Abstract

Reliable fetal brain biometry is important for prenatal neurodevelopmental evaluation and disease diagnosis, yet current clinical practice relies largely on manual annotation, which is time-consuming and susceptible to inter-observer variability. While deep learning has been applied to automated biometry, pre-trained models tend to underperform when deployed in unseen clinical environments due to domain shifts. Motivated by the lack of domain adaptation methods specifically tailored to biometry tasks, we propose BioTTA, a backbone-agnostic, source-free unsupervised test-time adaptation (TTA) framework that leverages geometric consistency between landmark heatmaps and regressed biometric scalars as a powerful self-supervisory signal for online adaptation. Specifically, BioTTA jointly optimizes three complementary self-supervised objectives at test time, namely length consistency loss, entropy minimization loss, and boundary-gradient loss, all further constrained by atlas-guided anatomical priors, to enable robust adaptation without target labels or access to source data. Extensive experiments on multi-center datasets demonstrate that BioTTA achieves measurement errors on target domains comparable to or lower than those of supervised fine-tuning, while maintaining expert-level diagnostic performance. Code: https://github.com/birthlab/BioTTA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0246_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/0246_supp.zip

Link to the Code Repository

https://github.com/birthlab/BioTTA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiYij_Maximizing_MICCAI2026,
        author = { Li, Yijin AND Liu, Mingxuan AND Chang, Le AND Yang, Hongjia AND Anmahapong, Kasidit AND Hu, Xiaotian AND Wang, Ziang AND Hao, Yingqi AND Li, Haoxiang AND Chen, Yifei AND Li, Xuesheng AND Jia, Fenglin AND Liao, Yi AND Qu, Haibo AND Tian, Qiyuan},
        title = { { Maximizing Domain Generalization in Automated Fetal Brain Biometry } },
        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 an important topic, namely the reduced performance of AI-based methods for fetal brain biometry analysis in cross-domain settings, due to variations in image characteristics across different sources, as well as intrinsic changes in the fetal brain related to gestational age and pathological conditions. To solve this problem, the authors propose the BioTTA framework, an unsupervised test-time adaptation pipeline that adapts pre-trained models to data from other sources during inference, without the need for manual annotations. Particularly, a backbone is first trained for predicting biometric measurements and landmark heatmaps. Afterwards, the pre-trained backbone is adapted through self-supervised optimization, where the model adapts to unlabeled test samples by minimizing entropy, length consistency, and boundary-gradient losses, incorporating also atlas-informed anatomical priors for landmark heatmap prediction. The authors also developed an automated web-based system that reads a 3D image, extracts the biometric measurements, and generates standardized reports for clinical analysis/evaluation.

  • 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 presents several strengths. First, the methodological design is well developed and carefully thought out, resulting in a novel framework for test-time adaptation. In particular, the approach combines different self-supervised objectives that not only leverage landmark heatmap predictions but also incorporate their relationships with biometric lengths and their expected positions along structural edges. Additionally, a structural prior based on an atlas is used to enforce anatomical plausibility. Another strength is the comprehensive evaluation. The authors train the backbone model(s) on a single dataset and perform two types of evaluations, each across two different datasets: a quantitative evaluation (analyzing biometric measurement errors) and a diagnostic evaluation (assessing the final diagnostic performance of the method in conditions such as ventriculomegaly and germinal matrix hemorrhage). They also compare their method with state-of-the-art approaches for the quantitative evaluation and with observer variability for the clinical assessment. Finally, another strength is the application of the proposed method in a web-based system, demonstrating the authors’ intention to translate the approach into clinical practice.

  • 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.

    Concerning the weaknesses, despite the strong evaluation presented by the authors, where two out-of-domain datasets were used for quantitative assessment, in one of the datasets only three measurements were evaluated (LFC), whereas 11 measurements were considered in the other (FeTA). Although this still allows the assessment of the method across different datasets, this lack of uniformity in the analysis may raise questions about how the method would perform across a broader set of measurements on the LFC dataset. Another weakness is the modest discussion of the results. While the authors present the results and their main conclusions, the discussion lacks depth and could be improved by adding comprehensive insights.

