Pith. sign in

REVIEW 3 major objections 6 minor 24 references

Test-time Adaptation of Pelvic Bone Segmentation Models via Dynamic Reliability-Guided

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes ReGA, a test-time adaptation framework for pelvic bone CT segmentation that steers feature refinement, contrastive learning, and teacher-student updates with a pseudo-label reliability score, and reports higher Dice and l

desk verdict The SICE reliability gating is a sensible new piece for TTA in pelvic bone segmentation, and the KITS19 gain looks real; but the MSD T10 gain is noise, so 'consistently outperforms' is not established. read the letter →

arxiv 2608.00510 v1 pith:HVHVOY7Y submitted 2026-08-01 cs.CV

Ling Ren , Chao Deng , Ziming Wang , Yuecong Xu , Kai Zheng This is my paper
classification cs.CV
keywords test-timeadaptationpelvicbonesegmentation3DCTpseudo-labelreliabilitydomainshiftcontrastivelearningteacher-studentboundaryconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ReGA is a test-time adaptation method for pelvic bone segmentation in CT. It lets a model trained in one hospital be updated online on unlabeled scans from a new hospital, without source data, by scoring the reliability of its own pseudo-labels. The reliability score, SICE, combines region overlap and boundary deviation across dropout-sampled predictions, and this score then gates which features are stored, how strongly a sample's features are refined, how fast the teacher model updates, and how much each pseudo-label contributes to the loss. On three heterogeneous pelvic CT datasets, ReGA reports mean Dice of 32.1 versus the best prior TTA baseline's 24.9 in one cross-dataset setting, and 26.4 versus 25.8 in another, with lower surface-distance errors. The paper's point is that a single closed-loop reliability signal can make online adaptation stable enough for clinical deployment.

What carries the argument

The machinery is SICE, Segmentation Inference Consistency Evaluation, a score that measures how much a test sample's prediction agrees with its own dropout-sampled predictions in both region overlap and boundary position, then applies an entropy-based calibration. SICE carries the argument because it controls four gates: feature-bank admission, feature-refinement weighting, teacher EMA rate, and pseudo-label loss weighting.

What would settle it

On a held-out target subset with ground truth, plot SICE against true Dice per test volume: if the ordering breaks, meaning high-SICE volumes are not the accurate ones or low-SICE volumes are, the feature-bank gate and loss weighting are misdirected. A second check: run ReGA with an oracle quality score in place of SICE; if performance does not improve over the SICE version, the reported gain is not coming from reliability estimation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that pseudo-label reliability can be measured at test time from the model's own dropout ensemble and then used to close the loop on every part of adaptation. SICE is computed as a weighted combination of volumetric consistency, meaning Dice-like overlap between the main prediction and K dropout predictions, and boundary discrepancy, meaning percentile distances between edge maps extracted from those predictions, calibrated by predictive entropy. A trust-weighted adaptive feature refinement module stores high-SICE features in a first-in-first-out bank and fuses them with the current sample, weighted by SICE and cosine similarity. A region-lev

Load-bearing premise

The load-bearing premise is that the paper's reliability score tracks true segmentation quality on the new domain; if that correlation fails, harmful pseudo-labels enter the feature bank and the training loss is weighted incorrectly, and the paper itself notes its boundary-extraction step needs further work.

