Abstract

The cross-center deployment of primary gross tumor volume (GTV) segmentation models for nasopharyngeal carcinoma (NPC) on CE-T1 MRI is frequently compromised by protocol mixture. This induces systemic domain shifts, primarily manifesting as boundary drift—severe inflation of the 95% Hausdorff Distance (HD95) despite stable Dice scores.To address this, we introduce Protocol-Manifold Coverage Learning (PMCL), a source-free active adaptation framework. Operating within a minimal slice budget, PMCL selects target queries independently of source data or pseudo-labels. It quantifies protocol variation via a compact signature, isolates boundary-band Protocol Disagreement (PD) using protocol-aware perturbations, and enforces Group Prior Coverage to prevent sampling collapse. Evaluated on 1,289 volumes across six centers under a 10% annotation budget, PMCL outperforms state-of-the-art methods, improving average Dice by ∼0.10 and reducing mean HD95 by∼3.9mm.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4796_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/seatrees104/PMCL2026

Link to the Dataset(s)

CenterFS primary nasopharyngeal carcinoma MRI dataset: https://zenodo.org/records/13131827

BibTex

@InProceedings{LuJia_ProtocolManifold_MICCAI2026,
        author = { Lu, Jiawen AND Zheng, Yahan AND Guo, Qiaojuan AND Wu, Xiangjun},
        title = { { Protocol-Manifold Coverage Learning (PMCL): Source-Free Active Adaptation for Multi-Center Nasopharyngeal Carcinoma MRI Segmentation } },
        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 addresses the issue of source-free active domain adaptation (SFADA) for tumour segmentation in T1 MR images. The problem is similar to source-free domain adaptation (SFDA), the goal of which is to use a network trained on a source domain and to fine-tune it on a target domain without any labelled data. In the SFADA case, however, the objective is to select a small subset of data for expert annotation prior to model fine-tuning. More generally, domain adaptation is a crucial challenge in medical image analysis because networks are typically specific to the underlying data distribution of their training/source dataset and may perform poorly on test/target data during deployment. Furthermore, the source dataset may be unavailable due to data privacy concerns, and the budget for target domain annotation may be limited. SFADA appears to be a suitable approach for addressing each of these challenges.

    The authors propose selecting a subset of the target data based on hand-crafted features and the disagreement/uncertainty of the segmentation boundaries. In particular, hand-crafted features enable the selection of data representative of the target domain, while segmentation boundary uncertainty allows challenging cases to be detected and included for manual annotation and fine-tuning. The authors claim that by selecting cases with high boundary uncertainty, this improves the accuracy of the segmentation boundaries produced by the fine-tuned network.

    Experiments were conducted on a multi-centre dataset with a 10% annotation budget. The dataset from each of the 6 centres was defined by specific acquisition parameters, hence the need for domain adaptation. The proposed method outperformed all other domain adaptation approaches.

  • 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 addresses an important issue of interest to the medical imaging community. The proposed SFADA approach is not limited to tumour segmentation alone; it could also be adapted for other segmentation tasks. The paper is well written and organised. The methods and experiments are straightforward to follow, and the mathematical notation is clear and consistent. Despite the simplicity of the proposed approach, the extensive experiments indicate improved performance compared to state-of-the-art approaches.

    Overall, the problem setting is relevant and the experimental scope is reasonable.

  • 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 paper is closely related to references [ref13] and [ref14], with similar SFADA setups and datasets. Unlike [ref13] and [ref14], which use features from the final layer of the segmentation network to generate a latent space for selecting target domain images, the proposed approach relies on basic handcrafted features (voxel size and Fourier transform statistics). It is unclear what criteria led to this selection of features and whether alternatives were tested. Notably, these features do not incorporate any information relating to anatomical shape. Therefore, it is unclear whether the selected target data for annotation are representative of the underlying shape distribution.

    The ablation study does not clarify whether using boundary segmentation uncertainty for target data selection leads to more accurate delineation, as the Dice and HD95 scores are similar (with overlapping standard deviations) between the PMCL (full), w/o Protocol Perturb. And w/o Boundary Band rows. The lack of statistical analysis is also problematic.

    The results show that the method achieves Dice scores in the range 64–78% and HD95 scores between 3.9 and 15.5 mm across all datasets. It is unclear whether this performance is adequate in a real-world scenario; the HD95 scores are particularly high.

  • 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

    Methods:

    It is unclear why the raw probability maps are not used to compute boundary uncertainty directly.

    The authors use an intensity-based transformation to create multiple views for evaluating boundary uncertainty, but is the network trained using intensity-based data augmentation?

    In equation 4, the symbol ⊮ is not explicitly defined.

    Experiments:

    Is the target-only network trained on 10% or 100% of the target domain data? This should be clarified.

    Results:

    The authors claim ‘statistically significant increases in boundary deviation’, but the statistical test used is unclear (Student’s t-test?).

