List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation and quantitative thickness analysis of retinal layers in optical coherence tomography (OCT) are crucial for early diagnoses of ocular disorders. To address the clinical needs of diagnosing various ocular and systemic diseases, numerous multi-granularity OCT datasets are constructed. While deep learning achieves impressive results in retinal layer segmentation, general training paradigms require separate models for datasets with different annotation granularities. Universal models have been developed to diverse datasets and tasks via advanced techniques such as prompt learning, but they overlook across-granularity information and struggle to generalize to new granularities. In this paper, we propose a universal OCT segmentation model, named UniOCTSeg, which builds its basis upon Hierarchical Prompting Strategy (HPS) and Progressive Consistency Learning (PCL). HPS employs a granularity-merging strategy, where prompts at various granularities are constructed based on the finest-grained prompts. Meanwhile, PCL leverages an Exponential Moving Average teacher model to generate pseudo-supervision signals, guiding the model through easy-to-hard progression to ensure consistency across hierarchical levels. Extensive experiments across eight publicly available OCT datasets involving six distinct granularity levels demonstrate UniOCTSeg’s superior performance compared with state-of-the-art methods, while also illustrating its high flexibility and strong generalizability. Our code and data are available at https://github.com/MICCAI2025-1237/UniOCTSeg
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1237_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Halcyon1010/UniOCTSeg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhoJia_UniOCTSeg_MICCAI2025,
author = { Zhong, Jian and Lin, Li and Miao, Chaoran and Wong, Kenneth K. Y. and Tang, Xiaoying},
title = { { UniOCTSeg: Towards Universal OCT Retinal Layer Segmentation via Hierarchical Prompting and Progressive Consistency Learning } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {637 -- 647}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose UniOCTSeg, a universal retinal layer segmentation framework for OCT images that effectively handles datasets with different annotation granularities. They proposed a strategy, Hierarchical Prompting Strategy (HPS), to construct multi-granularity prompts from fine-grained base prompts, enabling the model to flexibly adapt to both seen and unseen granularities across different datasets. The paper also used Progressive Consistency Learning (PCL) that gradually enforces consistency across tasks from coarse to fine granularity levels, enhancing robustness and segmentation accuracy. Extensive experiments on 8 public OCT datasets with 6 different granularity levels show that UniOCTSeg outperforms existing 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.
-
The paper presents a novel approach for handling OCT datasets with varying annotation granularities. Experiments on 8 public datasets demonstrate that the proposed method consistently outperforms existing state-of-the-art models.
-
The authors introduce Progressive Consistency Learning (PCL), which combines a teacher-student framework with Exponential Moving Average (EMA) and a progressive learning schedule. This design helps stabilize training and promotes consistency across different granularities. Ablation studies validate the contribution of each component within PCL.
-
The authors have released the code anonymously on GitHub, enhancing the reproducibility of their work.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The explanation of the Hierarchical Prompting Strategy (HPS) is unclear, especially in Section 2.2, which took the reviewer a considerable amount of time to fully understand.
- The authors did not clarify the rules for merging granularities. It would be better to explicitly state:
- That the nine layers defined in reference [21] are used as the finest-grained hierarchy;
- That the varying granularities across the eight public datasets can all be expressed using these nine basic layers;
- That only adjacent granularity levels can be merged.
- These assumptions are not obvious to readers without prior knowledge of OCT anatomy and should be explicitly stated in the manuscript. (Please correct if any of these points are inaccurate.)
-
The vector dimensions of the prompts are not clearly explained. The merging mechanism is also ambiguous. In particular, Equation (1) is difficult to interpret—the meaning of the circle with a dot operator is unclear. Readers would expect Equation (1) to define the merge operation M explicitly as a function of the prompts, e.g., something like: 𝑀(𝑃𝑖,𝑃𝑗,… )=Conv(𝐺𝑖(𝑃𝑖),… ) This would help make the meaning of M(⋅) in Equation (2) more understandable.
- Equation (3) is also confusing. While it seems the authors intend to convey that 𝑃𝑀 is composed of three basic layers, the expression is overly complicated and obfuscates a relatively simple idea.
Additional Minor Issues:
-
In Figures 1 and 2, some text is too small, making it difficult to read.
-
In Figure 1(d), although the layout is visually appealing, the axes are not normalized, which can mislead readers.
-
In Figure 3, the second-to-last column should be labeled “Ours” or “UniOCTSeg” for consistency and clarity
- The authors did not clarify the rules for merging granularities. It would be better to explicitly state:
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the proposed method shows promise and is supported by comprehensive experiments, the paper relies heavily on established paradigms such as teacher-student EMA frameworks and prompt-based segmentation. Although the Hierarchical Prompting Strategy (HPS) introduces an interesting idea, its formulation is poorly explained and may not constitute a substantial novelty beyond prior work. As a result, the combination of limited methodological novelty and lack of clarity in technical presentation leads me to recommend a weak reject.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a universal OCT segmentation model with a hierarchical prompt strategy to leverage priors across retinal layers. Extensive experiments demonstrate the effectiveness of the proposed method.
- 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 hierarchical prompt strategy is novel and effective. It intergrate the priors from different retinal layers with diverse prompt types. The experimental design is sufficient and result is promising.
- 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.
Some details are missing and should be addressed. For example, it is unclear which ViT the authors employ—whether it is SAM-based ViT, CLIP-based ViT, or another pre-trained resource.
Additionally, the authors should validate the effectiveness of the number of basic prompts N at each level. Furthermore, providing computational complexity comparisons or analyses is essential for reviewers to evaluate the fairness of the performance comparisons.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Novel algorithm and promising results.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
Based on the paper, the main contribution is the introduction of UniOCTSeg, the first universal segmentation framework specifically for OCT retinal layer analysis. This framework uses a Hierarchical Prompting Strategy (HPS) to handle diverse and even unseen segmentation granularities based on finest-grained prompts, and Progressive Consistency Learning (PCL) to improve segmentation robustness and consistency across these different granularities.
