List of Papers Browse by Subject Areas Author List
Abstract
Volumetric segmentation of Optical Coherence Tomography (OCT) images is essential for diagnosing ocular diseases but requires labor-intensive voxel-wise annotations. While semi-supervised learning (SSL) can reduce annotation costs, most existing methods process data slice-by-slice, failing to exploit the inherent 3D spatial context. In this paper, we propose PC-Seg (Progressive Cross-view Segmentation), a novel curriculum learning framework that lifts sparse 2D annotations to high-precision 3D segmentation models. Unlike conventional multi-view approaches, our method employs a single 2D model to learn cross-view consistency from both standard B-scans and orthogonal slices, generating reliable volumetric pseudo-labels. These labels are then distilled into a 3D model, followed by a co-training phase where 2D and 3D models mutually refine each other through ensemble pseudo-labeling. Experiments on the MSHC dataset and the Duke DME dataset demonstrate that our method achieves accuracy comparable to fully supervised learning using only about 0.7\% of the labeled data, outperforming both state-of-the-art semi-supervised and retinal layer segmentation methods. Our code is publicly available at https://github.com/gsisaoki/pc-seg-official.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6555_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/gsisaoki/pc-seg-official
Link to the Dataset(s)
MSHC dataset: https://iacl.ece.jhu.edu/index.php/Resources
Duke DME dataset: https://people.duke.edu/~sf59/Chiu_BOE_2014_dataset.htm
BibTex
@InProceedings{KonTsu_PCSeg_MICCAI2026,
author = { Konno, Tsubasa AND Ninomiya, Takahiro AND Zhou, Yukun AND Ito, Koichi AND Wagner, Siegfried K. AND Lin, Yiqun AND Keane, Pearse A. AND Nakazawa, Toru AND Aoki, Takafumi},
title = { { PC-Seg: Progressive Cross-View Consistency for 3D OCT Segmentation from Sparse 2D Annotations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes PC-Seg, a progressive cross-view segmentation framework for training 3D OCT segmentation models from sparse 2D annotations. The method adopts a multi-stage curriculum learning strategy: a single 2D model is first trained to enforce cross-view consistency between standard B-scans and orthogonal slices, which is then used to generate volumetric pseudo-labels. These pseudo-labels are distilled into a 3D model, followed by a co-training stage in which the 2D and 3D models further refine each other through ensemble pseudo-labeling. Experiments on the MSHC and Duke DME datasets suggest that the method can achieve performance comparable to fully supervised learning while using only a very small fraction of labeled data.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a practically important problem: how to train a 3D OCT segmentation model under sparse 2D annotation conditions. Rather than directly relying on dense voxel-wise supervision, the method uses cross-view consistency to progressively lift limited 2D supervision into 3D training signals, which is a meaningful direction for annotation-efficient OCT analysis. 2.A notable strength is that the method uses a single 2D model to exploit information from both standard B-scans and orthogonal slices, instead of introducing multiple view-specific models. This design is conceptually neat and provides a feasible route from sparse 2D supervision to 3D modeling. 3.The method makes good use of unlabeled data. By leveraging consistency across views, it generates volumetric pseudo-labels and thereby increases the utility of unlabeled orthogonal slices, reducing dependence on costly 3D annotation. 4.The paper is also well aligned with practical clinical needs. Training a 3D model from only sparse 2D slice annotations is attractive in settings where voxel-level annotation is prohibitively expensive, and the reported results suggest that the proposed framework can support 3D analysis under highly limited supervision. 5.The experimental design is reasonably comprehensive. The paper evaluates the method on two public datasets, includes comparisons under different annotation ratios, and benchmarks against both fully supervised and semi-supervised methods. This makes the empirical evaluation relevant to the low-annotation setting that motivates the work. 6.The progressive multi-stage training strategy is another appealing aspect of the paper. The staged transition from 2D learning to 3D modeling appears to stabilize training and to reduce error amplification from early pseudo-labels, which is particularly important in sparse-supervision settings.
