Pith. sign in

REVIEW 2 major objections 6 minor 18 references

Exploring Test Time Adaptation for Subcortical Segmentation of the Fetal Brain in 3D Ultrasound

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Test-time adaptation lets a pretrained fetal-brain segmentation model adapt to new ultrasound scans without manual labels, and adding a normative-atlas volume prior makes the adaptation more effective.

desk verdict Useful TTA benchmark for fetal US segmentation, but the EntropyKL claim is undermined by an evaluation that shares the same atlas as its prior. read the letter →

arxiv 2502.08774 v1 pith:C6IONDZT submitted 2025-02-12 cs.CV cs.AIcs.LG

Joshua Omolegan , Pak Hei Yeung , Madeleine K. Wyburd , Linde Hesse , Monique Haak , Intergrowth-21st Consortium , Ana I. L. Namburete , Nicola K. Dinsdale This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords TestTimeAdaptationUltrasoundSegmentationFetalbrainSubcorticalstructuresEntropyminimizationNormativeatlasDomainshift
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

The paper asks whether a segmentation model trained on one set of 3D fetal-brain ultrasound scans can be adapted at test time to new scans that look different, without any manual labels for the new data. It shows that test-time entropy minimization improves Dice scores over the unadapted model under simulated rotations, scaling, smoothing, and gamma changes, and also across gestational weeks and unseen scanner vendors. The paper's proposed EntropyKL method adds a KL-divergence term that penalizes deviation of predicted class proportions from a normative fetal-brain atlas, and reports that this prior yields the most consistent segmentations, especially on unseen scanners. If correct, this would let automated subcortical segmentation models be deployed to new acquisition setups without retraining or re-annotation.

What carries the argument

The central objects are the entropy-minimization objective of TENT, which updates only batch-normalization parameters to reduce prediction entropy on target data, and the proposed EntropyKL extension, which appends a KL divergence between the predicted class proportion $\hat{\tau}_t$ and the atlas proportion $\tau_t$ from the normative fetal-brain atlas, with weight $\lambda$. A second variant, LayerInspect, selects the $m$ layers whose activation magnitudes differ most between source and target using a Taylor approximation, and updates only those layers. The atlas prior is the load-bearing addition: it supplies an expected volume ratio for the four subcortical classes and prevents adaptation from collapsing to trivial solutions, while still allowing the model to learn locations.

What would settle it

Evaluate EntropyKL on the unseen scanner volumes using manual annotations, or on the simulated-shift volumes using manual annotations instead of atlas-propagated labels; if the Dice advantage over TENT shrinks or disappears, the atlas prior is not contributing independent anatomical information.

Watch

Extended reading notes

Core claim

The central claim is that test-time adaptation can overcome domain shift in 3D fetal ultrasound segmentation, and that the proposed EntropyKL objective — TENT's entropy loss plus $\lambda \, \mathrm{KL}(\hat{\tau}_t \| \tau_t)$, where $\hat{\tau}_t$ is the average predicted class probability and $\tau_t$ is the class proportion from a normative atlas — adapts a frozen pretrained 3D U-Net better than TENT or LayerInspect alone. The authors report that all TTA methods significantly outperform the source model on simulated domain shifts, that TTA improves performance at gestational ages far from the training distribution, and that EntropyKL gives the most anatomically plausible segmentations on unseen scanner data, where no manual labels exist.

Load-bearing premise

The evaluation of EntropyKL depends on the assumption that the atlas-propagated annotations used as ground truth in the simulated and gestational-age experiments are valid and independent of the same atlas that provides the EntropyKL prior; if they are not independent, the reported Dice improvements for EntropyKL could be inflated.

Editorial extensions

If this is right

  • TTA with entropy minimization significantly improves Dice over the unadapted source model for simulated ultrasound domain shifts such as rotations, scaling, smoothing, and gamma correction.
  • Adapting per single volume rather than per batch yields better results, matching the observation that each freehand ultrasound scan effectively forms its own domain.
  • EntropyKL, with the atlas class-ratio prior, gives the most consistent segmentations among the TTA methods, including on data from two unseen scanner vendors where only qualitative evaluation is possible.
  • For gestational-age shifts, TTA improves Dice by about 0.1 on the earliest weeks where the source model performs worst, and never reduces performance.
  • Histogram matching, a standard preprocessing fix, fails on morphometric shifts such as gestational-age differences, whereas TTA handles them.

