List of Papers Browse by Subject Areas Author List
Abstract
The Segment Anything Model has demonstrated strong cross-domain generalization for image segmentation, making it a promising foundation for medical imaging. However, deploying SAM in retinal fundus analysis remains difficult because existing datasets are fragmented and task-specific, typically annotated for vessels, lesions, or optic disc/cup only, preventing effective multitask training. In addition, many retinal targets are spatially diffuse, making manual prompt engineering impractical in real workflows. We present PseudoRET, a semi-supervised framework that unifies disjoint fundus annotations to enable generalizable, multitask segmentation. PseudoRET adopts a task-agnostic training strategy and introduces a pseudo memory that aggregates and propagates supervision across tasks to mitigate annotation sparsity. To improve robustness to label noise, we further propose a confidence-aware pseudo loss that adaptively reweights pseudo labels stored in the memory, retaining reliable signals while suppressing noisy ones during training. We validate PseudoRET on 13 public datasets and 4 relabeled datasets covering vessel, lesion, and optic disc/cup segmentation. PseudoRET consistently outperforms state-of-the-art baselines across tasks and datasets. Ablation studies show that both the task-agnostic paradigm and CAP Loss contribute substantially, and their combination yields the strongest gains. Our work provides a recipe for training a SAM-style multi-target model under sparse annotations, and we release code and relabeled datasets at https://github.com/Wzhjerry/PseudoRET.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0846_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Wzhjerry/PseudoRET
Link to the Dataset(s)
DRIVE: https://www.isi.uu.nl/Research/Databases/DRIVE/
FIVES: https://figshare.com/articles/figure/FIVES_A_Fundus_Image_Dataset_for_AI-based_Vessel_Segmentation/19688169
HRF: https://www5.cs.fau.de/research/data/fundus-images/
STARE: http://cecas.clemson.edu/~ahoover/stare/
G1020: https://www.dfki.uni-kl.de/g1020
GAMMA: https://gamma.grand-challenge.org/
ORIGA: https://figshare.com/articles/dataset/Retinal_Fundus_Glaucoma_Image_dataset/24549217
Papila: https://figshare.com/articles/dataset/PAPIL
BibTex
@InProceedings{WanZho_PseudoRET_MICCAI2026,
author = { Wang, Zhonghua AND Li, Sijia AND Ju, Lie AND Feng, Wei AND Zhou, Sijin AND Hu, Ming AND Huang, James AND Huang, Anthony AND Dong, Li AND Ge, Zongyuan},
title = { { PseudoRET: A General Retinal Segmentation Framework with Multi-target Pseudo Embeddings } },
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 PseudoRET, a unified retinal segmentation framework that aims to train a single model across disjoint, task-specific fundus datasets with incomplete annotations. The core idea is to replace task-by-task training with a task-agnostic pseudo-supervision scheme built around a pseudo memory that stores recent predictions and reuses them as soft supervision, together with a Confidence-Aware Pseudo Loss that downweights unreliable pseudo labels. The paper also contributes a broad experimental benchmark spanning 13 public datasets and 4 relabeled datasets covering vessel, lesion, and optic disc/cup 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 motivation is strong: public retinal datasets are fragmented by task, so learning a single segmentation model is difficult when labels are missing or inconsistent across datasets. This is a practically important setting for fundus imaging, and the paper frames the problem clearly. 2) Rather than using static pseudo labels once, the method stores predictions from recent epochs in a FIFO memory and averages them to generate soft supervision for unlabeled tasks. This is a reasonable and practically useful way to stabilize pseudo supervision under incomplete labels, especially in a multi-task setting. 3) Pseudo labels can be noisy, especially early in training. The proposed confidence-aware weighting uses variability across recent pseudo labels to reduce the influence of unstable pseudo supervision, which is conceptually sound and well aligned with the problem. 4) The paper evaluates on datasets spanning multiple retinal targets and reports improvements over several strong baselines, including SAM-based methods and RETFound . The ablations are also useful: they isolate the effect of task-agnostic pseudo embedding, CAP Loss, and the pseudo-memory size.
- 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 methodological novelty is moderate rather than strong. The main ingredients of the method are related to established ideas in semi-supervised learning: temporal aggregation of predictions is closely connected to Temporal Ensembling and Mean Teacher, while reliability-based weighting of pseudo supervision is also well explored in semi-supervised medical segmentation. As a result, the paper’s novelty appears to lie more in the combination and application to fragmented retinal segmentation than in a fundamentally new learning principle. 2) The paper does not sufficiently position itself against prior semi-supervised segmentation literature. Because pseudo-memory plus consistency-style supervision is central to the method, the paper would be stronger if it explicitly compared against representative teacher-student or temporal-ensemble baselines, rather than only SAM-family methods and RETFound. Without such comparisons, it is difficult to isolate whether the gain comes from the proposed retinal formulation or from a more standard semi-supervised effect. 3) The evaluation protocol relies heavily on relabeled datasets, but annotation reliability is not deeply analyzed. The paper states that four public datasets were relabeled and that added annotations were reviewed by three retinal specialists, which is positive, but there is no reporting of inter-rater agreement, adjudication protocol, or annotation uncertainty. Since the relabeled benchmark is central to the claimed superiority, this missing detail weakens confidence in the fairness and robustness of the evaluation.