- 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 overall idea is interesting, but the methodological novelty appears to be more in the integration of existing ideas than in a fundamentally new modeling framework. Cross-view consistency, pseudo-labeling, co-training, and curriculum learning have all been explored in medical image segmentation. The main contribution here lies in how these components are combined for sparse-annotation 3D OCT segmentation. 2.The rationale for the five-stage curriculum design is not sufficiently explained. The paper does not clearly justify why five stages were chosen, nor does it provide enough analysis of whether this design is necessary or optimal. Additional details on stage duration, switching criteria, and model selection during pseudo-label generation would improve both clarity and reproducibility. 3.Some implementation details remain unclear. For example, the paper could more clearly explain how orthogonal slices are aligned within the 3D volume, how pseudo-labels are transferred between the 2D and 3D models, how the feedback loop between the two models is implemented, and how the confidence thresholds are applied in practice. These omissions make the method harder to reproduce. 4.The motivation in the introduction for why existing multi-view methods developed for CT or MRI may not transfer well to OCT is currently not fully convincing. In particular, the discussion would benefit from a clearer explanation of anisotropy and why it changes the applicability of prior approaches. 5.The framework figure is informative but visually dense, which makes the overall pipeline somewhat difficult to follow at first reading. 6.The paper would benefit from additional analysis of confidence threshold selection for pseudo-label generation, since these thresholds are likely to influence the stability and final performance of the multi-stage training pipeline. 7.Some parts of the experimental presentation could be improved. For example, Table 2 does not report statistical significance indicators and does not explain missing entries clearly. In Table 3, the distinction between fully supervised and semi-supervised methods, as well as the annotation ratios used by each method, should be made more explicit to ensure fair and reproducible comparison. 8.Figure 2 mainly presents comparisons with relatively weaker segmentation results and does not include stronger or more representative baselines. Including comparisons against stronger-performing methods would make the visual evidence more convincing. 9.The full pipeline is relatively complex, involving multiple training stages, repeated pseudo-label generation, and alternating 2D/3D optimization. While this design may improve stability, it also increases implementation burden and hyperparameter sensitivity compared with simpler end-to-end approaches. 10.The conclusion section would be stronger if it more explicitly discussed current limitations and future directions, rather than largely restating the paper’s main findings.
- 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 paper addresses an important and practical problem: how to perform 3D OCT segmentation under sparse annotation conditions. Its main strength lies in using cross-view consistency to progressively lift sparse 2D supervision into volumetric pseudo-labels, and then transferring this supervision into a 3D model through distillation and co-training. This provides a meaningful solution to the annotation bottleneck in OCT segmentation and appears particularly well suited to settings with strong anisotropy and limited labeling resources. Although the overall framework is composed of several known ingredients, I think the way they are combined is thoughtful and practically useful for this task. The empirical results are also encouraging, especially given the very low annotation ratio used in the experiments. My main reservations concern limited methodological novelty, insufficient detail on the curriculum and pseudo-labeling mechanisms, and some aspects of the experimental presentation and reproducibility. Still, I find the problem important, the approach technically reasonable, and the results strong enough to place the paper slightly above the acceptance threshold.
- 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 rebuttal clarifies several important aspects of the work, particularly regarding the OCT-specific motivation, anisotropy, staged training rationale, and implementation details. The authors also provided additional information on pseudo-label generation, confidence thresholds, orthogonal alignment, and statistical significance analysis, which improves the overall clarity and reproducibility of the method. The rebuttal further clarifies why the proposed staged framework is particularly suitable for anisotropic OCT data, and the overall methodological design appears technically sound and practically meaningful for sparse-annotation OCT segmentation. The experimental results also provide reasonable empirical support for the effectiveness of the proposed framework. Although some aspects of the experimental presentation could still be improved, I believe the paper provides sufficient empirical validation and domain-specific contribution to merit acceptance at MICCAI.
Review #2
- Please describe the contribution of the paper
The paper proposes a five-stage curriculum learning framework, named PC-Seg, for semi-supervised 3D OCT segmentation using sparse 2D annotations. The method progressively propagates knowledge from 2D annotated B-scans to 3D volumes through 2D warm-up, orthogonal slice adaptation, orthogonal cross-teaching, knowledge distillation to a 3D model, and final 2D/3D co-training. Experimental results on the MSHC and Duke DME datasets demonstrate that the proposed method achieves competitive performance with extremely limited annotated data compared to recent state-of-the-art semi-supervised methods.
- 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 motivation to bridge the gap between 2D slice-wise annotations and 3D spatial context in OCT while alleviating the overhead of training 3D models from scratch is clear and practically valuable.
2.The writing is straightforward, and the pipeline logic is well-illustrated.