Editorial extensions

If this is right

  • A deployed pelvic bone segmentation model can keep improving while running on a new hospital's routine CTs, without labels or source data transfer.
  • Boundary-sensitive reliability scoring translates into lower Hausdorff distances, meaning cleaner separation between adjacent pelvic bones for surgical planning.
  • A single reliability score can coordinate feature-bank memory, feature fusion, model averaging, and loss reweighting, one quality signal driving multiple mechanisms.
  • Teacher-student adaptation with an adaptive EMA rate based on sample reliability is a general pattern for class-imbalanced 3D segmentation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If SICE tracks true accuracy, the closed-loop design should transfer to other boundary-ambiguous segmentation tasks, such as other bone or organ boundaries under domain shift; this is an extension the paper does not test.
  • A direct SICE-versus-true-Dice calibration curve on a validation partition would separate whether the score itself works or whether the surrounding machinery is responsible for the gains; the paper's ablations only vary the framework internally.
  • The static choices, feature-bank length, percentile admission threshold, and number of dropout passes, could be made adaptive over the test stream, and a self-tuning version might behave differently in longer deployments.
  • Because SICE is computed on a central patch of each volume, multi-scale or full-volume scoring could be a more reliable gate; the paper does not compare patch-selection strategies.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes ReGA, a test-time adaptation (TTA) framework for pelvic bone segmentation from CT. ReGA introduces a pseudo-label reliability score, SICE, computed from dropout-based ensemble predictions using volumetric and boundary consistency (Eqs. 1–5). SICE is then used to gate a feature bank and refine current features (Eq. 6), to weight region-level contrastive learning (Eq. 7–8), to set the teacher EMA rate (Eq. 9), and to weight the total loss (Eq. 10). Experiments transfer a source model trained on CLINIC to KITS19 and MSD T10, comparing with five TTA baselines. The central claim is that ReGA consistently outperforms state-of-the-art TTA methods, with the strongest result on CLINIC→KITS19 (32.1 vs 24.9 mean Dice) and a much smaller gain on CLINIC→MSD T10 (26.4 vs 25.8).

Significance. If validated, ReGA would be the first TTA method specifically designed for pelvic bone segmentation and introduces a novel reliability criterion (SICE) that is cheaply computable from dropout passes. The code is made publicly available, which supports reproducibility. However, the current evidence is not yet sufficient to establish the headline comparative claim: the gain on one of the two transfer settings is small relative to the reported per-case variance, and the pivotal SICE score is never calibrated against ground truth or an external proxy. The framework is coherent and the ablation study is a useful first step, but the missing inferential statistics and the unverified reliability proxy are load-bearing for the paper's conclusions.

major comments (3)
  1. [§3, Table 1; Abstract] The claim that ReGA 'consistently outperforms state-of-the-art TTA methods' is not statistically supported. In CLINIC→MSD T10, the reported gain over the best baseline is 26.4±12.0 vs 25.8±10.3, a margin of 0.6 Dice with overlapping standard deviations; a two-sample comparison yields a z-score around 0.5. No paired significance tests, confidence intervals, or multi-seed runs are reported anywhere in the paper. Since the central contribution is comparative, the authors should provide paired tests (e.g., Wilcoxon signed-rank) and ideally repeated runs or bootstrap confidence intervals for both transfer settings. Additionally, on CLINIC→KITS19 ReGA's HD95 (160.4±42.8) is worse than source-only (112.7±20.8); this unexpected trade-off is not discussed and further complicates the 'consistently outperforms' statement.
  2. [§2.1, Eqs. (1)-(5); §2.2–2.4, Eqs. (6), (9), (10); Table 2] SICE is the load-bearing reliability signal: it controls feature-bank admission, feature refinement, teacher EMA rate, and both loss weights. Yet the paper provides no direct evidence that SICE tracks pseudo-label quality in the target domain. The only supporting ablation is w/oE (entropy instead of SICE), which compares two internal heuristics and does not establish correlation with true segmentation accuracy. The authors should add a validation experiment on a held-out partition (or on source-domain samples with synthetic shifts) showing, for example, a SICE-vs-true-Dice curve or a correlation coefficient, and demonstrate that the τ-based admission and SICE-based weighting behave sensibly across that curve. Without such evidence, the reported gains could in principle arise from an incidental property of SICE rather than from reliable pseudo-labels.
  3. [§2.1–§2.4; §3, implementation details] The method introduces several hyperparameters (λ1, λ2, α in Eq. 4; temperature T in Eq. 8; τ for bank admission; β for contrastive loss; dropout rate, K, M) but reports no sensitivity analysis or ablation over the most critical ones. In particular, τ and β directly control how much target information enters the adaptation signal, and λ1/λ2 determine the reliability score itself. Given that the improvements in Table 2 are small in several comparisons (e.g., w/oLre 32.0 vs 32.1 Dice), the absence of a sensitivity study makes it difficult to assess robustness. Please report a sweep over at least τ, β, and the λ weights, and state how these values were chosen.
minor comments (6)
  1. [§2.3, before Eq. (7)] Typo: 'SICE sore' should be 'SICE score'.
  2. [§2.1, Eq. (3)] The phrase 'the revise direction' is unclear; use 'the reverse direction' or 'the opposite direction'.
  3. [§2.1, Fig. 2 and text] The 'central patch' xt_ip is not quantitatively defined. Its size and how it is cropped from the full volume should be stated, as it affects the SICE computation.
  4. [§2.1, Fig. 2 and Eq. (3)] The Canny operator parameters (thresholds, sigma) are not given. Since boundary extraction is a component of SICE, these details should be provided for reproducibility. The conclusion also notes boundary extraction 'needs further work', which should be reflected as a limitation in the main text.
  5. [§3, Table 1] The paper reports per-case standard deviations but does not distinguish between per-case variance and run-to-run variance. Please clarify that the ± values are over test cases, not over adaptation runs, and consider reporting metrics over multiple random seeds.
  6. [§2.1, Eq. (5)] The notation logC for maximum entropy is used without defining C. Define C as the number of classes or the maximum entropy constant explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ReGA's guided self-training is evaluated against external ground truth and baselines.

