List of Papers Browse by Subject Areas Author List
Abstract
Accurate corneal endothelial cell (CEC) quantification is essential for the diagnosis and management of ocular diseases; however, current learning-based CECs segmentation is often hindered by a significant domain gap between homogeneous training sets and heterogeneous clinical data. Existing models frequently fail to generalize to images from unseen devices, centers, pathologies or species. We present ProtoCEC, a prototype-driven framework designed for robust domain generalization in confocal microscopy image segmentation. The architecture comprises a prototype collector for robust feature extraction and a prototype enhancer to improve class discriminability. At the inference stage, a prototype allocator facilitates dynamic feature-to-prototype matching to guide the segmentation mask generation. Our results confirm that ProtoCEC provides a significant performance boost over existing state-of-the-art architectures. We demonstrate the model’s versatility through rigorous testing in four out-of-distribution scenarios: cross-center, cross-disease, cross-device, and cross-species, proving its efficacy as a generalized solution for confocal microscopy image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0275_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{LiHon_PrototypeDriven_MICCAI2026,
author = { Li, Hongshuo AND Li, Wenlong AND Zhou, Shijia AND Zheng, Yalin AND Qi, Hong AND Zhao, Yitian},
title = { { Prototype-Driven Robust Domain Generalization in Corneal Endothelial Cell Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses domain generalization in corneal endothelial cell segmentation, a task that is particularly challenging due to variability across imaging conditions, devices, and datasets. The method is based on superpixel-guided clustering (SGC), a prototype-based approach originally proposed for few-shot learning, which is extended by introducing two types of prototypes: (i) cell prototypes, extracted from correctly segmented regions during training, and (ii) error prototypes, extracted from regions where the model fails during training. The underlying assumption is that features corresponding to correctly segmented regions capture more stable and transferable characteristics, while features associated with errors are more likely to reflect ambiguous or domain-specific patterns. By separating and treating these two types of features differently, the method aims to improve robustness when applied to unseen domains.
- 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.
S1.The paper addresses corneal endothelial cell segmentation, which is clinically relevant for diagnosis and monitoring. Moreover, it tackles the problem of domain generalization, which is pervasive in medical image computing, where datasets are often limited and models must remain robust to variations across centers, devices, and pathologies.
S2.The proposed method demonstrates strong empirical performance across multiple domain shifts, including cross-center, cross-disease, cross-device, and cross-species scenarios. This is a comprehensive and meaningful evaluation of domain generalization.
S3.The paper proposes a novel and non-trivial extension of (SGC) by introducing error prototypes to model difficult regions. The overall framework is modular and could, in principle, be adapted to other segmentation tasks or integrated into different architectures.
S4.The method appears reasonably robust to hyper-parameters, as supported by the sensitive analysis. This supports the practical applicability of the method.
S5.The use of an error-focused segmentation loss can be interpreted as a form of spatial hard-example mining, encouraging the model to focus on regions where previous predictions were incorrect. This is a simple and intuitive mechanism to improve performance on challenging areas.
- 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.
W1.I find the connection between the proposed method and its intended goal of domain generalization to be somewhat unclear. While the use of prototype-based approaches (such as SGC) is well motivated for few-shot learning, it is less clear how extending this framework with cell and error prototypes should improve robustness to domain shifts. The authors argue that prototypes guide segmentation through similarity matching and thereby “decouple feature recognition from domain-specific noise,” but it is not entirely clear how this decoupling is actually achieved in practice.
In particular, the method does not appear to explicitly ensure that the cell prototypes capture domain-invariant features. Similarly, error prototypes, defined based on regions where the model fails during training, may not necessarily correspond to domain-dependent features. As a result, it remains unclear why cell prototypes should transfer across domains, or why error prototypes would consistently identify unreliable features in unseen settings, as the method does not appear to include explicit mechanisms that encourage this behavior.