Reading between the lines

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

  • If the same atlas is doing double duty as both prior and evaluation label, then the reported Dice gains for EntropyKL over TENT may partly reflect label leakage; an independent manual-annotation evaluation would settle how much of the gain is real.
  • The authors note the atlas encodes healthy anatomy; a direct extension would test EntropyKL on fetuses with structural anomalies, where the volume prior will compete with the entropy signal and may mask pathology.
  • The single-volume adaptation result implies a practical deployment recipe: adapt each incoming freehand scan before inference, a protocol that could be validated prospectively without changing the acquisition workflow.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper investigates test-time adaptation (TTA) for subcortical segmentation of the fetal brain in 3D ultrasound. Using a pretrained 3D UNet as the source model, the authors evaluate TENT, a proposed layer-selection method (LayerInspect), and a proposed atlas-prior method (EntropyKL) under simulated domain shifts, gestational-age variation, and unseen scanner data. They report that TTA methods improve over the source model, with EntropyKL often performing best, and they release their code.

Significance. If the results hold, the paper would demonstrate a practical approach to adapting fetal-brain ultrasound segmentation models to new scans or scanners without manual labels, which is valuable given the acquisition variability in freehand ultrasound. The code release and the use of a normative atlas as a prior are constructive contributions. However, the central quantitative evidence for EntropyKL is compromised by a likely circularity between the atlas used as the prior and the atlas-propagated labels used for evaluation, and the unseen-scanner evidence is qualitative and selected. The paper's broader claim that TTA improves over the source model is plausible and supported by paired tests, but the specific novelty (EntropyKL) needs stronger, independent validation.

major comments (2)
  1. [Section III (Simulated Domain Shifts and Gestational Age) and Eq. (3)] The quantitative evaluation of EntropyKL is confounded by the shared origin of the atlas prior and the reference labels. The KL term in Eq. (3) penalizes deviations of the predicted class proportions from the atlas [11] ratios, while the Dice scores for the simulated-shift and gestational-age experiments are computed against 'propagated annotations from an atlas [6], [11]' (Section III). If those reference labels are produced by warping the same atlas segmentation to each volume, then the ground-truth class proportions are the very proportions EntropyKL is regularized toward; TENT and LayerInspect receive no such information. This can inflate EntropyKL's measured Dice improvement independently of anatomical correctness. Please clarify the provenance of the propagated annotations; if they derive from the same atlas, either evaluate on manual annotations, or demonstrate that the prior remains beneficial when the reference class ratios are deliberately varied, or restrict the claim of EntropyKL's advantage to the qualitative unseen-scanner results.
  2. [Section IV (Unseen Datasets) and Fig. 5] The claim that EntropyKL gives the best segmentations on unseen scanners is not quantitatively supported. The comparison is qualitative, shows only one slice per scanner, and the slices are selected as the highest-entropy predictions of the base model, which may preferentially display the effect of entropy minimization. Manual segmentations or a blinded reader study would be required to substantiate the advantage of EntropyKL over the other TTA methods in this setting. Without this, the abstract's statement that EntropyKL demonstrates improvements across real domain shifts rests on the simulated experiments, which are subject to the circularity concern above.
minor comments (6)
  1. [Section III, Simulated Domain Shifts] The phrase 'propagated annotations from an atlas [6], [11]' is ambiguous; specify which atlas (e.g., [11]) and whether the same atlas is used in the EntropyKL prior.
  2. [Section IV, Gestational Age] The text states that 'TTA methods significantly increases the performance' but reports no test statistic or p-value for this figure; please add error bars and significance tests for the gestational-age comparisons.
  3. [Fig. 4 caption] The caption says the right-hand panel shows 'The same plane from the US atlas [11] at different ages'; indicate whether these slices are the source of the propagated annotations used as ground truth.
  4. [Section V, Discussion] The phrase 'normative altas' is a typo for 'normative atlas'.
  5. [Eq. (2)] Define N, l, n, and the notation used for the Taylor approximation; also clarify that the activations are L2-normalized before the difference is computed.
  6. [Section III, Implementation] The hyperparameter choices (λ=1, m=1, number of backward passes) are selected based on sensitivity experiments in Fig. 3; state whether these selections were made on a validation set separate from the test data.