full rationale

The paper's derivation chain is a self-training/test-time adaptation objective, not a self-referential prediction. SICE (Eqs. 1-5) is an internal consistency heuristic computed from the model's own dropout predictions; it gates feature-bank admission (Eq. 6), the teacher EMA rate (Eq. 9), and loss weights (Eq. 10). This is a feedback loop, but not circular in the prohibited sense: the reported improvements in Dice, HD95, and ASD are measured on target-domain ground truth (Table 1) and compared against external baselines (TENT, CoTTA, etc.). No parameter is fit to the target metric and then reported as a prediction of that same metric. The paper contains no load-bearing self-citations: the supporting citations (e.g., [7,25]) are external prior work on dropout uncertainty, and no uniqueness or ansatz is imported from the authors' own prior publications. The lack of direct SICE-versus-true-Dice calibration is a legitimate validity concern, and the small CLINIC→MSD T10 margin raises statistical-significance questions, but these are correctness risks, not circularity. The boundary-extraction limitation noted in the conclusion is an acknowledged weakness, not a circular step. Therefore, the central claim rests on independent empirical evaluation, and no circular step can be exhibited from the paper's own equations.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim rests on one new heuristic (SICE) plus a set of hand-set hyperparameters, listed above. The supporting axioms are standard TTA premises borrowed from cited work: dropout variance correlates with error, features from confident samples transfer, centroid separation preserves structure, and mean-teacher averaging stabilizes updates. No new physical entities are introduced; SICE is a new composite metric validated only through the pipeline's end-task performance.

free parameters (8)
  • lambda1, lambda2 (SICE region/boundary weights, Eq 4) = not reported
    Balance volumetric consistency against boundary deviation in the reliability score; values are absent from the implementation details.
  • alpha (boundary deviation sensitivity, Eq 4) = not reported
    Scales exp(-Sb/alpha); no value or sensitivity analysis is given.
  • Temperature T (Eq 8) = not reported
    Scales cosine similarity in the region-level contrastive loss; omitted from implementation details.
  • tau (feature bank admission percentile) = 70
    SICE threshold for inserting features into the bank; hand-set with no sensitivity analysis.
  • gamma (boundary distance percentile) = 95
    Robust Hausdorff percentile used in Sb computation; hand-set.
  • beta (contrastive loss weight) = 0.01
    Tradeoff weight for Lrc in Eq (10); hand-set with no sensitivity analysis.
  • K, M, dropout rate = K=5, M=10, dropout rate=0.5
    Number of dropout passes, feature bank length, and dropout rate; reported but no sensitivity analysis is provided.
  • w(.) normalization in Eq (6) = unspecified
    Maps SICE to the feature blend weight; the exact normalization operation is never defined.
assumptions (5)
  • domain assumption MC dropout prediction variance correlates with segmentation error on target-domain data
    Borrowed from [7,25]; the entire SICE reliability estimate rests on this correlation. Invoked in Sec 2.1.
  • ad hoc to paper SICE is a valid uncalibrated proxy for pseudo-label quality
    SICE gates feature bank, EMA rate, and losses (Eqs 6, 9, 10). No external calibration or correlation analysis against true Dice is provided.
  • domain assumption Canny edge detection yields meaningful boundaries of predicted segmentation maps for boundary-consistency scoring
    Used in Eq (3) to compute Sb; fragile on noisy soft probability maps, as the conclusion partially concedes.
  • domain assumption Class centroids of sacrum, hips, and lumbar spine remain separable and stable enough under domain shift for centroid contrastive training
    Sec 2.3; the RCL loss only repels centroids (numerator in Eq 8 is self-similarity), so consistency gains rely on centroid stability.
  • domain assumption Mean-teacher EMA with the SICE-adaptive rate (Eq 9) preserves stability of the student
    Standard EMA stability argument [14,15] is extended with a variable rate; no stability analysis is given.
invented entities (1)
  • SICE score (Segmentation Inference Consistency Evaluation)
    purpose: Reliability criterion that gates sample selection, feature refinement, EMA rate, and loss weighting (Eqs 4, 6, 9, 10)
    It is a new composite metric with no external calibration: the paper does not validate SICE against true Dice on held-out data, and its only support is the internal w/oE ablation. The main experiments validate the end task, not the metric itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Test-time Adaptation of Pelvic Bone Segmentation Models via Dynamic Reliability-Guided." pith.science (2026). https://pith.science/paper/HVHVOY7Y