Overall, while the intuition behind separating “reliable” and “unreliable” regions is reasonable, the link between this design and improved domain generalization remains largely implicit and is not convincingly supported by the methodology. This raises the question of whether the proposed approach should be considered a domain generalization method, or rather a form of hard-example mining (thanks to the error-focused segmentation loss and the error prototypes) that incidentally improves performance on the CEC segmentation task.
W2.Related to the previous point, a more detailed ablation study would be helpful to better understand what drives the observed performance gains, and in particular whether they stem from the incorporation of hard-example mining strategies. The current ablation appears relatively coarse, as multiple components are introduced simultaneously (e. g. , the prototype enhancer, ablated as a whole, combines contrastive learning and prototype refinement, while the prototype allocator introduces confidence maps and noise mitigation). This makes it difficult to attribute performance gains to specific design choices.
More importantly, the contributions of the different loss terms are not fully disentangled. The optimized objective combines a standard segmentation loss, an error-focused (hard-mining) segmentation loss, and a contrastive loss, but their individual impact is not explicitly evaluated in isolation. In particular, it would be informative to assess how much of the improvement is due to the hard-mining component, and whether applying a similar strategy to baseline methods would yield comparable gains.
W3.The experimental evaluation focuses primarily on cross-domain scenarios, but does not explicitly report performance on the training domain. It would be useful to assess how the proposed method performs in-domain, in order to better understand the trade-off between improving domain generalization and maintaining performance on the source distribution. Also, reporting in-domain performance would help better understand whether the observed improvements are specific to domain generalization, or instead reflect more general effects of the training strategy, such as the use of error-focused supervision.
W4.The t-SNE visualizations provide a qualitative illustration of how the proposed method separates cell and error features, but provide limited evidence to support the domain generalization claims. First, since the method is explicitly designed to distinguish between cell and error prototypes, the observed separation between them may be expected and may not necessarily demonstrate improved generalization. Second, the plots do not provide insight into how features from different domains are distributed. In particular, samples are not stratified by domain, which makes it difficult to assess whether the proposed method effectively reduces the domain gap or aligns representations across domains. Using domain-specific color coding could provide better evidence of improved generalization.
W5.[Minor] Section 2.3 indicates that the pool of prototypes is N=128.In Section 3.1, it is set to N=3.This could be clarified. Also, it would be helpful to specify whether N=3 refers to the total number of prototypes or to the number of prototypes per type (cell and error).
W6.[Minor] The description of the prototype update mechanism in Section 2.3 is not completely clear. It would be helpful to clarify whether the update rule is applied in the same way to both cell and error prototypes. Additionally, it is not specified whether cell and error prototypes are treated as negative pairs in the contrastive learning loss, even when they are considered similar according to the similarity function.
W7.[Minor] The method relies on an iterative procedure in which the predicted segmentation from the previous epoch is used to estimate error regions for the next epoch. It is not entirely clear how this process is initialized, particularly during the first epoch when no prior prediction is available.
- 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
The paper presents promising empirical results on a relevant and challenging problem. However, the mechanism by which the proposed formulation contributes to domain generalization remains insufficiently characterized. It is unclear whether the observed improvements arise from the prototype-based modeling itself or are primarily driven by the error-focused (hard-example mining) components of the training objective. A more targeted ablation isolating these factors (for example, evaluating the impact of the error-focused loss independently of the prototype mechanism) would help clarify the source of the gains.
It would also be valuable to provide more direct evidence that the method learns domain-invariant representations. For instance, by comparing prototype assignment statistics and similarity distributions across domains, we could assess whether the learned features are truly domain-invariant rather than domain-specific. Visualizations such as t-SNE stratified by domain could further support this analysis.
- 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 presents strong empirical results on a clinically relevant and challenging problem, and proposes an interesting extension of prototype-based segmentation. However, I find that the central claim of domain generalization is not fully supported by the method design and the current analysis. In particular, the method does not explicitly enforce domain invariance, and it remains unclear whether the observed improvements stem from the proposed prototype formulation or from the error-focused training strategy.