Circularity Check

1 steps flagged · score 6.0 of 10

EntropyKL is evaluated against atlas-propagated labels from the same [11] atlas that supplies its class-ratio prior, so the reported quantitative advantage is partly built in.

  1. self definitional [Section III (Simulated Domain Shifts; Gestational Age), with the prior defined in Section II-C, Eq. (3)]
    "We used 59 pre-aligned volumes at 21 gestational weeks from the INTERGROWTH-21st study [12], with propagated annotations from an atlas [6], [11]. ... This gave a dataset of 529 labelled images with atlas-propagated annotations."

    The EntropyKL loss in Eq. (3) is L_TENT + lambda KL(tau_hat || tau_t), where tau_t is the proportion of the atlas that was class k, taken from the normative atlas [11]. The simulated-shift and gestational-age Dice evaluations use 'propagated annotations from an atlas [6], [11]' as the reference labels. Thus the reference standard is produced from the same atlas that defines the prior, while the TENT and LayerInspect baselines are not regularized toward that atlas. EntropyKL is therefore penalized for deviating from the class proportions of the very labels against which it is scored, giving it a built-in statistical advantage that can inflate its measured Dice improvement. The unseen-scanner experiments are only qualitative, so they do not independently break this circularity.

full rationale

The central novel claim is that EntropyKL improves test-time adaptation for fetal subcortical segmentation. The decisive quantitative evidence for this claim is compromised: Section II-C defines the atlas prior using [11], and Section III evaluates the method on 'atlas-propagated annotations' citing the same [11] atlas. Because the ground-truth labels inherit the class-ratio statistics of the atlas, and EntropyKL explicitly penalizes deviations from those ratios, the comparison against TENT and LayerInspect is biased in EntropyKL's favor. This is not a full identity reduction, since Dice also depends on spatial overlap and the KL term only constrains global class proportions, so the score is set at 6 rather than higher. The remaining components of the paper, including TENT and LayerInspect baselines, are not circular; the issue is localized to the evaluation of the proposed method against labels derived from its own prior.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central quantitative claims rest on atlas-propagated labels, a fixed source model, and several hyperparameters chosen on the test set. The prior and the ground truth share the same atlas, which creates the main circularity burden. No new physical or conceptual entities are postulated.

free parameters (5)
  • KL weight lambda = 1
    Set to 1 based on the Figure 3 sensitivity experiment; controls the strength of the atlas prior in L_EntropyKL.
  • number of adaptation backward passes = 25 for unseen datasets, 1 for others
    Chosen per experiment as the smallest value with a substantial observed change on the target/test data.
  • LayerInspect number of updated layers m = 1
    Selected from the Figure 3(b) sensitivity experiment.
  • LayerInspect learning rate = 1e-4
    Set lower than the TENT and EntropyKL learning rate of 1e-3, presumably to keep layer updates stable.
  • batch size = 2
    Chosen due to memory constraints; affects batch normalization statistics during adaptation.
assumptions (6)
  • domain assumption Atlas-propagated annotations are valid ground truth for subcortical segmentation in the simulated and gestational-age experiments.
    Quantitative Dice scores use labels propagated from the INTERGROWTH atlas [6],[11]; no independent manual verification is reported for these test subjects.
  • domain assumption Entropy minimization at test time improves segmentation under domain shift.
    Adopted from TENT [16] and prior medical imaging TTA [15]; it is the baseline adaptation mechanism.
  • domain assumption Updating only batch normalization layers is sufficient and prevents collapse during adaptation.
    TENT restricts updates to BN parameters [16]; the paper relies on this to avoid trivial solutions.
  • domain assumption The normative atlas gives accurate expected tissue proportions for the healthy fetal brain at the evaluated weeks.
    EntropyKL uses the atlas [11] as prior; the authors acknowledge it may bias pathological cases.
  • domain assumption Simulated augmentations capture realistic US acquisition shifts.
    Ranges are chosen to exceed training augmentation but no clinical validity check is provided.
  • domain assumption The source model from [6] is a reasonable fixed pretrained model to adapt.
    All experiments start from this model; no comparison with other source models is made.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Test Time Adaptation for Subcortical Segmentation of the Fetal Brain in 3D Ultrasound." pith.science (2026). https://pith.science/paper/C6IONDZT