- 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 core concern is the novelty of the method. According to my understanding, PC-seg is a general pipeline which is not only suitable for OCT, but also for other volumetric images, like CT, MRI. If you have searched deep enough in sparsely annotated semi-supervised segmentation/barely-supervised medical image segmentation, you will find the core idea of PC-seg is not original or novel. Especially for SGTC (AAAI 2025) [1] and DeSCO (CVPR 2023) [2], which also exploit orthogonal 2D sparse annotations to generate pseudo labels and then co-train 3D models (semi-supervised pipeline). But the authors didn’t include these up-to-date methods into their baseline for comparison. Since from the methodology perspective, PC-Seg is a general method, why do the authors only conduct experiments on OCT, not MRI or CT? To sum up, the authors should extend their proposed PC-Seg into different datasets from other modalities like CT, MRI and include more relevant baseline methods for comparison. Why not directly use well-pretrained model to generate pseudo-label but choose warm-up training? More up-to-date semi-supervised methods for 3D medical image segmentation should also be included.
2.The information in Fig.1 is so insufficient. When cross-referencing with the methodology section, it becomes apparent that many critical details are not visually represented, such as the predictions under four perturbations (weak, feature-level, and two strong) and the cross-teaching strategy. Consequently, the core innovation is obscured, instead of a clear highlight, the figure presents a cumbersome five-stage training strategy.
3.Lack of significance test: In highly sparse annotation settings (e.g., 6 or 30 labels), deep learning models are notoriously sensitive to the random sampling of labeled data. The paper exclusively reports single-run mean values in Tables 1, 2, and 3 without variance/standard deviation or statistical significance tests (e.g., t-test). The marginal improvements in some settings (e.g., 0.9135 to 0.9174 for 60 labels) are mathematically unconvincing without robust statistical validation.
[1] SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation [2] Orthogonal Annotation Benefits Barely-supervised Medical Image Segmentation
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please refer to the major weaknesses.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors didn’t answer my questions: (1) Why not directly use well-pretrained model to generate pseudo-label but choose warm-up training? (2) More up-to-date semi-supervised methods for 3D medical image segmentation should also be included. Moreover, as the most important figure of the whole paper, the information of Figure 1 is too insufficient. And to be honest, the performance improvement of the proposed method PC-Seg is marginal. So I still hold on my original opinon.
Review #3
- Please describe the contribution of the paper
This paper presents weakly supervised method for volumetric segmentation of optical coherence tomography (OCT) images named PC-Seg (Progressive Cross-view Segmentation). PC-Seg lifts sparse 2D annotations to high-precision 3D segmentation by a five-stage curriculum learning strategy. The method starts with training 2D Residual U-Net (Res UNet) model on sparsely annotated B-scans. The Res UNet is also a baseline model for 3D version. Predicted segmentation masks are also retained as pseudo-labels for Orthogonal B-scans in Stage 2.In Stage 2 two 2D Res UNet models are trained: one for B-scans and one for Orthogonal B-scans (with pseudo-labels from Stage 1). Ortho B-scans with weights with the highest accuracy are used to generate B-scan pseudo-labels for Stage 3.At Stage 3 cross-teaching strategy between orthogonal slices is used such that pseudo-labels from each plane are mutually generated from predictions on its orthogonal counterpart. At the end the ensemble predictions from both planes are used to generate high-quality 3D pseudo-labels. In Stage 4 semi-supervised learning of 3D model is conducted on sparsely labeled and unlabeled volumes guided by the 3D pseudo-labels from Stage 3.In Stage 5, using the highly accurate pseudo-labels from Stage 4, 2D and 3D models are trained alternatively updating dynamically ensemble pseudo-labels. Proposed model is validated on two public OCT datasets: the OCT MS and Healthy Control (MSHC) dataset and the Duke DME dataset. By using F1 score as a metric proposed method was compared with: five domain specific supervised methods, four domain specific unsupervised methods, and four general semi-supervised frameworks. The ensemble version of proposed method outperformed all the competitors. Specifically, experiments on the MSHC dataset demonstrated that PC-Seg achieves accuracy comparable to fully supervised methods using only 0.7% of the labeled data. Furthermore, on the Duke DME dataset, PC-Seg outperformed state-of-the-art semi-supervised methods, particularly in detecting complex fluid lesions.