  • 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 state that the WCT and FeTA datasets were annotated for all measurements, whereas this was not the case for the LFC dataset, where only landmark annotations for three measurements were used. The authors should clarify why the same level of annotation was not performed for the LFC dataset.
    • The authors present Figure 3 to illustrate results on example images with different gestational ages (GA). However, no specific discussion is provided in the manuscript regarding this figure. Including such a discussion could strengthen the paper, particularly by highlighting what conclusions can be drawn about the performance of the method across different GA, as well as its advantages compared to other methods under varying GA conditions.
    • The results section would benefit from more in-depth analysis and interpretation. For example, Table 2 shows that BioTTA does not outperform the senior radiologist (although it is comparable) on the OoD-GMH dataset, while it does outperform on FeTA. This difference in performance should be discussed. Is one dataset more challenging than the other for the method or for the radiologist?
  • 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?

    This paper presents a complete and well-structured study addressing an important and timely problem. The proposed methodology is clearly described and introduces an innovative test-time adaptation framework that effectively integrates multiple self-supervised objectives with anatomical priors, resulting in a coherent and well-motivated pipeline. The recommendation is further supported by the strength of the experimental validation and the inclusion of a web-based application, which highlights the potential for clinical translation. While the discussion of the results could be more in-depth (which can be addressed in the rebuttal), and some aspects of the evaluation could benefit from greater uniformity, these limitations do not significantly detract from the overall quality, completeness, and contribution of the paper.

  • 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 responded reasonably well to some of the concerns raised in the initial review, particularly the more direct questions regarding the dataset and the insights obtained from the comparison with the radiologist. These clarifications addressed part of the reviewers’ concerns.

    However, the authors did not sufficiently address (or justify) the previously mentioned lack of depth in the discussion section. In particular, the manuscript would still benefit from a more critical analysis of the results and limitations, as highlighted in the first review.

    Nevertheless, as also stated in the initial evaluation, the topic addressed by the paper is highly relevant, and the methodological design is well developed and carefully thought out, resulting in a strong framework for test-time adaptation. Considering the overall quality, relevance, and potential impact of the work, the reviewer maintains the recommendation for acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper addresses an important problem in fetal brain biometry, namely cross-domain generalization, and attempts to solve it under a source-free unsupervised test-time adaptation setting. The authors propose BioTTA, which is intended as a biometry-aware TTA framework. Concretely, the method uses several self-supervised objectives, including a length consistency loss, to adapt the model at test time without target labels. The paper also goes beyond reporting technical metrics and attempts to demonstrate clinical usefulness through comparison with clinical experts.

  • 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 paper tackles an important problem in FeTA biometry and adopts a reasonable constraint setting, namely source-free unsupervised test-time adaptation. The use of a length consistency loss to design a biometry-aware TTA framework is a clear methodological contribution. In addition, the authors attempt to move beyond laboratory-style AI evaluation by comparing BioTTA with clinical experts, which reflects an interest in practical clinical utility. Finally, the ablation study is well designed and clearly shows the contribution of the main components of the proposed method.

  • 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 main weakness of this paper is that the comparison framework is not sufficiently fair. In domain adaptation, different paradigms allow access to different kinds of information, and comparison methods should therefore be matched under the same information constraints. For clarity, one can distinguish four settings: (A) OOD detection, which aims to identify whether an input is outside the source distribution; (B) UDA, which uses target data during training; (C) TTA, which uses target input during inference; and (D) SDG, which assumes no access to target information at all. This paper clearly belongs to (C), namely source-free unsupervised test-time adaptation, so the core evaluation should be organized within that setting.

