List of Papers Browse by Subject Areas Author List
Abstract
Retinal vessel segmentation is critical for diagnosing ocular conditions, yet current deep learning methods are limited by modality-specific challenges and significant distribution shifts across imaging devices, resolutions, and anatomical regions. In this paper, we propose GrInAdapt, a novel framework for source-free multi-target domain adaptation that leverages multi-view images to refine segmentation labels and enhance model generalizability for optical coherence tomography angiography (OCTA) of the fundus of the eye. GrInAdapt follows an intuitive three-step approach: (i) grounding images to a common anchor space via registration, (ii) integrating predictions from multiple views to achieve improved label consensus, and (iii) adapting the source model to diverse target domains. Furthermore, GrInAdapt is flexible enough to incorporate auxiliary modalities—such as color fundus photography—to provide complementary cues for robust vessel segmentation. Extensive experiments on a multi-device, multi-site, and multi-modal retinal dataset demonstrate that GrInAdapt significantly outperforms existing domain adaptation methods, achieving higher segmentation accuracy and robustness across multiple domains. These results highlight the potential of GrInAdapt to advance automated retinal vessel analysis and
support robust clinical decision-making. Our code is at https://github.com/YuekaiXuEric/GrInAdapt.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0903_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YuekaiXuEric/GrInAdapt
Link to the Dataset(s)
BibTex
@InProceedings{LiuZix_GrInAdapt_MICCAI2025,
author = { Liu, Zixuan and Honjaya, Aaron and Xu, Yuekai and Zhang, Yi and Pan, Hefu and Wang, Xin and Shapiro, Linda G. and Wang, Sheng and Wang, Ruikang K.},
title = { { GrInAdapt: Source-free Multi-Target Domain Adaptation for Retinal Vessel Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15964},
month = {September},
page = {218 -- 228}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes GrInAdapt, a novel framework for source-free multi-target domain adaptation of fundus images. It consists of three key steps, namely, Grounding, Integrating and Adapting. Experiments on some novel domain adaptation setting show its effectiveness.
- 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.
-
This paper establishes a new evaluation protocol for cross-domain and cross-modality fundus image segmentation.
-
Clear performance improvement over the baseline.
-
The authors have made the source code publicly available.
-
- 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 technical innovation of the proposed method may not be especially devised for cross-domain and modality adaptation.
-
Besides, the first two components, i.e., mask proposal and regional merging, are not uncommon for existing image segmentation methods and the corresponding design.
-
The network module, technical design and insight is relatively old and out-of-date. The better generalization from recent foundation model or AIGC is not leveraged at all.
-
Lack sufficient comparison with state-of-the-art methods. The compared methods are [4] and [24], one is from 2021 while the other from 2024. Significantly more recent medical image segmentation, 3D medical image segmentation and domain adaptation medical image segmentation methods, in 2022-2024, should be involved for comparison.
-
Lack a direct comparison with recent foundation model driven medical image segmentation methods, including but not limited to SAM, medical SAM and etc.
-
There is no visual prediction result between the baseline and the proposed method.
-
How does the lambda hyper-parameter in Eq.2 impact the overall performance? More ablation study is needed.
-
The title is not very rationale, too long and need to be condensed significantly. In my view, both ‘multi-site’ and ‘multi-device’ can be attributed together to ‘multi-domain’. Or the title can directly use ‘source-free multi-target domain adaptation’.
-
- 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?
This paper has some merit to the field, but is relatively limited. The technical design and innovation is not strong, and the modules are out-of-date. Besides, the compared state-of-the-art are out-of-date and limited. Therefore, the reviewer votes to reject this paper.
- 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.
Thanks for the rebuttal.
My concerns on novelty and comparison with more cutting-edge techniques are not properly addressed in this rebuttal.
Therefore, I maintain my initial rating.
Review #2
- Please describe the contribution of the paper
A framework，GrInAdapt, is proposed in the manuscript to segment CAVF in OCTA images. The main contribution is the design for adaption to different fundus domains. The images in different domain are aligned first via vessel segmentation. The prediction labels are refined by integration of prediction in multiple views. The teacher-student learning approach is employed in the adaption to target 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.
A suitable framework for OCTA segmentation is proposed. The multiple views are employed for improving performance. Experiments shows the performance of the adaption in the proposed framework is satisfactory.
- 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 title is too long and redundant. The adaption is mainly among different source of data, which is domain adaption. There is not much about multi-modality adaption. There is no explanation how the registration is evaluated. Does successful registration mean the images can be registered, but not necessary registered well?
- 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.
(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?
The proposed framework is applicable in clinical images and the framework works well for domain adaption. The method proposed in simple.
- Reviewer confidence
Somewhat confident (2)
- [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 addresses most of my concerns.
Review #3
- Please describe the contribution of the paper
This paper proposes a method for retinal vessel structural map segmentation that consists of three key steps: grounding, label refinement, and adaptation, aimed at addressing domain adaptation across multiple clinical sites. The approach effectively improves segmentation performance in multi-site 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.
The motivation is clear, and the addressed problem is both relevant and high-impact in the field of biomedical image analysis. The problem formulation is well-defined, and the manuscript is clearly written with effective visualizations. The proposed method demonstrates strong performance and significantly enhances multi-domain adaptation.