As a result, while the method is technically sound and promising, the lack of clear evidence linking the design choices to domain generalization limits the strength of the contribution. I would therefore place the paper slightly below the acceptance threshold, but remain open to revising my assessment based on 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.
I appreciate authors’ effort to clarify the motivation and positioning of the method. I still believe the paper should moderate some of its domain generalization claims. The rebuttal largely reiterates that the framework “explicitly models and mitigates domain gaps,” but the methodological justification for this claim remains unclear. In particular, I would like to highlight that the strong improvement reported on the source domain in the rebuttal is further evidence suggesting that the observed gains are likely driven in substantial part by hard-example mining and improved representation learning. I therefore continue to view ProtoCEC primarily as a robust prototype-regularized segmentation framework that empirically generalizes well, rather than as a method providing deeper insight into domain generalization itself.
Nevertheless, the empirical evaluation presented in the manuscript itself remains extensive, the application is clinically relevant, and the observed robustness across cross-center, cross-disease, cross-device, and cross-species scenarios constitutes a meaningful practical contribution. Although I still find the theoretical claims relatively limited, I believe the practical contribution and experimental validation are sufficient for acceptance at a primarily application-oriented conference such as MICCAI.
Review #2
- Please describe the contribution of the paper
This paper presents ProtoCEC, a prototype-driven framework for domain-generalizable corneal endothelial cell (CEC) segmentation in confocal microscopy images. The core idea is to explicitly extract and maintain two prototype pools, one for valid cell features and one for segmentation error patterns, and use these pools to clean up skip connections in a U-Net architecture before they reach the decoder. The framework comprises three components: a Prototype Collector (difficulty-aware feature mining), a Prototype Enhancer (contrastive learning to separate prototypes), and a Prototype Allocator (bidirectional feature rectification via standardization and suppression). The method is evaluated across four out-of-distribution scenarios: cross-center, cross-disease, cross-device, and cross-species.
- 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.Clinically motivated The domain generalization problem in CEC segmentation is real, important, and underexplored. The paper does an excellent job motivating why existing methods fail, they overfit to single-center, single-device training distributions, and Fig. 1 powerfully illustrates these failure modes with concrete Dice drop numbers. 2.Novel idea The prototype strategy is clever. Rather than only learning what cells look like (as standard prototypical networks do), ProtoCEC explicitly learns what segmentation failures look like via the Error Prototype Pool. This is an elegant and well-motivated design choice that distinguishes the work from prior art. 3.Comprehensive evaluation Testing across four distinct distribution shift scenarios; cross-center, cross-disease, cross-device, and cross-species; is a robust way to test different cases. 4.Consistent performance improvements The reported gains are consistent and meaningful: +3.2%, +4.5%, +7.7%, and +3.6% over the second-best method across the four scenarios. These are not marginal improvements.
