List of Papers Browse by Subject Areas Author List
Abstract
Deep learning (DL) methods are challenged to demonstrate robust performance across different segmentation datasets due to domain shifts, but active domain adaptation techniques enhance their generalization performance by querying a few samples from target domains for adaptation training. However in clinical practice, target domains often include private classes of new anatomical structures or pathologies that are not presented in the source data, and existing methods implement closed-set segmentation where source and target domains have the same segmentation classes. Additionally, source data are often inaccessible during adaptation due to strict data privacy regulations. To address these limitations, we propose an Active Source-free Open-set Domain Adaptation (ASFOSDA) method which is the first work to implement active learning for adapting DL models in open-set medical image segmentation without the access to source data. This method employs an active open-set query strategy to select the most informative target samples for training models based on Class-aware Decomposed Uncertainty (CDU) and Class-agnostic Prototype Discrepancy (CPD). CDU measures sample aleatoric uncertainty and model epistemic uncertainty by employing test time augmentation in stochastic processes. CPD measures cross-domain and self-domain discrepancy for selecting diverse samples. Subsequently, to boost the adaptation performance by enhancing training samples, a Target-refined Self-training strategy is proposed to generate high-quality pseudo labels for unselected samples, thus combining them with labeled samples for a semi-supervised training. We evaluated our method on cross-domain open-set volumetric medical image segmentation tasks, and it outperformed state-of-the-art adaptation methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2507_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanJin_Active_MICCAI2026,
author = { Yang, Jin AND Zhang, Yichi AND Qiu, Peijie AND Yu, Xiaobing},
title = { { Active Source-free Domain Adaptation in Open-set Medical Image Segmentation via Decomposed Uncertainty and Prototype Discrepancy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes an Active Source-free Open-set Domain Adaptation (ASFOSDA) framework for medical image segmentation. The work deal with a realistic setting where source data are inaccessible in adaptation, the target domain contains new classes, and only a small annotation budget is available in the target domain. This framework selects informative target samples for annotation using two criteria: (1) Class-aware Decomposed Uncertainty (CDU), which combines class-aware aleatoric uncertainty (CAU) and class-aware epistemic uncertainty (CEU) to focus on unknown target classes; (2) Class-agnostic Prototype Discrepancy (CPD), which measures both cross-domain discrepancy between low-uncertainty “source-like” target samples and high-uncertainty target samples, and self-domain discrepancy among candidate target samples to reduce redundancy. To further improve performance, the paper adds a Target-refined Self-training stage. Experiments on open-set CT and MR adaptation tasks show improved performance over SFDA and ASFDA baselines under a 5% target annotation budget.
- 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 problem setting is more realistic than standard closed-set DA in many clinical settings, where new organs/pathologies may exist in target data and source data may be unavailable due to privacy constraints. This problem formulation is meaningful and timely.
2.The decomposition into CAU and CEU is sensible: CAU targets uncertainty related to target-pattern ambiguity, and CEU targets uncertainty related to model/domain mismatch.
3.The method does not stop at querying; it also includes target-refined self-training to further improve the performance.
4.Overall, ASFOSDA consistently outperforms SFDA and ASFDA baselines across all three tasks and both backbones. The gains over the strongest active baselines such as UGTST are nontrivial. Testing on open-set CT and MR segmentation with both 3D U-Net and Swin UNETR strengthens the generality of the empirical findings.
5.The ablation studies are through. Table 2 and Fig. 3 provide evidence that: combining CAU and CEU is better than either alone; CPD improves over uncertainty-only querying; self-training adds further gains; the choice of prototype/candidate number matters. These analyses help support the proposed design.
- 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.In true source-free open-set adaptation, the source-trained model would normally not have explicit output channels for classes unseen in the source labels. The paper does not explain clearly how the segmentation head is initialized to support target-only classes before target annotations are acquired.
2.Some comparisons seem not be fully fair because the baselines many compared ASFDA baselines appear designed for standard active DA rather than the proposed open-set source-free setting. Since the proposed method is tailored to detect and prioritize unknown-class information, it benefits from design assumptions that the baseline methods do not share. This does not invalidate the comparison, but the paper should acknowledge more clearly that these are not equally matched open-set baselines.