@misc{pith2026250208774,
  author       = {Pith},
  title        = {Pith review of: Exploring Test Time Adaptation for Subcortical Segmentation of the Fetal Brain in 3D Ultrasound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6IONDZT}},
  note         = {Machine review of arXiv:2502.08774}
}
read the original abstract

Monitoring the growth of subcortical regions of the fetal brain in ultrasound (US) images can help identify the presence of abnormal development. Manually segmenting these regions is a challenging task, but recent work has shown that it can be automated using deep learning. However, applying pretrained models to unseen freehand US volumes often leads to a degradation of performance due to the vast differences in acquisition and alignment. In this work, we first demonstrate that test time adaptation (TTA) can be used to improve model performance in the presence of both real and simulated domain shifts. We further propose a novel TTA method by incorporating a normative atlas as a prior for anatomy. In the presence of various types of domain shifts, we benchmark the performance of different TTA methods and demonstrate the improvements brought by our proposed approach, which may further facilitate automated monitoring of fetal brain development. Our code is available at https://github.com/joshuaomolegan/TTA-for-3D-Fetal-Subcortical-Segmentation.

Figures

Figures reproduced from arXiv: 2502.08774 by the authors.

Figure 1
Figure 1. TTA framework schematic. The source model is the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. EntropyKL with varying λ (a). LayerInspect perfor￾mance for various rotations, varying the number of layers updated (b) and the learning rate (c). with propagated annotations from an atlas [6], [11]. All images are of size 160 × 160 × 160 voxels, with an isotropic voxel size of 0.6mm. The considered domain shifts were: rotations, scaling, Gaussian smoothing and contrast changes (Gamma Correction), implemented using … view at source ↗
Figure 2
Figure 2. Results from adapting to simulated augmentations. Dice [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: LHS: Average Dice score across all structures at each [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

  1. [6]

    NeuroImage 254, 119117 (2022)

    Hesse, L.S., Aliasi, M., Moser, F., Haak, M.C., Xie, W., Jenkinson, M., Namburete, A.I.: Subcortical segmentation of the fetal brain in 3d ultrasound using deep learning. NeuroImage 254, 119117 (2022). https://doi.org/https://doi.org/10.1016/j.neuroimage.2022.119117

    work page arXiv 2022

  2. [11]

    Nature 623 (10 2023)

    Namburete, A., Papie ˙z, B., Fernandes, M., Wyburd, M., Hesse, L., Moser, F., Cheikh Ismail, L., Gunier, R., Squier, W., Ohuma, E., Carvalho, M., Jaffer, Y ., Gravett, M., Wu, Q., Lambert, A., Winsey, A., Restrepo-M´endez, M., Bertino, E., Purwar, M., Kennedy, S.: Normative spatiotemporal fetal brain maturation with satisfactory development at 2 years. Na...

    work page doi:10.1038/s41586-023-06630 2023

  3. [1]

    Medical Image Analysis 82, 102617 (2022)

    Bateson, M., Kervadec, H., Dolz, J., Lombaert, H., Ben Ayed, I.: Source-free domain adaptation for image segmentation. Medical Image Analysis 82, 102617 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102617

    work page arXiv 2022

  4. [2]

    Neuron 110, 3866–3881 (2021), https://api.semanticscholar.org/CorpusID:235829550

    Dinsdale, N.K., Bluemke, E., Sundaresan, V ., Jenkinson, M., Smith, S., Namburete, A.I.L.: Challenges for machine learning in clinical translation of big data imaging studies. Neuron 110, 3866–3881 (2021), https://api.semanticscholar.org/CorpusID:235829550

    work page 2021

  5. [3]

    NeuroImage 228, 117689 (2021)