- 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.Prototype collector There is some ambiguity in the paper regarding the Cell Prototype Pool and Error Prototype Pool, is it “derived from the predicted mask P” and the Error Prototype Pool is “extracted from the error mask M_ER”. However, it is never explicitly explained how the superpixel-guided clustering (SGC) extracts prototypes from these two sources in practice: Does SGC operate on the feature map masked by P, or on the full feature map with P as a soft weight? For the error branch, does the shared encoder process I_ER as a completely separate forward pass, or are features reused from the main forward pass? If it is a separate forward pass through the shared encoder, what is the computational cost during training? At epoch 1, there is no previous prediction — how is the cold start handled? The paper is completely silent on this 2.The prototype allocator’s standardization step Is very aggressive When scell > serror, the paper completely replaces the pixel feature f with its nearest cell prototype: f’ = arg max sim(f, p) p ∈ Pcell This is a hard replacement, the original feature is discarded entirely. This raises serious concerns: Fine-grained spatial details present in f but not captured by any prototype are permanently lost With only 128 prototypes per pool, can the library truly represent all valid cell morphologies across cross-species scenarios? No ablation compares hard replacement vs soft blending (e.g., f’ = α × nearest_proto + (1-α) × f) I would suggest test the algorithm with soft blending if possible 3.The bottleneck features are never guided The PA module only operates on skip connections at Block 2, 3, and 4.The deepest bottleneck features flow directly to the decoder without any prototype guidance: Encoder bottleneck → Decoder (unguided, potentially domain-noisy) Block 2,3,4 skip → PA → Decoder (guided and cleaned) The bottleneck arguably contains the most semantically rich and domain-sensitive features. Is there any reason that the authors decided to avoid that in the bottleneck layer, that’s need to be added in the Ablation studies section 4.Hyperparameter justifications are weak Several key hyperparameters are fixed without adequate justification: τ (temperature) 0.1 No sensitivity analysis provided η (momentum) 0.9 No sensitivity analysis Dual thresholds 0.3 / 0.8 No ablation on these critical values K per level 128 for all same capacity for all encoder levels seems suboptimal
5.SIM needs more Justification The Structural Importance Mapping function is directly borrowed from a reference. The paper doesn’t justify why SIM specifically is the right choice over simpler alternatives such as Gaussian blur of the error mask No ablation compares SIM against these simpler alternatives, making it unclear whether the SDF-based exponential formulation is truly necessary. 6.Dataset sizes are concerning for generalization claims The external test sets are quite small: Cross-device (Rodrep): 52 images ← very small! Cross-species (Alisarine): 30 images ← extremely small! With only 30 images for the cross-species evaluation, the +3.6% improvement over the second-best method may not be statistically reliable. No statistical significance testing (e.g., Wilcoxon signed-rank test) is reported for any of the comparisons. 7.Lack of related work The paper claims ProtoCEC is “the first prototype-driven framework tailored for corneal endothelial cell segmentation”, but this claim is not carefully scoped. Prototype learning has been extensively applied in medical image segmentation more broadly. The authors should either: More carefully scope this claim to CEC segmentation specifically, or provide a proper related work section discussing prototype-based methods in medical imaging. Notably, there is no related work section at all. Minor Comments What does robust mean in the title ?, I recommend to remove that word Fig. 2 is very dense and difficult to read and understand The conclusion is very brief and adds little beyond what is already in the abstract
- 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 an important clinical problem, which is prototype-oriented, built on top of the U-Net architecture. There are some significant gaps with the technical details, and also the size of the datasets and the missing statistical analysis that I mentioned earlier, but it is still a nice approach with variations to be explored.
- 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.
After considering the authors’ rebuttal, I am updating my recommendation to accept. The rebuttal addressed several of my main concerns, particularly regarding the role of the prototype allocator, the distinction between hard-example mining and domain-generalization mechanisms, the handling of prototype initialization, and the rationale for design choices such as hard replacement and skip-connection guidance. The additional clarifications on in-domain performance, statistical significance, annotation consistency, inference speed, and comparison with foundation-model baselines also strengthen the paper. Although some details should still be clarified in the final version, I believe the work presents a clinically relevant problem, a novel and thoughtful technical approach, and strong empirical evidence across multiple domain-shift settings. Therefore, I support acceptance.
Review #3
- Please describe the contribution of the paper
The paper introduces ProtoCEC, which is a prototype-driven deep learning framework designed to achieve domain generalization for corneal endothelial cell segmentation. The main contribution is the architectural design that explicitly decouples invariant anatomical features from domain-specific artifacts. This is achieved through three modules: a Prototype Collector that mines both “cell” and “error” features a Prototype Enhancer that refines these features via contrastive learning; and a Prototype Allocator that performs bidirectional feature rectification to standardize cell features and suppress error features during inference. They show outperforming state-of-the-art methods across four scenarios: cross-center,cross-disease, cross-device, and cross-species settings.