3.The Target-refined Self-training stage is useful, but conceptually less novel than the active open-set query strategy. The paper should emphasize more clearly what is novel in the pseudo-label refinement itself, versus using a generally standard two-stage supervised + pseudo-labeling recipe.
4.The paper is readable overall, but some notation and assumptions are not sufficiently precise. Especially, how should we treat known vs unknown classes? How logits/probability maps are produced for target-only classes before adaptation? The authors should explain these much more explicitly.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My major concern is: regarding how unknown target classes are represented by the model before adaptation and whether the full target class set is assumed known in advance.
- 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 novel active domain adaptation approach without access to source data. The authors use a confidence-based approach to identify out-of-domain samples, which leverages test-time augmentation to estimate uncertainty (confidence). Domain adaption is an impactful problem to address in the medical image analysis community. Solutions to this task have the potential to enhance translational impact of automated deep learning segmentation approaches that otherwise suffer from challenges of domain shift.
- 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.
- Evaluation is performed using 4 publicly available datasets, which is good for reproducibility by the community.
- Experiments use 2 commonly used network architecture (U-net and Swin UNETR) to test the proposed approach, so results should be informative and relevant to the community that uses these models frequently.
- Experiments are used to quantify both lower and upper bound segmentation performance using the zero-shot source only evaluation and fully supervised target domain evaluation, respectively.
- Overall, the proposed approach demonstrates what appears to be significantly better segmentation results (Dice and 95HD) compared to baseline domain adaptation methods and does so across the two tested network architectures.
- Ablation studies are performed and help to present evidence in support of the approach design decisions.
- 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.
- It is unclear if the code implementation will be made available, which potentially limits reproducibility.
- (minor) Based on comparison to the upper bound, domain adaptation methods still have a long way to go toward realizing the full performance of fully supervised training. Nonetheless, this method is still a good step forward compared to other baselines.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The results demonstrate that the proposed approach does indeed improve segmentation performance compared to baseline methods. However, I want to raise the following general question. There is a fundamental question with respect to the approach: low uncertainty (high confidence) does not guarantee anything about correctness. A model can be 100% confident (no uncertainty) in otherwise incorrect prediction. How is such as case handled in any of these adaptation frameworks? In other words, how does one disentangle confidence and correctness through a measure such as entropy?
Grammatical: Page 5: “is employ” -> “is employed”
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting methodological approach that demonstrates good domain adaptation segmentation performance gains compared to a rigorous set of baselines. Such an approach has good potential to realize impactful translational significance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
This work combines novel TTA (Test Time Augmentation) strategies to estimate data and model uncertainty using entropy and energy formulations with embedding-based discrepancy scoring for source-free open-set domain adaptation in segmentation models. The methods are mostly well-explained. However, important experimental details are unclear, making objective evaluation difficult.
- 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 results outperform every other DA (Domain Adaptation) technique across multiple datasets, and the ablation study clearly demonstrates the additive effect of each pipeline component.
2.Results at 5% target query budget are impressive compared to full target set training. DSC, HSD, and standard deviations show consistent improvement across all experiments.