- 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.
- 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?
see major strengths and major weaknesses
- 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 author addressed the concerns.
Review #2
- Please describe the contribution of the paper
The paper proposes PseudoRET, a semi-supervised multi-target fundus image segmentation framework. It also establishes a benchmark for multi-target 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 introduces a task-agnostic pseudo-embedded training scheme and a confidence-aware dynamic pseudo loss. 2.It builds the benchmark based on public datasets and states that the code and data will be released. 3.The experimental results are better than those of current SAM-based methods and existing foundation models for fundus images.
- 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 empirical comparison is not sufficiently comprehensive. The paper lacks comparisons with mixed-domain semi-supervised methods and other semi-supervised methods. In addition, universal domain adaptation may also be relevant for handling label-space mismatch, and it is worth considering whether such methods should be included in the comparison. 2.In Table~2, what do Base and Large refer to? It is also unclear which SAM variants are used in the proposed method and in the compared methods.
- 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.
- 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 problem setting and benchmark construction in this work have strong practical value, and the experimental results also show clear advantages. However, the scope of the comparisons and the description of the experimental setup are still not sufficiently clear and should be further clarified 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.
Reject
- [Post rebuttal] Please justify your final decision from above.
I appreciate the authors’ clarifications regarding the problem formulation, the meaning of Base/Large, and related implementation details, which partially improve the clarity of the paper.
However, my main concern about insufficient empirical comparison remains unresolved. I understand that new experiments are not allowed at the rebuttal stage, but the authors neither provide comparisons with representative mixed-domain semi-supervised methods or related domain-adaptation approaches, nor sufficiently justify why these relevant baselines were excluded. The internal ablation in Fig. 3 helps analyze the proposed components, but I do not think it can replace direct comparison with closely related prior methods.
Since this concern was raised by multiple reviewers and is important for assessing both the empirical strength and the claimed novelty of the method, I do not think the rebuttal adequately addresses the core weakness of the submission. Therefore, despite the practical value of the problem and benchmark, I recommend rejection.
Review #3
- Please describe the contribution of the paper
The paper presents a framework for fundus multi-task segmentation and reports competitive performance on several 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 paper presents a semi-supervised framework for task-agnostic fundus multi-task segmentation. 2.The experimental design is well structured. 3.The qualitative results show good segmentation performance on fundus multi-task segmentation.
- 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 method is presented as a semi-supervised, task-agnostic framework, but the actual setting seems closer to multi-task learning with incomplete annotations across multiple datasets. This distinction is not clearly explained, and the problem setting should be described more precisely. 2.The proposed method mainly uses a pseudo-memory mechanism and simple consistency-based designs. The overall design is relatively straightforward, and the novelty of key components, such as the pseudo-memory and CAP loss, appears limited. 3.The paper states that it is evaluated on 13 datasets, but multi-task results are only reported on the relabeled datasets with complete annotations. The role of the remaining datasets appears to be mainly to provide additional supervision during training. This setup should be clarified to avoid confusion.
- 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.
- 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 shows promising results on fundus multi-task segmentation, but suffers from limited methodological novelty, unclear dataset usage, and insufficient clarity in writing and explanation.
- 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.
Rebuttal is satisfactory
Author Feedback
We thank R1-R3 for the constructive reviews and address the main concerns below.
(1) Problem setting and dataset usage (R3, R2). PseudoRET targets cross-task annotation incompleteness: each image has GT for only one task, while other tasks are unannotated. This can be viewed as (i) multi-target learning with incomplete annotations, or (ii) semi-supervised learning along the task-label axis, where the unlabeled dimension is which task is annotated, not which pixels. The two views are complementary; we will clarify this in §1. For training vs evaluation (R3), all training uses only original single-task labels. For DRIVE, HRF, GAMMA, and IDRiD, only their original single-task annotations enter training. The relabeled multi-target annotations are used solely for evaluation. Together with Maples-DR, whose native multi-target annotations are used only on its test split, these form the unified multi-target evaluation benchmark. They never serve as auxiliary training supervision for one another. The §3.4 ablation reports per-task DSC on each dataset’s own test split. We will tighten §3.1 and §3.3. (2) Backbone, Base/Large, and SAM variants (R2). PseudoRET is “SAM-style” only in producing one model for multiple targets without prompts; it does not use a SAM backbone. The architecture is SwinUNETR. PseudoRET-Base/-Large denote Swin ViT-B/-L. All compared SAM-family methods (Med-SAM-Adapter, MedSAM2, SAM3, SAM2-UNet) use their official configurations from the released repositories without modification. This will be stated in Table 2’s caption.