@misc{pith2026260800510,
  author       = {Pith},
  title        = {Pith review of: Test-time Adaptation of Pelvic Bone Segmentation Models via Dynamic Reliability-Guided},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HVHVOY7Y}},
  note         = {Machine review of arXiv:2608.00510}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2608.00510 by the authors.

Figure 1
Figure 1. Overall process of the proposed ReGA. During adaptation, the features of a testing sample x t i extracted by the encoder of the PBS model θ t (initialized at the training institute) are refined by TAFR (see [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CSCS and TAFR module. In CSCS, the test sample x t i is centrally cropped to obtain x t ip , which is passed through θ t with K-times dropout for multiple inferences. These inferences, along with the edges extracted using the Canny operator, are combined to compute SICE. In TAFR, latent features of high SICE are dynamically accumulated into a feature bank M. The current feature is then adaptively ref… view at source ↗
Figure 3
Figure 3. Visualization of segmentation results for different TTA methods. that Sb of the SICE score is removed, while w/o E denotes that entropy is used in place of SICE. Ablation results across the three datasets consistently show that each module incrementally enhances segmentation performance. From the obvious improvement in HD95 and ASD compared to the variant w/o B and w/o E, the proposed SICE can successfully evaluate … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages

  1. [1]

    Medical Image Analysis101(103440) (2025)

    Cai,Z.,Xin,J.,You,C.,Shi,P.,Dong,S.,Dvornek,N.,Zheng,N.,Duncan,J.:Style mixup enhanced disentanglement learning for unsupervised domain adaptation in medical image segmentation. Medical Image Analysis101(103440) (2025)

    work page 2025

  2. [2]

    In: CVPR

    Chen, Z., Pan, Y., Ye, Y., Lu, M., Xia, Y.: Each test image deserves a specific prompt: continual test-time adaptation for 2d medical image segmentation. In: CVPR. pp. 11184–11193 (2024)

    work page 2024

  3. [3]

    Dong, H., Konz, N., Gu, H., Mazurowski, M.: Medical image segmentation with in- tent:Integratedentropyweightingforsingleimagetest-timeadaptation.In:CVPR. pp. 5046–5055 (2024)

    work page 2024

  4. [4]

    In: ICML

    Gal,Y.,Ghahramani,Z.:Dropoutasabayesianapproximation:representingmodel uncertainty in deep learning. In: ICML. pp. 1050–1059. PMLR (2016)

    work page 2016

  5. [5]

    arXiv preprint arXiv:1904.00445 (2019)

    Heller, N., et al.: The kits19 challenge data: 300 kidney tumor cases with clin- ical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445 (2019)

    arXiv 1904

  6. [6]

    Nature methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

    2021

  7. [7]

    In: CVPR

    Lee, T., Chottananurak, S., Gong, T., Lee, S.: Aetta: label-free accuracy estimation for test-time adaptation. In: CVPR. pp. 28643–28652 (2024)

    work page 2024

  8. [8]

    In: International conference on machine learning

    Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hy- pothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028–6039. PMLR (2020)

    2020

Show all 24 references
  1. [9]

    In: CVPR

    Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty esti- mation for source-free unsupervised domain adaptation. In: CVPR. pp. 7640–7650 (2023)

    2023

  2. [10]

    IEEE Transactions on Medical Imaging44(1), 79– 91 (2025)