- 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 methodology does not clearly describe how the source model was configured to produce probability maps or logits corresponding to unknown classes within the target dataset. Clarifying this setup is essential for understanding how the model handles class discrepancies across domains. 2.The comparison with prior domain adaptation (DA) methods omits details on the degree of hyperparameter tuning or optimization conducted to ensure fair performance comparison. Including this information would strengthen the validity of the reported results. (specific details in the comments) 3.The composition of target images selected for annotation—by the proposed approach versus baseline DA techniques—should be explicitly reported. Quantifying the overlap and differences (e.g., number of distinct images selected) would provide an objective evaluation beyond model accuracy alone. 4.The study does not include cross-modality domain adaptation experiments (e.g., CT→MRI or MRI→CT), which limits the generalizability and robustness of the proposed method’s 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 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
1.Normalization in CDU calculation: The normalization applied to CEU and CAU when computing CDU is not clearly explained. Please provide more detail on the specific role and implementation of the normalization term (“Norm”) in Equation 7.2.Loss definition in Section 2.2: The manuscript defines two loss functions, L_sup and L_unsup, for supervised and unsupervised samples, but it is unclear whether these are implemented differently or simply represent the same BCE + Dice formulation written separately. Please clarify whether there is any functional distinction between them in practice. 3.Model expansion – unknown class handling (Methods §2.1): The source model F(Θ), pre-trained on source classes Cs, generates probability maps for the full set of target classes Ct (where Ct includes Cs and unknown target-specific classes Cu). Please clarify the model adaptation for unknown classes Cu. 4.Datasets and splits: Please specify the train/test split sizes used for each of the five datasets. For full target supervision (the upper-limit experiment), what is the test set size? At 5% labeling, what is the number of labeled training samples per dataset? 5.Baselines and training details: The comparison with prior domain adaptation (DA) methods lacks clarity regarding hyperparameter tuning and training duration. For the proposed Target-refined Self-training approach, does the final model train for 1000 epochs total (500 + 500) or 500 epochs in total (250 + 250)? If the other baseline techniques were limited to 500 epochs total, the proposed technique should be evaluated at 500 total including the target-refined self-training step. 6.DSC performance drop at 5% labeling: The Dice score decreases beyond 20% queried target samples, despite implemented mechanisms for redundancy reduction. Please clarify this non-monotonic behavior, as performance would typically be expected to increase or saturate with more labeled data. 7.Additional labeling ratios: It would be useful to report results at additional labeling ratios such as 2.5% and 10% to demonstrate consistency of performance trends relative to the next-best DA technique. 8.Cross-modality experiments: The manuscript lacks cross-modality adaptation experiments (e.g., CT-to-MRI or MRI-to-CT). Within MRI, please also clarify whether the method generalizes across sub-modalities (for example, between T1- and T2-weighted data). 9.Query selection comparison: Consider adding a table summarizing how many queried target images (or exams) differ between the proposed approach and prior DA techniques at key query thresholds (e.g., at 20 queries, how many samples differ from the next-best method?). 10.Random seed consistency: Was the first randomly initialized sample fixed across all techniques? If not, variations in the initial sample could introduce confounding factors affecting performance comparison. Please clarify this experimental detail.
- 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 methodology is well grounded in previous work and the results are promising, but there are several experimental details that should be clarified for completeness.
- 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
Author Feedback
We thank the reviewers and meta-reviewers for the feedback, and address main concerns below: -Clarity on comparison with SOTA methods (R1Q2, R3Q2): To the best of our knowledge, our work is the first active source-free domain adaptation framework designed for open-set medical image segmentation. Therefore, there are limited existing active source-free domain adaptation methods specifically developed for this task available for comparison. Furthermore, to fully validate the effectiveness of our approach, we compared our method with various baselines, including unsupervised (source-free) domain adaptation methods, traditional active learning strategies, and active (source-free) domain adaptation methods. To ensure fair comparisons, all methods were initialized using the same samples and initialization settings, and were trained using the same adaptation protocols and hyperparameters. Following the reviewers’ suggestions, we updated and expanded the descriptions of the baselines and experimental settings in the manuscript. -Clarity on notations and assumptions (R1Q1, R1Q4, R3Q1): we updated our descriptions about known and unknown classes in our manuscript to improve clarity and readibility. -Clarity on experimental details (R3Q1): to improve the reproducibility and clarity, we updated the descriptions of normalization in CDU calculation, loss function, and training epochs. -Generalizability on cross-modality adaptation (R3Q4): To demonstrate the generalizability of our approach, we evaluated our method on five different medical image segmentation tasks with varying numbers of segmentation classes and adapted two different segmentation architectures (e.g., U-Net and Swin UNETR). Evaluation on cross-modality adaptation remains an important future direction, and we plan to implement it in future work. -Grammar error (R2): we fixed the grammar errors and typos (e.g., “is employ” -> “is employed”).
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The three reviewers are consistently positive about the paper (all scores: 4), highlighting the realistic problem setting of source-free open-set domain adaptation, strong empirical performance across multiple datasets and backbones, and well-supported design through comprehensive ablations. However, they also raise several important concerns, including unclear handling of unknown target classes, potential fairness issues in baseline comparisons, and missing experimental details that affect reproducibility and clarity . Overall, while some aspects of the formulation and evaluation require clarification, the contribution is meaningful, timely, and demonstrates clear improvements over prior work. Based on the strengths of the method and the reviewers’ generally favorable assessments, I recommend acceptance.