    Dinsdale, N.K., Jenkinson, M., Namburete, A.I.: Deep learning- based unlearning of dataset bias for mri harmonisation and confound removal. NeuroImage 228, 117689 (2021). https://doi.org/https://doi.org/10.1016/j.neuroimage.2020.117689

    work page arXiv 2021

  6. [4]

    Medical Image Analysis 81, 102583 (2022)

    Dinsdale, N.K., Jenkinson, M., Namburete, A.I.: Stamp: Simultane- ous training and model pruning for low data regimes in medical image segmentation. Medical Image Analysis 81, 102583 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102583

    work page arXiv 2022

  7. [5]

    In: Proceedings of the 32nd International Conference on International Conference on Machine Learning - V olume 37

    Ganin, Y ., Lempitsky, V .: Unsupervised domain adaptation by back- propagation. In: Proceedings of the 32nd International Conference on International Conference on Machine Learning - V olume 37. p. 1180–1189. ICML’15 (2015)

    work page 2015

  8. [7]

    Single Image Test-Time Adaptation for Segmentation

    Janouskova, K., Shor, T., Baskin, C., Matas, J.: Single image test-time adaptation for segmentation. ArXiv (2023). https://doi.org/https://arxiv.org/abs/2309.14052

    work page Pith review arXiv 2023

Show all 18 references
  1. [8]

    In: Proceedings of the 37th International Conference on Machine Learning

    Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 6028–6039....

    2020

  2. [9]

    part 1: performance of screening examination and indications for targeted neurosonography

    Malinger, G., Paladini, D., Haratx, K., Monteagudo, A., Pulu, G., Timor- Tritsch, I.: Isuog practice guidelines (updated): sonographic examination of the fetal central nervous system. part 1: performance of screening examination and indications for targeted neurosonography. Ul...

    2020

  3. [10]

    In: 5th In- ternational Conference on Learning Representations, ICLR 2017 (2017)

    Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J.: Pruning convolutional neural networks for resource efficient inference. In: 5th In- ternational Conference on Learning Representations, ICLR 2017 (2017)

    2017

  4. [12]

    American journal of obstetrics and gynecology 218(2), S630–S640 (2018)

    Papageorghiou, A.T., Kennedy, S.H., Salomon, L.J., Altman, D.G., Ohuma, E.O., Stones, W., Gravett, M.G., Barros, F.C., Victora, C., Purwar, M., et al.: The intergrowth-21st fetal growth standards: toward the global integration of pregnancy and pediatric care. American journal ...

    2018

  5. [13]

    Com- puter Methods and Programs in Biomedicine p

    P ´erez-Garc´ıa, F., Sparks, R., Ourselin, S.: Torchio: a python li- brary for efficient loading, preprocessing, augmentation and patch- based sampling of medical images in deep learning. Com- puter Methods and Programs in Biomedicine p. 106236 (2021). https://doi.org/https://...

    2021

  6. [14]

    In: Computer Vision – ECCV 2016 Workshops

    Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: Computer Vision – ECCV 2016 Workshops. pp. 443–450 (2016). https://doi.org/10.1007/978 − 3 − 319 − 49409 − 835

    2016 doi

  7. [15]

    In: Medical Imaging with Deep Learning

    Valanarasu, J.M.J., Guo, P., Vibashan, V ., Patel, V .M.: On-the-fly test- time adaptation for medical image segmentation. In: Medical Imaging with Deep Learning. pp. 586–598. PMLR (2024)

    2024

  8. [16]

    In: International Conference on Learning Representations (2021)

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021)

    2021

  9. [17]

    In: Annual Conference on Medical Image Understanding and Analysis

    Wyburd, M.K., Jenkinson, M., Namburete, A.I.: Cortical plate segmenta- tion using cnns in 3d fetal ultrasound. In: Annual Conference on Medical Image Understanding and Analysis. pp. 56–68. Springer (2020)

    2020

  10. [18]

    Medical Image Analysis 94, 103147 (2024)

    Yeung, P.H., Hesse, L., Aliasi, M., Haak, M., Xie, W., Namburete, A.I., et al.: Sensorless volumetric reconstruction of fetal brain freehand ultrasound scans with deep implicit representation. Medical Image Analysis 94, 103147 (2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.