codebase introduces natural metric variations, yet TrustSyn remains highly competitive. 2) Failure analysis: Per R3, we analyzed Domain 2 failure cases. In severely degraded images with low contrast and blurry boundaries, TrustSyn acts conservatively: UW suppresses high-entropy signals and FD remains inactive. This limits aggressive adaptation but prevents confirmation bias, yielding the highest overall robustness. We will incorporate this discussion. 4.Hyperparameters & Stats (R2, R3) We verify parameter robustness and statistical significance. 1) Parameter robustness (R2): SRA is insensitive to window/overlap partitions due to Hanning weighting. Coefficients (α, β) are stable across broad ranges. UW’s soft-gating prevents hard thresholding drops, rendering T_c fluctuations negligible (0.80-0.95). 2) Statistical significance (R3): Tab. 2 gains are statistically significant (paired t-test, p < 0.05). Though progressive DSC gains appear modest, the modules synergistically drive substantial reductions in geometric errors, effectively correcting vital boundaries. 5.Notation (R3, R4) λ_h in Eq. 5 scales entropy-based modulation in UW. We will explicitly define this. 6.Circularity (R2) To prevent circularity from shared biases in ambiguous regions: 1) EMA temporally decouples the teacher. 2) FD structurally enforces feature divergence, reaching consensus via distinct pathways. 3) UW down-weights supervision in high-entropy regions, avoiding false agreement memorization. 7.Reproducibility (R4) We commit to open-sourcing the full codebase upon acceptance.
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.
The reviewers consistently recognize the paper’s strengths: a timely and important problem (mixed-domain semi-supervised fundus segmentation), a well-motivated integration of MedSAM and U-Net, and solid ablation studies. While Reviewer #3 gives a Weak Reject, the concerns raised — lack of clarity on differences from prior work (SynFoC), weaker performance on two domains, missing statistical testing, and insufficient hyperparameter sensitivity analysis — are all addressable during rebuttal. Importantly, Reviewer #3 explicitly states that the score may be raised if concerns are adequately addressed. No reviewer identifies a fatal methodological flaw. Therefore, the paper should proceed to the rebuttal phase, with the expectation that the authors provide a thorough response addressing the noted concerns and strengthening the experimental analysis.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper does not convincingly distinguish itself from the prior SynFoC framework. The rebuttal’s claim that SynFoC uses only “image level” confidence is inaccurate (SynFoC operates at the pixel level), so the asserted novelty of “spatial” reliability is unfounded. Moreover, the method performs substantially worse than existing SOTA on two of the four test domains, and the rebuttal’s explanation (that the model deliberately suppresses adaptation in difficult cases) actually confirms a lack of robustness. Reviewer #4 also notes that key components largely reuse existing ideas. These core issues cannot be fixed during rebuttal.
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.
After the rebuttal, all reviewers agreed that the paper should be accepted. The Area Chair supports this consensus and recommends acceptance.
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.
Two accept and one missing after rebuttal.
I agree and would accept the paper.