- 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 concept of explicitly mining and utilizing “error prototypes” alongside valid cell prototypes is innovative. By actively learning the feature signatures of typical failure modes and using them to suppress domain-specific artifacts via the Prototype Allocator, the authors provide a structurally elegant solution to domain shift.
2) The evaluation design is strong. Testing the framework against four distinct out-of-distribution (OOD) scenarios (cross-center, cross-disease, cross-device, and cross-species) directly addresses the clinical reality of heterogeneous data and shows the model’s versatility.
3) ProtoCEC demonstrates clear, significant margins over a wide array of recent state-of-the-art baselines (+3.2% to +7.7% Dice improvements across OOD sets).
4) The paper provides a convincing ablation study that isolates the contributions of individual modules (PC, PE, PA) and maps the model’s sensitivity to key hyperparameters, demonstrating that the architectural choices are justified and not arbitrary.
5) The method is not limited to the proposed application and can be extended and adapted to other segmentation tasks.
- 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 OOD evaluation is extensive, the paper lacks detail on how the ground truth annotations were obtained for the four external target datasets (DATA-I, DATA-II, Rodrep, Alisarine). It is critical to clarify whether these were annotated by the same clinical experts as the source dataset, as varying annotation standards across datasets could introduce bias into the cross-domain evaluation.
2) The source dataset (CED) and the primary target datasets (DATA-I, DATA-II) appear to be private/in-house. The manuscript does not state whether the code, pre-trained models, or datasets will be made publicly available upon publication. Given the complex nature of the training pipeline, releasing the code is highly recommended.
3) The authors claim the inference is “streamlined for real-time application”, but no metrics (e.g., inference time per image) are provided to substantiate this claim or compare the computational overhead against the baseline U-Net/ResNet models.
4) The authors should clarify if dice score is computed over the boundary overlap or over cell area?
5) How does this pipeline compare against methods that use foundation models in their pipeline, such as Cellpose SAM?
- 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 relevant clinical bottleneck, model fragility under domain shift in corneal endothelial cell segmentation, and its rigorous evaluation across four distinct out-of-distribution scenarios (cross-center, cross-disease, cross-device, cross-species) is a major strength. The proposed ProtoCEC architecture is logically sound, and the ablation studies effectively validate the contribution of each module. However, there are several critical points that must be resolved in a revision. To secure a higher score, the authors must clarify the ground-truth annotation protocols for the external target datasets, explicitly define their Dice score formulation (boundary vs. area), and provide quantitative computational metrics to substantiate their claims of “real-time” inference. Furthermore, discussing the model’s performance relative to modern foundation models (e.g., Cellpose, SAM) and addressing reproducibility are necessary.
- 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 cleared up my major concerns. They have clarified the annotation protocol for the external datasets and confirmed that the code will be released. I also really appreciate getting the metrics for the inference speed (20ms/image) and the clarification on the Dice score (based on boundary overlap). The new baseline comparisons against Cellpose and SAM are also appreciated.
Overall, given the sound methodology and rigorous multi-scenario OOD evaluation, this is a clinically relevant paper that will be worth sharing with the community.
Author Feedback
1.Domain Generalization (R1, R2) While the clinical motivation and evaluation of this work are anchored in CCM, our architectural innovations constitute a targeted domain generalization strategy. Rather than relying on standard hard-example mining that incidentally improves segmentation, our framework explicitly models and mitigates domain gaps across distinct scenarios (cross-center/disease/device/species). However, we agree with the reviewer that our claims should be precisely scoped, and will revise our tone, framing ProtoCEC primarily as a robust framework tailored for CCM.
In practice, this robustness is achieved because the decoder never processes raw, domain-specific features. The Prototype Allocator acts as a structural adapter: it strictly standardizes valid features by mapping them to purified canonical prototypes, while utilizing error prototypes to actively attenuate OOD noise.