    Major weakness 1.Essential baselines for source-free unsupervised TTA are missing

    In this setting, I believe the following comparisons are necessary for a fair evaluation.

    (i) Source Model This is the basic baseline and is required to show how the source-pretrained model performs on the target domain without adaptation.

    (ii) A generic source-free TTA baseline TENT is the most direct example. The authors are clearly aware of it and mention it in the paper, yet it is omitted from the reported comparisons. I would be interested in hearing the authors’ reasoning here. For a paper proposing a source-free test-time adaptation method, comparison against a generic baseline such as TENT appears highly relevant unless there is a clear reason why such a comparison would not be appropriate in this setting.

    (iii) Common source-free TTA methods used for medical brain image inputs Although this paper uses dense spatial prediction through heatmaps and landmark localization rather than segmentation, it would still be strongly desirable, or at least worth discussing explicitly, to compare against source-free TTA methods that are commonly used in medical brain imaging. Examples include Karani et al., “Test-time adaptable neural networks for robust medical image segmentation,” atlas-guided TTA for 3D medical segmentation, and one-pass or plug-and-play medical TTA methods such as TestFit. Even if the authors believe that these methods cannot be applied directly to FeTA biometry, the paper should explain why they were not considered appropriate. At present, such an explanation is not provided.

    (iv) A source-free TTA method with a similar key idea or with the same target domain, i.e., FeTA This point is particularly important. Through reference [10], the authors are already aware of a TTA study that is directly related to FeTA or fetal brain MRI morphometry. That study shares substantial overlap with the present paper: fetal brain MRI morphometry, target domain = FeTA, TTA, and the use of anatomy or atlas priors. It is therefore a strong comparable method and should be included experimentally. Differences in backbone, annotation protocol, or exact label setting may exist, but such differences are usually not sufficient to justify complete exclusion. At minimum, the paper should explain why this comparison was not performed.

    (v) Ablation over the proposed components I mention this only to complete the evaluation framework. This is not a weakness. On the contrary, the authors handled this part well.

    By contrast, the methods that are actually reported do not match the proposed paradigm.

    • FetalSynthGen [24] is based on synthetic data augmentation to improve robustness during training. It is not an inference-time online adaptation method(C), and therefore it does not belong to the same source-free TTA setting as the proposed approach. (maybe this is B or D)
    • DGMLD [8] uses 10 labeled target samples and is therefore target-supervised. It may serve as an upper-bound reference, but it is not an appropriate core baseline for source-free unsupervised TTA. (this is B)

    In other words, the key baselines in (ii), (iii), and (iv) are missing, while the reported comparisons mostly involve methods that do not share the same information constraints as the proposed method. As a result, the current experiments do not convincingly show how competitive BioTTA is as a source-free TTA method. Instead, they read more like a list of methods from different paradigms placed in the same table.

    Major weakness 2.Table 2 is interesting, but it does not directly validate the claimed advantage of the TTA method

    Table 2 is clinically interesting. In our field, it is important not only to compare technical metrics, but also to demonstrate actual clinical usefulness. If designed rigorously and fairly, this kind of evaluation could add substantial value even if the purely technical comparisons were somewhat limited. For that reason, I view Table 2 as potentially very meaningful. However, as currently described, it is not fully clear to me what exactly this table is validating, and I would be interested in the authors’ clarification.

    From the paper, my understanding is the following:

    • BioTTA predicts landmarks and biometric measurements.
    • Those biometrics are then converted into VM diagnosis, VM grading, and GMH grading using rule-based clinical criteria.
    • The radiologists, in contrast, directly interpret the MRI scans.

    If this interpretation is correct, then Table 2 may not directly test the advantage of BioTTA as a TTA method. Rather, it would compare a full TTA-based biometry pipeline against human readers. That is a different question. In that case, the table would read more as an AI-pipeline-versus-human-reader experiment than as a controlled validation of the proposed TTA method itself. This would not make the evaluation unimportant, but it would make its role in the paper different from what I initially expected.