  • 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 is closely related to references [ref13] and [ref14], featuring similar SFADA setups and datasets. Even if the results indicate improved performance compared to state-of-the-art approaches, the methodological contribution remains incremental. The paper would benefit from a clearer justification of the handcrafted feature selection, stronger ablation evidence for the boundary uncertainty contribution, and statistical validation of the results.

  • 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 authors have addressed my concerns, I recommend accepting the paper.



Review #2

  • Please describe the contribution of the paper

    The paper proposes PMCL, a source-free active adaptation framework that effectively addresses boundary shift in multi-center nasopharyngeal carcinoma segmentation under low annotation budgets by explicitly quantifying protocol disparities via compact signatures and targeting boundary drift with protocol-aware perturbations.

  • 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.

    Overall, this paper addresses an important issue in active learning by proposing a dedicated sample screening strategy. The motivation is clear, the method is reasonably designed, and the experimental results show improvements over the entropy-based baseline. The work is generally well-organized and can be accepted with minor revisions.

  • 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.Why are the comparison experiments only conducted under the 10% annotation budget? How does the proposed method perform under other budget settings such as 5%, 15%, and 20%? Can the superiority of the proposed method be maintained across different annotation budgets? 2.Although the proposed strategy outperforms conventional sampling, a more detailed analysis of selected samples, for example, sample difficulty, distribution, or visualization, would help readers better understand why the strategy works. 3.Related works on query based active learning could be briefly discussed to better position the contribution of this paper.

  • 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?

    The paper targets an important and practical problem in active learning, with a clear and well-founded research motivation. The proposed sample selection strategy is logically designed and well integrated into the overall framework. The experimental results validate the effectiveness of the method, showing consistent improvements over the entropy-based baseline. The manuscript is well organized and clearly written, making it easy to follow and understand.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    In this paper, the authors address the domain shift problem in NPC GTV segmentation caused by protocol mixture across different centers. To tackle this challenge, the authors propose a source-free active domain adaptation framework, namely Protocol-Manifold Coverage Learning (PMCL). The framework consists of three key components: (1) representing each imaging volume with a compact protocol signature, (2) quantifying boundary uncertainty via Protocol Disagreement (PD) computed from a protocol-aware perturbation bank, and (3) introducing Group Prior Coverage to balance informative sample selection and protocol diversity. Experimental results demonstrate that PMCL achieves strong domain adaptation performance, particularly in improving boundary robustness under protocol variations.

  • 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 clinically meaningful problem and provides insightful observations with real-world impact. Shifting the focus in GTV segmentation from volumetric overlap to boundary drift is particularly important for radiotherapy planning. Moreover, the source-free setting aligns well with realistic clinical data governance constraints, where access to source data or features is often restricted. (2) The three core components in this work including Group Prior Coverage, Protocol Perturbation, and Boundary Band, are logically well-designed. The experimental evaluations are clear and interpretable, effectively demonstrating the contribution and necessity of each component. (3) The method is evaluated on a large multi-center dataset consisting of 1,289 volumes from six centers, across ten transfer tasks. The scale of evaluation is extensive, and the consistency in fine-tuning and testing across tasks, as well as the selection of baseline methods, is appropriate and well-justified.

  • 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) All comparative experiments are conducted only under a 10% annotation budget. The lack of performance curves across different budget levels (e.g., 5%, 10%, 20%) limits the practical insight for real-world deployment. In addition, although the paper compares with UDA methods that use source data (e.g., ADVENT, AdaptSeg), it does not provide an upper-bound result such as full supervision on the target domain (100%), making it difficult to assess the remaining performance gap. (2) Figure 3 presents qualitative results only for CenterA. It would be more informative to include examples from the most challenging target domain (e.g., CenterFS with the largest protocol gap) to better demonstrate the effectiveness of domain adaptation. (3) The current protocol signature uses only voxel area and frequency sharpness, which capture resolution and blur but ignore other important acquisition factors such as field strength, contrast agent, and reconstruction kernel. It remains unclear whether the proposed framework can effectively handle these variations, and thus its robustness across broader protocol differences is not fully validated.

  • 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?

    This paper makes a valuable contribution to a clinically meaningful problem. The three core components are well-designed and effectively achieve domain adaptation while preventing sampling collapse. The main limitations lies in the lack of annotation budget sensitivity analysis, limited coverage of protocol-related factors, and insufficient qualitative visualization. Hope the authors can address these issues in revision. Therefore, I give a weak accept recommendation.

  • 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 authors’ rebuttal addresses my main concerns. It clarifies the novelty of the paper compared with prior SOTA methods, justifies the compact protocol signature, clarifies the role of boundary protocol disagreement, and revises unsupported statistical significance claims. The explanation of the 10% active budget setting and the target-only upper bound also resolves ambiguity in the experimental protocol. Although budget sensitivity analysis and richer acquisition metadata remain limitations, partly due to page constraints, I maintain my recommendation for acceptance.



