List of Papers Browse by Subject Areas Author List
Abstract
Reliable pelvic bone segmentation (PBS) from CT is essential for robot-assisted pelvic trauma surgery, yet deploying a source-trained model to a new hospital suffers from severe performance degradation due to cross-center domain shifts. While test-time adaptation (TTA) enables online model adaptation without accessing source data, existing methods show limited effectiveness for PBS, facing challenges including boundary degradation, anatomical inconsistency under domain shifts, and voxel-level class imbalance. To address these challenges, we propose a novel closed-loop dynamic Reliability-Guided TTA framework (ReGA) for PBS. Specifically, we introduce a pseudo-label reliability criterion termed Segmentation Inference Consistency Evaluation (SICE), which jointly measures region overlap and boundary deviation via dropout-based ensemble predictions. Based on SICE, a trust-weighted refinement module adaptively updates features to mitigate boundary errors in pseudo-labels. Furthermore, a confidence-weighted region-level contrastive learning strategy is proposed to enforce anatomical consistency. Finally, ReGA follows the teacher-student (TS) scheme to alleviate voxel-level class imbalance.
Experiments on three heterogeneous 3D pelvic CT datasets demonstrate that ReGA consistently outperforms state-of-the-art TTA methods, enabling effective adaptation of the source-trained PBS model to unseen clinical domains. The code is available at https://github.com/Ren-ling/ReGA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1573_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Ren-ling/ReGA
Link to the Dataset(s)
CLINIC: https://zenodo.org/records/4588403#.YEyLq_0zaCo
KITS19: https://github.com/neheller/kits19
MSD_T10: https://drive.google.com/file/d/1m7tMpE9qEcQGQjL_BdMD-Mvgmc44hG1Y/view
BibTex
@InProceedings{RenLin_Testtime_MICCAI2026,
author = { Ren, Ling AND Deng, Chao AND Wang, Ziming AND Xu, Yuecong AND Zheng, Kai},
title = { { Test-time Adaptation of Pelvic Bone Segmentation Models via Dynamic Reliability-Guided } },
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
The paper proposes a test-time adaptation method (ReGA) for improving pelvic bone segmentation under domain shift. It introduces a reliability score (SICE) based on dropout-based ensemble predictions to guide feature refinement and model updates during inference, combined with contrastive learning. This enables improved segmentation performance for adapted source-trained models on unseen clinical 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.
-
Background is well formulated and the problem is of genuine interest to the community
-
Evaluation covers important datasets in the field
-
Visualisations for the method and results are high quality
-
- 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.
-
A few terms could benefit from an expanded explanation to make this work self-contained. One example is ‘teacher-student schemes’, which is a very wide area of research and requires a brief 1-2 sentences explanation to allow readers to fully understand the shortfalls.
-
Several claims are unsupported by citations/evidence, please add citations to previous work to support claims: TS schemes are confined by false prediction and error accumulation at these boundaries enforcing anatomical consistency is crucial for the reliable adaptation existing methods neglect the self-supervised abilities
-
Some maths terms are undefined at first use e. g. , \theta^t is missing a definition in the first paragraph of sec 2.
-
Maths definitions are mixing model and parameters; specifically theta is used to denote parameters but later on terminology such as ‘fed into \theta^t’ implies it is a model; this is confusing
-
MC-dropout is not explicitly defined with k remaining ambiguous; how is k iterated?
-
Some of the results seem to be within 1 St. D. of the proposed method so claims of superiority seem overstated; additionally, statistical tests are missing so it is difficult to determine whether performance improvements are meaningful or not
-
There are several hyper-parameters (e. g. , dropout rate, k, M, beta etc. ) however, performance sensitivity to these is not evaluated in depth
-
The overall novelty is somewhat limited as MC-dropout based uncertainty, contrastive learning etc. are well-established in previous TTA and SFDA literature. Their combination is novel in this application but to be positioned as more application-focused the evaluation needs to be stronger (see comments above).
-
- 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.
- 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?
As highlighted in the weaknesses section, there is some lacking in clarity of maths notation, which is coupled with some missing depth of analysis in the results. Furthermore, as this is more application focused, a strong evaluation (e.g., robust statistical testing, hyperparameter evaluation etc.) is necessary to reach the threshold of publication at MICCAI.