    A fair TTA-oriented evaluation would require at least the following comparisons:

    • source model biometry → rule-based diagnosis
    • generic TTA biometry → rule-based diagnosis
    • BioTTA biometry → rule-based diagnosis

    This would allow the reader to assess whether the proposed TTA actually improves downstream diagnostic usefulness relative to other TTA methods or to the source model alone. From a clinical workflow perspective, an even better comparison would be radiologist alone versus radiologist + model measurement assistance. As it stands, Table 2 appears potentially interesting and important, but I am not yet certain how it should be interpreted in relation to the central TTA claim. I would therefore welcome clarification from the authors.

    Minor weakness

    The authors partially explain why they chose setting (C), but they do not fully answer why this setting is more appropriate than (A), (B), or (D) for the actual clinical deployment scenario. Questions remain, such as: how many optimization steps are acceptable per patient at test time; whether this conflicts with real-time or near-real-time workflow; whether per-patient online updating is operationally acceptable; and whether the setting assumes a continuous target stream or independent one-shot patients. That said, I do not regard this as a major weakness in a short paper, and this point alone would not lead me to reject the submission.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    What matters most to me is not whether the method looks technically fancy, but whether the paper addresses an important problem and evaluates it in a fair and convincing way. This paper clearly addresses an important problem, and I do not think the chosen constraint setting is fundamentally misguided. However, I do feel that several core elements of a fair comparison framework are currently missing or insufficiently explained, especially for a paper that positions itself as a source-free TTA method.

    I do not believe that a conference paper must compare against every baseline that a reviewer can imagine. It would be excessive to reject a paper simply because it did not include every recent study under the same constraint setting. However, I do care a great deal about whether the authors selected fair comparison methods within the framework they themselves claim to operate in. If the comparison framework is substantially ambiguous, or if several of its key components are missing, acceptance becomes difficult.

    If the authors can convincingly explain why the important comparison groups in (ii)–(iv) were not included, and if they can clarify that Table 2 is indeed meaningful for evaluating BioTTA in a way that differs from my current interpretation, then I believe the paper could still have publishable value. At this stage, my judgment is weak reject, and I would welcome the authors’ rebuttal.

  • Reviewer confidence

    Very confident (4)

  • [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.

    After reading the rebuttal and the other reviews, I would like to clarify why I maintain my Reject. My concern is not that the paper lacks merit. On the contrary, I agree that the clinical problem is important, that the source-free test-time adaptation setting is well aligned with clinical deployment, and that the proposed method is task-specific and technically meaningful. In particular, the length-consistency idea, differentiable landmark extraction, and the overall biometry-aware adaptation design are reasonable contributions, and the ablation study supports the importance of the proposed components.

    However, precisely because the paper is positioned as a source-free unsupervised TTA method, I believe the central evidence should show whether BioTTA is better than appropriate methods under the same information constraints. My concern is not that the authors failed to include every possible baseline or every recent TTA paper. That would be an unreasonable expectation for a MICCAI short paper. Rather, my concern is that the comparison framework does not clearly define a minimal set of comparable methods within the same source-free, unlabeled-target, test-time adaptation regime, nor does it clearly explain why the selected methods are the appropriate ones for this regime.

    For example, I would have expected a comparison framework with a logic such as: “among generic TTA methods, we compare against methods whose information constraints and task form are reasonably compatible, such as TENT and possibly other applicable entropy- or stability-based TTA baselines(EATA and SAR); methods requiring a different training protocol, such as auxiliary-task training(eg, TTT), or methods assuming a continual target stream(eg, CoTTA), are excluded for these specific reasons; segmentation-specific medical TTA methods are discussed but not used as direct baselines because their output structure is incompatible with landmark heatmap biometry.” I am not saying that these exact methods must all be included. The key point is that the paper should provide a principled comparable-method selection under the claimed TTA setting.