2.Cross-Domain Alignment & t-SNE (R1) Beyond 2D visual projections, we rely on quantitative alignment evidence: tracking allocated features from the four unseen domains revealed that 117 of 128 cell prototype clusters embed mixed-domain features (initially omitted for space). This proves cross-domain alignment. Furthermore, separating cell and error prototypes in Fig. 5 is not merely a loss artifact; trapping deceptive OOD patterns drives our robustness.
3.In-Domain Performance (R1) Error-focused supervision provides a general uplift without sacrificing source Dice. On the held-out source test set (omitted for space), ProtoCEC achieved 0.858 Dice, outperforming SOTA (SPG-Net) by 4.4%, proving standardization suppresses noise while enhancing refinement.
4.Hard-Mining vs. Standardization (R1) While hard-example mining strengthens intra-domain boundaries, our framework goes further by projecting unseen domain features onto a stable, discrete manifold. Decoupling loss alone without other components (L_seg: 0.688, L_error: 0.705, L_con: 0.730) confirms standardization overwrites domain-sensitive textures with invariant geometric configurations (e.g., SGC-captured repetitive hexagonal structures). This yields a 2.5% OOD gain over L_error (3.3% when applied to SPG-Net), demonstrating that it drives optimal OOD generalization across all shift scenarios, rather than merely performing hard mining.
5.Module Designs, Aggressiveness & Scope (R1, R2) SGC operates strictly on hard-masked regions (binarized P and M_ER) without soft weights. Processing the error image via a separate pass doubles training computation but adds zero inference overhead. A 30-epoch warm-up reliably prepares robust initial prototypes (apologies, omitted). To filter domain interference, hard replacement overwrites noise for optimal OOD generalization (0.765 Avg. Dice) vs. soft blending at α=0.5 (0.742). Fine details are preserved via unguided bottleneck/decoder pathways. Guiding the deepest bottleneck causes over-regularization (0.749) vs. skip connections only (0.765). Parameters (τ=0.1, η=0.9, thresholds 0.3/0.8, pool capacity K=128, image-specific per-class prototypes N=3) were optimized via grid search. SIM uses an SDF-based exponential formulation, assigning unit weights inside and smoothly decaying outside to enhance boundary contrast. Wilcoxon tests confirm significant gains across all OOD sets (p<0.001). We will revise the “first” claim. Sub-component analysis: PE (contrastive only 0.730, +update 0.735) and PA (suppression only 0.738, +confidence 0.748).
6.Annotation, Speed & Foundation Models (R3) External annotations (DATA-I/II) share the exact CED source pipeline, cross-verified by three experts. Rodrep and Alisarine are public. Code will be open-sourced; ProtoCEC achieves real-time inference (20ms/image, RTX 3090), faster than SPG-Net (25ms). Dice is computed over boundary overlap, which is critical for CEC morphometry. ProtoCEC (0.765) outperforms Cellpose (0.732) and SAM (0.678) on OOD sets, as generic models lack structural priors for low-contrast CECs.
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 work proposes a prototype-driven domain generalization method for corneal endothelial cell segmentation. The recommendation is mixed among reviewers. The main disagreements are on empirical evaluation, methodology effectiveness, lack of sufficient evidences and explanations for some claims, and comparison to related works. Based on my reading, I think that the reviewers raised effective concerns. Please try to fix these concerns in the rebuttal.
- 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 three reviewers agree to accept. I read the rebuttal and reviewers’ comments after the rebuttal. I think that the rebuttal solves main concerns, and found not inappropriate comments from the reviewers.
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.
This work deals with domain generalization in corneal endothelial cell segmentation, which is driven in a prototype-learning manner. All the reviewers agreed on the importance of this work, but raised several clarifications during the first round. The authors rebuttal seemed to be quite convincing for the reviewers and all of them have unanimously voted for acceptance.
I think this paper is a good methodology as well as application driven work and will encourage good discussions amongst the community.
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.
All reviewers have reached a consensus in favor of accepting the paper.