- 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 sufficiently addressed the feedback.
Review #2
- Please describe the contribution of the paper
The paper proposes a test-time adaptation (TTA) framework using a teacher-student scheme to enhance anatomical consistency and improve bone boundary segmentation for pelvis bone segmentation from CT scans. The authors introduce a segmentation inference consistency evaluation (SICE) that depends on a dropout-based ensemble segmentation of regions and boundaries. Later, SICE is combined with a weighted adaptive module to refine features. The proposed methods enhance the structural shape (anatomical consistency) of the segmented bone compared to other TTA approaches, but still face challenges in differentiating segmentation classes. Authors validate the method on three 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.
While the concept of sampling from dropout is used to define reliability or uncertainty, the authors employ it here to integrate reliability into the metric being optimized for improved boundary and local segmentation. Additionally, they incorporate dropout samples into contrastive learning.
- 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 proposed method seems closely related to the TEGDA paper, with the only new aspect being the use of SICE in defining the metrics for optimization.
- 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.
- 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 authors did not explain how the proposed method differs from existing methods, such as TEGDA, which appears ideologically similar to this work.
How does LRC enhance the outcome beyond just lowering the model’s computational cost?
- 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?
Limited novelty: the proposed method seems closely related to the TEGDA model, with the only new aspect being the use of SICE in defining the metrics to be optimized.
- 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.
While the proposed method shares similarities with TEGDA, it also remains an application to the PBS. However, it provides a theoretical formulation and an evaluation of 3 datasets (sub-datasets). The proposed work will find a place at a satellite event (workshop).
Review #3
- Please describe the contribution of the paper
- This paper aims to improve the segmentation reliability and anatomical consistency in test-time adaptation for pelvic bone segmentation (PBS). To this end, the authors proposed a new closed loop framework (ReGA) that incorporates a new consistency-based testing metric (SICE) and built a feature refinement model based on this metric (TAFR).
- 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 central ideas of the paper are well motivated in the introduction, and the proposed method have attempted to address all the limitations (data privacy, boundary degradation in TTA, not utilizing self-supervised potentials) in the model and pipeline design
- The method is clearly defined mathematically
- The authors have open sourced their code
- In the comparison section, the authors not only included other TTA models but also included benchmarks for source only and target only metrics, giving a clear picture of the expected range of results. This provides a good quantitative context to the readers as to why TTA is crucial in this particular application.
- 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.
- Minor concern on the dataset size: The number of 3D CT volumes used for training seems to be rather small. In the context of TTA, I’m not entirely sure whether this is intentionally designed to model domain shift across institutions (in which case this should be fine). Hopefully future work can include additional datasets and larger data pools if possible.
- 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.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall this is a well motivated and well organized paper, with clear maths defining the method.
- 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 concern has been sufficiently addressed
Author Feedback
R1Q1, Q3, Q4: Terminology, Notation, and Technical Details. We have added relevant citations and refined the statements. Specifically, the teacher-student schemes refer to TTA methods where an EMA-updated teacher generates pseudo-labels for supervising the student on target domain data. We clarify that θᵗ denotes the source-trained PBS model parameters, while f(θᵗ) denotes the corresponding segmentation model. Ambiguous expressions are corrected, and θᵗ is now defined at its first occurrence in Sec. 2.During TTA, dropout is kept active, and each target sample is forwarded K times to obtain stochastic predictions P̂ᵢₖ, where k indexes the stochastic forward pass. Following [7], we analyze K=1,5,10,15 and set K=5 in all experiments. R1Q2: Evidence and Literature Support for Claims. The prior TS schemes with pseudo-label supervision have shown that unreliable predictions can be propagated and accumulated under domain shifts in the cited works [15,25]. Table 1 shows that these methods suffer from poor HD95, indicating