Author Feedback

We thank the reviewers and AC for the constructive comments. We focus on novelty, protocol-signature design, boundary-uncertainty evidence, budget, and reproducibility. Novelty vs. prior SFADA (R1/AC). PMCL shares the SFADA setting with STDR/Versatile-SFADA, but addresses a different query problem. Prior methods mainly select samples via latent/reference representations for representativeness or calibration.PMCL does not perform latent-space clustering/reference-star selection and requires no source images, source features, or source-derived reference vectors during query selection. Its novelty is not replacing deep features with handcrafted ones, but selecting queries for protocol-induced boundary drift by coupling acquisition-protocol coverage with boundary-band disagreement under protocol-aware perturbations. Protocol signature and uncertainty design (R1/R3/AC). The compact signature is not an anatomical-shape descriptor. log(dx·dy) captures in-plane sampling resolution, and log(FFT_H/FFT_L) captures frequency response/sharpness. We retained this 2-D signature because it is stable, interpretable, consistently available, and tied to resolution/blur shifts; richer metadata are often missing or center-specific. Shape/segmentation difficulty is handled separately by boundary-band PD. Unlike raw single-view probability uncertainty, PMCL probes prediction stability under protocol-aware perturbations and aggregates disagreement only in the predicted boundary band. Field-strength/reconstruction metadata were not consistently available, but can be incorporated when available. Statistics and ablations (R1/AC). We agree that our statistical wording was unclear. The unsupported “statistically significant” wording in the ablation/boundary-deviation discussion will be removed. We will present Table 3 as directional component evidence rather than component-level significance claims. In the reported ablation tasks, full PMCL gives the best HD95 and best/tied-best Dice, but overlapping mean±std values make isolated component significance difficult to claim. Evidence that PMCL is not merely single-view uncertainty sampling comes from Entropy under the same target pool, 10% budget, source checkpoint, and fine-tuning/evaluation. Across the ten Table-2 tasks, PMCL improves Dice and reduces HD95 versus Entropy in all tasks; two-sided paired Wilcoxon tests on task-level summaries give p<0.005 for both metrics. Experimental setting (R1/R3). Target-only is the upper bound trained with 100% labeled target data, not the 10% active budget. All query methods use the same target pool, budget, source checkpoint, fine-tuning schedule, test split, and metrics. The perturbation bank is used only for frozen-model query scoring, not as PMCL-specific training augmentation. Eq. 4 will replace the indicator symbol with 1[·]. Budget, visualization, and clinical interpretation (R1/R2/R3/AC). The 10% budget was chosen a priori as a low-annotation SFADA setting and applied identically to all query methods; thus, our claims are limited to this fixed-budget regime. We agree that 5/15/20% curves would provide useful deployment insight and will state budget sensitivity as a limitation rather than introduce new budget results. The omitted protocol-space visualization re-plots existing signatures/selection lists; if space permits, we will restore a compact version. CenterA was a conservative example, and following R3 we will replace or complement it with a compact CenterFS example from existing predictions. PMCL reduces limited-budget adaptation error but does not claim autonomous clinical deployment. Reproducibility and related work (R1/R2/R3). We will clarify implementation and data/code availability. Center1/2 are private; CenterA/B/C follow M-NPC/STDR access; CenterFS is public. Upon acceptance, we will release cleaned code for PMCL query selection, PD/boundary scoring, baselines, and finetuning/evaluation with documented formats/commands.




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.

    The reviewers provide somewhat mixed assessments of this paper (scores: 3, 4, 4), with overall positive feedback on the problem setting, experimental scale, and practical relevance, but concerns regarding limited novelty, insufficient ablation evidence, and missing analysis of key design choices . In particular, questions remain about the justification of handcrafted feature design, the effectiveness of boundary uncertainty for sample selection, and the lack of evaluation across different annotation budgets. Given these points, I encourage the authors to address the raised concerns in the rebuttal—especially by clarifying methodological choices, strengthening experimental analysis, and providing additional evidence where possible. A clear and detailed response will be important for the final decision.

  • 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 a clinically meaningful and practically important problem of source-free active domain adaptation for multi-center nasopharyngeal carcinoma tumor segmentation under limited annotation budgets. The proposed PMCL framework is well motivated, combining protocol signatures, boundary protocol disagreement, and group prior coverage to select informative and diverse target-domain samples. Although the methodological contribution is somewhat incremental and concerns remain regarding budget sensitivity analysis, statistical validation, and the limited protocol factors captured by the compact signature, the reviewers generally agreed that the rebuttal adequately addressed the main concerns. The evaluation on a large multi-center dataset across multiple transfer tasks is convincing, and the method shows consistent improvements over relevant baselines. Therefore, I recommend acceptance, with the expectation that the authors will clarify the remaining experimental details and limitations in the final version.



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.

    I recommend accept: post-rebuttal unanimous Accept.



back to top