- 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 proposes a novel curriculum learning framework, named PC-Seg (Progressive Cross-view Segmentation) for volumetric 3D segmentation of optical coherence tomography (OCT) images. As opposed to existing semi-supervised learning methods that process data slice-by-slice, PC-Seg exploits the inherent 3D spatial context of OCT imaging data. Novel curriculum learning framework implemented in five stages lifts sparse 2D annotations to high precision 3D segmentation models. Unique contribution of PC-Seg is usage of single 2D model to learn cross-view consistency from B-scans and orthogonal slices to generate reliable volumetric (3D) pseudo-labels. These labels are then distilled into a 3D model, followed by a co-training phase where 2D and 3D models mutually refine each other through ensemble pseudo-labeling. On MSHC and Duke DME datasets, the ensemble version of proposed method outperformed five domain specific supervised methods, four domain specific unsupervised methods, and four general semi-supervised frameworks. It is significant that experiments on the MSHC dataset demonstrated that PC-Seg achieves accuracy comparable to fully supervised methods using only 0.7% of the labeled data. By my judgement proposed PC-Seg method is unique approach in attempt to handle OCT anisotropy and data scarcity. Another strength of proposed PC-Seg method is that it is model agnostic, i.e., the Res Unet model is used in the paper but five-stage curriculum learning strategy is universal.
- 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 studied paper carefully. I do not see any significant weakness of proposed PC-Seg method. By my judgement proposed PC-Seg method is unique approach in attempt to handle OCT anisotropy and data scarcity. The only objection is that authors neither provided link to the source code of PC-Seg nor mentioned they will provide it if the paper is accepted.
- 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
My only objection to the authors is that they neither provided link to the source code of PC-Seg nor mentioned they will provide it if the paper is accepted.
- 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 presents weakly supervised method for volumetric segmentation of optical coherence tomography (OCT) images named PC-Seg (Progressive Cross-view Segmentation). PC-Seg lifts sparse 2D annotations to high-precision 3D segmentation by a five-stage curriculum learning strategy. On MSHC and Duke DME datasets, the ensemble version of proposed method outperformed five domain specific supervised methods, four domain specific unsupervised methods, and four general semi-supervised frameworks. It is significant that experiments on the MSHC dataset demonstrated that PC-Seg achieves accuracy comparable to fully supervised methods using only 0.7% of the labeled data. By my judgement proposed PC-Seg method is unique approach in attempt to handle OCT anisotropy and data scarcity. Another strength of proposed PC-Seg method is that it is model agnostic, i.e., the Res Unet model is used in the paper but five-stage curriculum learning strategy is universal. I propose this paper for acceptance. The only reason I did not propose strong acceptance is because authors neither provided link to the source code of PC-Seg nor mentioned they will provide it if the paper is accepted. My confidence is above 75%.
- 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.
My initial recommendation for this paper was accept with a confidence over 75%. he only reason I did not propose strong acceptance is because authors neither provided link to the source code of PC-Seg nor mentioned they will provide it if the paper is accepted. However, in their rebuttal authors claimed they will provide source code with a trained weights. Thus, I propose this paper for acceptance with a high confidence. My decision is additionally supported by authors’ responses to comments raised by other two reviewers.
Author Feedback
We thank the AC and Reviewers for their constructive feedback. We are especially encouraged by Reviewer 3’s strong endorsement. Upon acceptance, we will release our complete source code and pretrained weights.
1.Novelty: Domain-Specific Paradigm vs. General Sparse 3D SSL (Rev 1, 2) Reviewer 2’s comparison to general sparse-annotation 3D SSL methods (e.g., SGTC) overlooks a fundamental physical discrepancy.
- Single-Planar Annotation & The Anisotropy Trap: Existing sparse-annotation 3D SSL methods require multi-planar annotations. However, in OCT, orthogonal (Ortho) B-scans exhibit severe sparsity and low inter-slice continuity, making reliable manual annotation impractical. With only sparse single-plane 2D labels, directly applying 3D models to such anisotropic data amplifies kernel biases, causing early training collapse. PC-Seg explicitly evades this “anisotropy trap.” Its core innovation is a staged dimension-lifting curriculum that requires labels from only a single plane (standard B-scan), robustly bootstrapping 3D representations without any Ortho labels.
- Model-Agnosticism & Efficiency: PC-Seg is backbone-agnostic. As stated in Sect. 2.1, our preliminary experiments compared ResUNet with ViT-based architectures (e.g., SegFormer, Swin-UNet). We found that 2D/3D ResUNet achieved comparable or superior performance with significantly fewer parameters. Thus, we prioritized ResUNet for computational efficiency in clinical deployment, though our paradigm also yields consistent improvements with Transformer backbones.