notable boundary delineation errors. We apologize for missing citations and included works (e.g., AdaMI [Bateson et al., 2022] and EDCP [Liu et al., 2026]) showing that enforcing anatomical consistency is crucial for reliable adaptation. Cited works [14,16] and TestFit [Zhang et al., 2025] show that self-supervised learning has been widely explored in TTA, but its explicit use for enforcing target-domain anatomical consistency remains underexplored. R1Q5, Q6, Q7: Strengthening Experimental Evidence. We agree that ReGA is positioned as an application-focused TTA framework for PBS. PBS is sensitive to boundary ambiguity and anatomical inconsistency under domain shifts; ReGA is designed to suppress unreliable boundary predictions and to promote anatomical consistency across pelvic regions. Further, in a new MSD_T10 → CLINIC setting, ReGA is also superior to source-only models and other TTA baselines, i.e., TENT, InTEnt, VPTTA, CoTTA, and TEGDA. Specifically, it improves Dice by 8.8 over the source-only model while reducing HD95/ASD by 18.8/4.7 and outperforms other TTA baselines by 3.3 - 4.8 Dice, with 16.8 - 26.7/3.5 - 4.1 reductions in HD95/ASD. Paired Wilcoxon signed-rank tests show that ReGA improves Dice (p≤0.008), HD95 (p≤0.001), and ASD (p≤0.002) over TTA baselines in CLINIC → KITS19, and achieves notable HD95/ASD gains (p≤0.05) with comparable Dice in CLINIC → MSD_T10. We conduct sensitivity analysis on key hyperparameters. ReGA is robust to M when M>10, with Dice variations within 0.3.K=5 is selected as significant computation overhead would occur for K>5 with a marginal gain of 0.1 in Dice. ReGA performs best at β=0.01 (range 0.005 - 0.020), other values degrade Dice minimally (0.6 max). Dropout rate of 0.5 (range 0.1 – 1.0) provides a balanced stochastic perturbation, yet the worst dropout rate results in only 0.6 degradation in Dice. R2Q1, Q2: Methodological Distinctions. While ReGA is partially inspired by the general adaptation philosophy of TEGDA, it targets a different task, i.e., TTA of PBS models, where boundary ambiguity and anatomical consistency are critical under domain shifts. To this end, we introduce a new metric SICE for optimization that jointly considers region overlap and boundary deviation and further uses Lrc for confidence- and SICE-weighted RCL. Thus, Lrc not only reduces computation but also promotes reliable anatomical-region feature. In Table 2, the variant w/o Lrc performs worse, especially in HD95, indicating the contribution of Lrc to preserving anatomical integrity. R3Q1: Choice of Dataset. TTA aims to adapt a source pre-trained model to the test data during inference in a source-free and online manner. Following this, the PBS model is trained on limited labeled 3D CT volumes from one institution and then adapted to unlabeled data from another institution. In the future, we aim to build a larger standardized multi-center 3D pelvic CT datasets.
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 mixed evaluations of this paper (scores: 3, 4, 5), with strengths noted in the problem motivation, methodological formulation, and empirical evaluation, but also concerns regarding limited novelty, clarity of presentation, and depth of experimental analysis . In particular, reviewers highlight missing statistical validation, insufficient analysis of hyperparameters and sensitivity, and unclear differentiation from closely related prior work. Given these differences, I invite the authors to address these concerns in the rebuttal—especially by clarifying methodological distinctions, strengthening experimental evidence, and improving clarity where needed. A clear and comprehensive 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 an important and clinically relevant problem in test-time adaptation for pelvic bone segmentation under domain shift. The proposed ReGA framework, including the SICE-based reliability estimation and feature refinement strategy, is well motivated and evaluated on multiple datasets. While concerns remain regarding the degree of novelty relative to existing TTA/SFDA methods such as TEGDA, the majority of reviewers considered the rebuttal sufficient and supported acceptance. The work is clearly presented, provides useful empirical validation, and offers a practical contribution for improving segmentation reliability and anatomical consistency in unseen clinical domains. Therefore, I recommend acceptance, with the expectation that the authors will further clarify the distinction from closely related methods, add missing statistical and hyperparameter analyses where possible, and temper overstated novelty claims 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.
The authors have addressed the reviewers’ comments well. Two reviewers acknowledged the novelty and significance of the work. Therefore acceptance of the paper is concluded.
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.
I recommend accept: post-rebuttal R1 raised to Accept and R3 maintained Accept (recommending Highlighted Poster), outweighing R2’s downgrade to Reject on TEGDA-similarity and workshop-venue assessments.