    Liu, J., Li, H., Zeng, B., Wang, H., Kikinis, R., Joskowicz, L., Chen, X.: An end-to- end geometry-based pipeline for automatic preoperative surgical planning of pelvic fracture reduction and fixation. IEEE Transactions on Medical Imaging44(1), 79– 91 (2025)

    2025

  3. [11]

    IJCARS16(5), 749–756 (2021)

    Liu, P., et al.: Deep learning to segment pelvic bones: large-scale ct datasets and baseline models. IJCARS16(5), 749–756 (2021)

    2021

  4. [12]

    Medical Image Analysis 102(103506) (2025)

    Liu, Y., et al.: Preoperative fracture reduction planning for image-guided pelvic trauma surgery: a comprehensive pipeline with learning. Medical Image Analysis 102(103506) (2025)

    2025

  5. [13]

    arXiv preprint arXiv:1902.09063 (2019)

    Simpson, A., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)

    1902 arXiv

  6. [14]

    arXiv preprint arXiv:2006.10726 (2020)

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020)

    2006 arXiv

  7. [15]

    In: CVPR

    Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: CVPR. pp. 7201–7211 (2022) 10 L. Ren et al

    2022

  8. [17]

    IEEE Transactions on Medical Imaging 43(9), 3098–3109 (2024)

    Wu, J., et al.: Fpl+: Filtered pseudo label-based unsupervised cross-modality adap- tation for 3d medical image segmentation. IEEE Transactions on Medical Imaging 43(9), 3098–3109 (2024)

    2024

  9. [18]

    Yang, H., Chen, C., Jiang, M., Liu, Q., Cao, J., Heng, P., Dou, Q.: Dltta: dy- namiclearningratefortest-timeadaptationoncross-domainmedicalimages.IEEE Transactions on Medical Imaging41(12), 3575–3586 (2022)

    2022

  10. [19]

    IEEE Transactions on Medical Imaging44(7), 2973–2988 (2025)

    Yang, Y., Zhuang, J., Sun, G., Wang, R., Su, J.: Boundary-guided contrastive learning for semi-supervised medical image segmentation. IEEE Transactions on Medical Imaging44(7), 2973–2988 (2025)

    2025

  11. [20]

    In: MICCAI 2023

    Yu, Q., Xi, N., Yuan, J., Zhou, Z., Dang, K., Ding, X.: Source-free domain adap- tation for medical image segmentation via prototype-anchored feature alignment and contrastive learning. In: MICCAI 2023. LNCS, vol. 14226, pp. 1–12. Springer (2023). https://doi.org/10.1007/978-3...

    2023 doi

  12. [21]

    IEEE Journal of Biomedical and Health Informatics29(12), 9060–9072 (2025)

    Zhang, G., Qi, X., Wu, J., Yan, B., Wang, G.: Iplc+: Sam-guided iterative pseudo label correction for source-free domain adaptation in medical image segmentation. IEEE Journal of Biomedical and Health Informatics29(12), 9060–9072 (2025)

    2025

  13. [22]

    In: CVPR

    Zhang, X., et al.: Mapseg: Unified unsupervised domain adaptation for heteroge- neous medical image segmentation based on 3d masked autoencoding and pseudo- labeling. In: CVPR. pp. 5851–5862 (2024)

    2024

  14. [23]

    In: MICCAI 2023

    Zhang, Y., Huang, K., Chen, C., Chen, Q., Heng, P.: Satta: semantic-aware test- time adaptation for cross-domain medical image segmentation. In: MICCAI 2023. LNCS, vol. 14221, pp. 160–171. Springer (2023). https://doi.org/10.1007/978-3- 031-43895-0_14

    2023 doi

  15. [24]

    Medical Image Analysis97(103275) (2024)

    Zheng, B., et al.: Dual domain distribution disruption with semantics preservation: Unsupervised domain adaptation for medical image segmentation. Medical Image Analysis97(103275) (2024)

    2024

  16. [25]

    In: MICCAI 2025

    Zhou, Y., Wu, J., Liao, W., Zhang, S., Zhang, S., Wang, G.: Tegda: test-time evaluation-guided dynamic adaptation for medical image segmenta- tion. In: MICCAI 2025. LNCS, vol. 15965, pp. 628–637. Springer (2026). https://doi.org/10.1007/978-3-032-04978-0_60

    2025 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.