    Regarding TENT specifically, I still do not fully understand the rebuttal explanation, as detailed in the additional comment below. I separated that point as an additional comment because it is not the central basis for my final recommendation. I am willing to accept the authors’ clarification that a TENT-like entropy-minimization baseline was intended to be represented in the ablation study. My main concern is broader: even if TENT is treated as present, the manuscript still does not provide a clear and principled comparator-selection logic for the central source-free TTA claim.

    Instead, even after considering the rebuttal, the main comparison remains difficult to interpret from this perspective. The reported methods mix different information regimes: the Source Model is a no-adaptation lower reference; FetalSynthGen is a training-time domain randomization / domain generalization method; DGMLD is presented in a target-supervised or fine-tuned form; supervised fine-tuning uses target labels and is therefore an upper/reference setting; and BioTTA is the only clearly source-free unsupervised TTA method in the main comparison. These methods are useful references, but they do not by themselves establish whether BioTTA is a better TTA method than appropriate alternatives under the same source-free test-time information constraint. In particular, FetalSynthGen and DGMLD may be relevant neighboring methods, but the manuscript does not sufficiently justify why they form the main comparator set for the central source-free TTA claim.

    This is why I still cannot conclude that BioTTA is well validated as a TTA method, even though I agree that it is well motivated and task-specific. The strongest statement I can make from the current evidence is that the authors propose a plausible method for an important clinical problem and show strong improvements over a mixture of no-adaptation, domain-randomization, and target-supervised reference settings. However, I do not think the current comparison framework is sufficient to establish that BioTTA is superior as a source-free unsupervised TTA method.

    The same issue also applies to the radiologist comparison. I agree that Table 2 is clinically interesting and valuable. It supports the claim that an AI-based biometry-to-rule pipeline can reach diagnostic performance comparable to clinical readers. This is useful evidence for clinical usability. However, as clarified in the rebuttal, Table 2 is not evidence of the specific gain from BioTTA as a TTA method. It does not compare source-model biometry followed by the same rule-based diagnosis, generic-TTA biometry followed by the same rule-based diagnosis, and BioTTA biometry followed by the same rule-based diagnosis. Therefore, Table 2 helps bridge the clinical motivation, but it cannot show that BioTTA improves downstream diagnosis relative to no TTA or to another comparable TTA method. The manuscript language, including statements that BioTTA is benchmarked against radiologists and provides diagnostic reliability comparable to senior radiologists, risks blurring the distinction between the clinical usability of the full pipeline and the TTA-specific contribution.

    I have considered the more positive interpretations from the other reviews and the meta-review. I agree with them that the paper addresses an important problem and contains meaningful technical contributions. My recommendation remains Reject not because I require a larger number of baselines, nor because I require one specific method that happens to be missing. Rather, it is because the paper does not yet provide the essential comparison logic needed to support its own central claim: that BioTTA is an effective source-free unsupervised TTA method for this clinical biometry problem.

    [Additional comment: TENT clarification] I have carefully reviewed your response regarding the inclusion of the TENT baseline. Even under these favorable interpretations, I still encounter inconsistencies that make the comparison framework difficult to understand. The rebuttal states that “the ablation setting with only entropy minimization is equivalent to TENT, which alone actually increases error.” I considered several possible interpretations of this statement.