(3) Relation to Mean Teacher and Temporal Ensembling (R1, R3). These methods operate in a fixed label space and address pixel-level sparsity for a single task via teacher-student consistency. PseudoRET is fundamentally different: its pseudo memory shares supervision across tasks for the same image, addressing label-space mismatch across datasets. For image x, tasks without annotations receive pseudo supervision from the model’s predictions on x for those tasks. The mechanisms are orthogonal, not substitutes. CAP Loss likewise differs from entropy/confidence weighting: it uses temporal dispersion over the FIFO memory to capture the stability of cross-task pseudo supervision across recent epochs.
Crucially, Fig. 3 already isolates these effects within our protocol. The “Task-specific Pseudo Embedded” variant restricts each image’s pseudo memory to its own annotated task, exactly a single-task temporal-ensemble baseline inside our framework, and yields only marginal gains over the baseline. The +2.1% average DSC gain from task-specific to task-agnostic embedding therefore comes from cross-task sharing, not a generic SSL effect; the further +1.2% from CAP Loss isolates dispersion-based weighting on top (combined +3.3% over baseline). Fig. 3 thus provides the controlled comparison requested.
(4) Missing SSL/mixed-domain/UDA baselines (R1, R2). MICCAI 2026 rebuttal guidelines disallow new experimental results. As argued in (3), Fig. 3 already separates cross-task sharing from a single-task SSL effect under an identical training corpus and protocol.
(5) Annotation protocol (R1). Data were divided into three disjoint subsets. Each subset was annotated by one of three retinal specialists under a unified standard. The other two then cross-checked it. Any disagreement was discussed by all three until consensus. This will be added to §3.1.We agree that quantitative inter-rater agreement would further strengthen the benchmark and will include such statistics with the public dataset release.
We hope these clarifications address the main concerns and respectfully ask for reconsideration.
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 paper addresses a practically relevant problem and reports promising empirical results; however, several important issues raised by the reviewers require clarification. First, the methodological novelty is considered moderate, as the approach largely builds upon established semi-supervised learning paradigms (e.g., temporal ensembling, Mean Teacher, and reliability-based pseudo-labeling). The authors should better articulate the key novelty and more clearly position the method against prior semi-supervised segmentation approaches, including appropriate teacher–student and temporal-ensemble baselines.
Second, the experimental evaluation and setup need to be strengthened. The current comparisons are not sufficiently comprehensive, and reviewers suggest including stronger baselines such as mixed-domain semi-supervised methods or related domain adaptation approaches. In addition, several details remain unclear, including the meaning of “Base” vs. “Large,” the specific SAM variants used, and the overall training protocol.
Third, while the relabeled datasets are a valuable contribution, the paper lacks sufficient analysis of annotation reliability (e.g., inter-rater agreement and adjudication process), which is critical to support the validity of the benchmark.
Finally, the problem formulation should be clarified, as the current setting appears closer to multi-task learning with incomplete annotations rather than standard semi-supervised learning. The roles of the multiple datasets and the contributions of key components (e.g., pseudo-memory, CAP loss) should also be more clearly explained.
- 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 reviewing the manuscript, the reviews, and the authors’ rebuttal, I recommend acceptance.
The paper addresses an important and practically relevant problem in retinal image segmentation under fragmented annotation settings. While the methodological novelty is moderate and builds upon existing semi-supervised learning concepts, the overall contribution extends beyond a single algorithmic component and includes a valuable benchmark, extensive relabeling effort, and large-scale evaluation across multiple retinal datasets.
The main concerns raised during the initial review process involved the problem formulation, the role of fragmented annotations, the novelty of the proposed components, and the evaluation protocol. I find that the rebuttal provides satisfactory clarification on these issues, which is reflected by the score increases from both Reviewer 1 and Reviewer 3. Reviewer 2 continues to express concerns regarding the lack of comparison with certain related semi-supervised and domain adaptation approaches. While additional comparisons would strengthen the paper, I do not consider this issue sufficient to outweigh the paper’s overall contribution and empirical value. The presented experiments and analyses are adequate to support the main conclusions.
Overall, the paper presents a sound and useful contribution to the retinal image analysis community. The remaining weaknesses are primarily related to experimental completeness rather than fundamental flaws in the methodology or evaluation. Therefore, I recommend acceptance.
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 considering the reviews and the rebuttal, I recommend acceptance. The paper addresses a practical problem setting and provides a benchmark with clear value. The rebuttal clarifies the dataset usage, training/evaluation protocol, Base/Large settings, SAM-related implementation details, and annotation procedure. It also explains the scope of the empirical comparison and the role of the internal ablation. I therefore recommend acceptance.