- 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 method introduces the novel combination of two core techniques for OCT retinal layer segmentation. The hierarchical prompting strategy allows the model to represent and adapt to any granularity, including those not seen during training, using a fixed set of basic prompts. The PCL enforces consistency between the different hierarchical/granularity levels
The paper has a particularly strong and comprehensive evaluation for a MICCAI submission.
- 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) Lack of Cross-Validation: The used 7:1:2 results in a very small test set compared to the training/validation set. In some cases (e.g. DukeDME dataset) this results in a test set consisting of only 2 volumes. A cross-validation would have been nice. (But it is out of scope for the revision) 2) Complexity: The framework involves several components. The authors provide little details about the implementation details, the computational complexity or inference time compared to simpler models. 3) Source code: While the authors provide a link for the source, the model present in the repository is empty. Due to the complexity, the source is crucial for reproducibility. 4) Novelty: While the paper claims to be the first universal segmentation framework for OCT retinal layer analysis leveraging specific hierarchical dependencies and consistency, the underlying concepts it builds upon (universal models, prompt learning, consistency learning, teacher-student paradigms) are existing techniques in the broader machine learning and medical imaging fields, which the paper acknowledges by citing relevant works. The core novelty lies in the specific integration and adaptation of these techniques (HPS, PCL) for the multi-granularity OCT segmentation problem, rather than the invention of each base component.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
It has a strong and convincing evaluation. This combination of different components is definitely novel. However, the source code is crucial for reproducability.
- 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 concerns were resolved in the rebuttal. I also don’t think the points raised by the other reviewers are serious issues for a miccai submission (given the page limitation).
Author Feedback
We appreciate the reviewers’ valuable feedback. They recognized the novelty of our work and the strong experimental results, such as “novel and effective” (R2), “adaptable to unseen granularities” (R3), and “strong performance on multi-granularity datasets” (R4), which are quite encouraging. Below, we address the main concerns: Basic Prompt Settings (R2-7.2 & R4-7.1/7.2): We set N=9 based on the definition of the nine anatomical retinal layers (Retinal layer parcellation., Data in Brief 2019), making it a fixed design rather than a tunable hyperparameter. Each prompt vector is of size [1, 64]. The operator in Eq. 1 denotes the Hadamard product, and we will make the merge operation more explicit in the final version. Merging Rules (R4-7.1): We appreciate the reviewer’s insight. The understanding is correct, and we will clarify the merging rules and their correspondence to OCT anatomical priors in the final version. Novelty (R3-7.4 & R4-12): We appreciate the reviewers’ comments. However, we respectfully clarify the following points: 1) We propose the first universal model specifically designed for OCT’s multi-granularity retinal layer segmentation. While building on existing concepts like universal models and prompt learning, we innovatively propose a prompt merging strategy HPS and a granularity-aware network that can accommodate datasets of diverse and even unseen granularities within a unified framework. 2) Although PCL adopts a teacher–student setup for pseudo-supervision, its core idea lies in enforcing cross-granularity consistency and anatomical alignment via output and prompt merging, guided by an easy-to-hard task progression strategy. This integrated design enhances training stability and representation robustness, and is crucial to our framework’s effectiveness. We respectfully do not consider this a minor adaptation. 3) Our model demonstrates superior performance, and the open-source code allows its concepts and components to be easily transferred to other hierarchically-structured scenarios. Computational Complexity (R2-7.3 & R3-7.2): We didn’t include such type of analysis, due to space limit. The results are as follows: UniOCTSeg (53.10 GFLOPs/92.53 M), vs. Hermes (25.67 GFLOPs/8.14 M), UniSeg (27.13 GFLOPs/42.05 M), nnUNet (234.06 [39.01 * 6] GFLOPs/123.84 [20.64 * 6] M), and UNet (591 [98.50 * 6] GFLOPs/142.5 [23.75 * 6] M). Although UniOCTSeg has relatively higher complexity, it is still comparable to baseline methods in scale and please note that it handles both seen and unseen granularities. In contrast, models like Hermes and UniSeg require reconfiguration and retraining for new/unseen granularities, while nnUNet and UNet must train extra models. These cumulative costs far exceed UniOCTSeg’s overhead. We will make targeted revisions accordingly. Pre-trained ViT Weights (R2-7.1): Our model uses the ViT-Base architecture with weights pre-trained on the ImageNet-21k dataset. We will clarify it in the final version. Cross-Validation Strategy (R3-7.1): We agree that k-fold cross-validation provides more robust evaluations. Our intention was to maintain a consistent single-split strategy across all datasets for comparability. We shall incorporate cross-validation experiments in future work. Implementation Details and Source Code (R3-7.2/7.3): We apologize for the limited implementation details due to space limit. Key settings will be added accordingly to improve reproducibility and code will be fully released upon acceptance for complete traceability. Axis Normalization in Fig. 1d (R4-7.4): We agree with the reviewer’s comment. However, please kindly understand omitting axis normalization is a common practice (MedCoss, CVPR 2024; Hermes, CVPR 2024; USFM, MedIA 2024), as it preserves dataset-specific distribution characteristics, which we followed for consistency and clarity. Edit and Formatting Issues (R4-7.3/7.4): We will address the issues in Eq.3 and Figs.1-3 in the final version.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
All reviewers acknowledge that this work achieves superior performance, supported by extensive experiments. Although it incorporates several existing modules, its task-specific designs highlight the contribution. Therefore, an acceptance is given.