- 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.
Subject-Level Image Pairing Assumption The proposed approach relies on the assumption that paired images for each subject are available across multiple domains. However, in real-world clinical settings, this may not be a realistic assumption, as patients typically undergo imaging at a single center. It would be helpful to understand how the model handles scenarios where such cross-domain paired data is unavailable. Can the authors make a comment about the general application of their method? Furthermore, the method assumes access to an auxiliary model capable of producing complementary segmentation maps. Is this assumption practical in deployment scenarios? Would any auxiliary segmentation task be sufficient, as long as there is some overlap in label space with the primary task?
Model Design and Assumptions The manuscript mentions that “optionally, paired data from auxiliary modalities such as color fundus photography (CFP) can also be incorporated,” yet later assumes the availability of both auxiliary inputs and a source model capable of generating auxiliary labels. Could the authors clarify whether these components are optional or required for the method to function effectively?
The proposed model involves multiple components. It would be valuable to include an ablation study showing the impact of each part on overall performance—for example, results without the confidence loss term (L_conf), or without the teacher-student framework. Since the teacher-student setup increases computational complexity, how does its resource requirement compare to other adaptation methods? Can the authors provide some insight into training cost or runtime overhead of this design?
- 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 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
I suggest the authors remove the coloring from the “Methods” column in Tables 1 and 2, as it is visually distracting and makes it harder to clearly differentiate between the individual methods.
- 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?
I have some questions and concerns regarding the complexity of the proposed method, it assumptions and applicability in real-world clinical settings, which I have outlined in the Weaknesses section, which I hope the authors clarify.
- 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.
After reading authors’ response, I recommend accepting the paper.
Author Feedback
We thank all reviewers for their constructive comments acknowledging our “clear performance improvement” (R3), “new and applicable clinical framework” (R1,3), and contribution to this “relevant and high-impact” problem (R2). Below we address main concerns.
(R1,3):Title, multimodal adaptation questioned We agree the title is too long and over-claims on multi-modality adaptation. We will change to “GrInAdapt: Source-free Multi-Target Domain Adaptation for Retinal Vessel Segmentation” (suggested by R3).
(R1):Missing details of registration evaluation We acknowledge this limited presentation. GrInAdapt works with adequate registration, evaluated by DSC score between target and registered vessel masks. The median value is 60% for cohorts with successful registration and 7% for failed cohorts.
(R2):Auxiliary data/model availability, label space overlap, deployment challenges Auxiliary data/model is indeed optional. GrInAdapt works without auxiliary CFP Artery/Vein (A/V) predictions. Still, extra A/V information helps generalization in optic disc and 12x12 regions. Label space overlap is the minimal requirement to enable the option. Auxiliary data usability depends on its prediction reliability. Since A/V are two reasonable, overlapping classes, and several public CFP A/V segmentation models are available, it’s simple to deploy. We will remove colors in tables as suggested.
(R2): Paired images assumption Only training needs subject-level pairs; the adapted model works on a single image from any domain. This is clinically realistic during development, as evidenced by the real-world multi-site dataset AI-READI containing such pairings. In principle, GrInAdapt can utilize paired multi-domain data whenever available, e.g., GrInAdapt would also work for longitudinal analysis (image pairs collected across time). Although unpaired DA needs additional techniques, GrInAdapt only needs two images per subject, which is not often hard to acquire in practice, making it suitable for a wide range of clinical applications.
(R2,3):Ablation studies and computation Removing the teacher-student module is effectively the modified DPL[4] baseline. Removing L_conf (README in code repo) gives slightly lower results than GrInAdapt. Adding a teacher-student module (latter) has light runtime & spatial overhead (6.5h vs. 7h/epoch, 8397 vs. 8763MiB, bs=2). We test λ=0.5 & 2 in Eqn.2. Results are stable.
(R3):Limited technical novelty, network design, lack of SOTA comparisons, missing visualization Consistently segmenting vessels from OCTA across devices is clinically important (R1,2). Yet, there is limited work on domain adaptation (DA) for this task. Previous DA methods [4,24] focused on adapting CFP optic disc/cup segmentation. Yet, adapting vessel structures on OCTA has different challenges. GrInAdapt is, to our knowledge, the first DA work for OCTA vessel segmentation. There is no existing DA framework directly applicable. Idea from [14] is hard to apply; we significantly modified [4,24] to work on OCTA vessel segmentation. Given it’s a new task, we intended not to go far on methodology design, but to design an effective pipeline using well-established components (mask proposal, regional merging, teacher-student) for this novel task. Comparison with general 3D/2D-only segmentation networks would not properly evaluate multi-target DA performance, and is not a common practice in recent DA works ([4,14,24]). See prediction visualizations at README in repo.
(R3):missing use of/comparison with FM GrInAdapt is independent of source model structure and may benefit from foundation models like SAM. However, CAVF segmentation is not easy for SAM models. Recent work (SAM-OCTA DOI:10.1016/j.bspc.2025.107698) shows that SAM only supports 2D inputs and needs extensive fine-tuning per class (4 separate models) and multiple 2D en face maps from different layers (not consistently available across devices), limiting its multi-target DA applicability (Visuals in repo).
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’
N/A