2.Reproducibility & 5-Stage Pipeline Justification (Rev 1, AC)
- Logical Necessity: As evidenced by our ablation study (Table 1), directly co-training multiple views from the initial stage causes unreliable pseudo-labels to propagate, degrading performance. PC-Seg’s 5-stage curriculum prevents this noise accumulation and stabilizes training by progressing through 2D B-scan warm-up, Ortho adaptation, cross-teaching, 3D distillation, and mutual refinement.
- Switching Criteria: For fair comparison, we utilized fixed epochs for each stage (e.g., [10, 5, 5, 50, 50] for MSHC), determined through preliminary experiments, which will be specified in the revision. Additionally, stage transitions can be dynamically triggered by validation loss convergence (patience=5), significantly optimizing total training time.
- Implementation: 3D pseudo-labels are formed by stacking 2D probability maps. Orthogonal alignment requires only simple coordinate permutation. We evaluated thresholds $\tau \in {0.90, 0.95, 0.99}$ and found $\tau=0.95$ optimal.
3.Statistical Significance and Robustness (Rev 1, 2) To validate reliability, we report variance analysis across 7 independent trials using different random seeds for the most practical setting (MSHC 60 labels). PC-Seg consistently achieved a Dice score of 0.9175 \pm 0.0012, outperforming the strongest baseline (2D ResUNet w/ UM: 0.9123 \pm 0.0029). The Wilcoxon signed-rank test confirmed these improvements are statistically significant ($p = 0.0156$). We will update Tables 1-3 with these standard deviations.
4.Figure 1 Clarity (Rev 2) To comply with page limits, Fig. 1 emphasizes the macro-level lifting workflow. In the revision, we will significantly expand the caption to explicitly map the abstracted “Train” arrows to the underlying SSL operations (e.g., four perturbations and cross-teaching).
PC-Seg provides a robust, domain-specialized solution using only single-planar supervision. We hope this response fully addresses the reviewers’ concerns and assists the Area Chairs in their final assessment.
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.
This work received mixed reviews, with major concerns regarding its technical novelty compared with related works and the heavy design of its five-stage training pipeline. Reproducibility and model details, such as the 2D-to-3D lifting, require further clarification. In addition, the figures should be improved for better clarity. Therefore, the authors are invited to submit a rebuttal addressing these issues.
- 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, most reviewers provided positive comments and acknowledged the contribution of using 2D cues for 3D segmentation. The authors’ response regarding the motivation is convincing, particularly for OCT segmentation. Although concerns remain about the heavy framework, the superior performance demonstrates a meaningful contribution to the relevant community. Therefore, final acceptance is recommended.
Meta-review #2
- 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.
After reviewing the authors’ rebuttal and the reviewers’ final recommendations, I lean on the side of rejection, primarily due to concerns regarding the novelty of the proposed approach. As noted by Reviewer 1, the method appears conceptually close to existing approaches such as SGTC, yet these related works are neither adequately discussed nor included in the experimental comparisons. The AC shares the concern that such methods could potentially be adapted to the problem setting considered in this paper, making it difficult to clearly assess the methodological contribution beyond existing literature. Consequently, the experimental validation does not sufficiently establish the distinctiveness and added value of the proposed approach. Given these concerns, the AC is unable to recommend 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.
After carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper addresses an important and practical problem of 3D OCT segmentation under sparse 2D annotation, and proposes PC-Seg, a progressive cross-view framework that lifts limited 2D supervision into volumetric pseudo-labels through staged cross-view consistency, 2D-to-3D distillation, and 2D/3D co-training. Although the method combines several existing ideas such as cross-view consistency, pseudo-labeling, curriculum learning, and co-training, the integration is well tailored to anisotropic OCT data and is practically meaningful for reducing annotation burden. The rebuttal clarified important concerns regarding the OCT-specific motivation, staged training rationale, pseudo-label generation, orthogonal alignment, confidence thresholds, statistical analysis, and code release. While some concerns remain about novelty relative to recent sparse-annotation methods and the complexity of the multi-stage pipeline, the majority of reviewers support acceptance after rebuttal, and the empirical results appear strong under very limited annotation settings. Therefore, I recommend acceptance, with the expectation that the authors will further clarify the annotation ratios, pipeline details, and limitations in the final version.