    First, if the authors refer to the Lent-only row in Table 3, then this does appear to be the natural entropy-only setting, since Lent is defined as the entropy minimization loss. However, this row does not increase error. It improves the average error from 2.761 mm for the source model to 1.995 mm, corresponding to a -27.74% relative change. Under this interpretation, the rebuttal’s claim that entropy minimization alone increases error contradicts Table 3.Second, if the authors instead refer to the row that increases error to 3.016 mm (+9.240%), that row includes both Atlas Guide and Lent. Therefore, it is not an entropy-only setting and cannot be directly equated with canonical TENT. Under this interpretation, the statement “increases error” is consistent with the number, but the identification with an entropy-only/TENT setting is not. Third, if the intended TENT-equivalent result is neither of these rows, then I cannot identify where the TENT-equivalent baseline is reported in the submitted manuscript. In that case, the concern about the absence of an explicitly reported generic source-free TTA baseline remains. Finally, even if the Lent-only row is accepted as the TENT-equivalent baseline, it was not reported as an explicitly named TENT baseline in the main comparison table. This matters because TENT is not a minor ablation component but the canonical generic comparator for source-free TTA. Presenting it only as an unnamed ablation row makes it difficult to assess the relative gain of BioTTA over a standard TTA method.



Review #3

  • Please describe the contribution of the paper

    This paper studies the domain-shift problem in automated fetal brain MRI biometry. The authors note that current clinical practice still relies on manual annotation, which is time-consuming and subject to inter-observer variability, while deep learning models often suffer substantial performance degradation when deployed across centers due to differences in scanners, imaging protocols, field strengths, SVR reconstruction methods, gestational age ranges, and pathology diversity. To address this gap, the paper proposes BioTTA, a backbone-agnostic, source-free unsupervised test-time adaptation framework. The method consists of supervised pretraining on a labeled source domain followed by sample-wise adaptation on unlabeled target data, without access to source data or target labels. Its key insight is that the regression branch, which integrates global features, is more robust to domain shift than the spatially sensitive heatmap branch. Based on this, the framework jointly optimizes three self-supervised losses at test time: a length consistency loss enabled by a sparse weighted differentiable landmark detector to overcome the non-differentiability of argmax and distill geometric knowledge from the regression branch to the heatmap branch; an entropy minimization loss to suppress background noise and sharpen confidence at anatomical landmarks; and a boundary gradient loss to anchor predictions to high-contrast anatomical boundaries. The framework further incorporates atlas-guided anatomical priors by registering gestational-age-matched atlases to the target image space and using the resulting anatomical masks to constrain predicted heatmaps. Experiments on 259 fetal brain MRI scans from five centers show substantial improvements on FeTA and LFC, with performance comparable to or even better than supervised fine-tuning using 10 labeled target samples, and additional diagnostic evaluation is reported on ventriculomegaly and germinal matrix hemorrhage. Overall, the paper presents BioTTA as the first TTA framework for fetal brain biometry, with the differentiable landmark detector and regression-to-heatmap distillation as its main methodological innovations, supported by multi-center validation on both CNN and Transformer backbones.

  • 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 methodological innovation is substantial and task-specific. The sparse weighted differentiable landmark detector resolves the gradient-blocking issue in converting heatmaps to coordinates, and the length consistency loss leverages the task-specific geometric structure of biometry by enforcing consistency between inter-landmark distance and regression output. This regression-to-heatmap distillation is the core algorithmic contribution rather than a simple combination of existing losses. The integration of atlas priors is also reasonable, as the gestational-age-matched fetal brain atlas provides structural constraints to keep landmarks within anatomically plausible regions.

    2) The experimental validation is extensive and forms a relatively complete evidence chain. The method is evaluated on 259 images from five centers using two target-domain datasets, FeTA and LFC. It is validated on both a CNN backbone (BioCNN) and a Transformer backbone (BioFormer), suggesting that the approach is not dependent on a specific architecture. The ablation study supports the central claims, with the full framework achieving a 47.12% error reduction over the source-model baseline. Moreover, without target annotations, the method reaches performance comparable to or even better than supervised fine-tuning with 10 labeled target samples.

    3) The paper has clear clinical deployment value. The problem definition targets a real bottleneck in fetal brain biometry, namely domain shift in practical cross-center deployment, rather than a purely academic setting. The source-free formulation is well aligned with clinical requirements because it avoids access to source data and target labels, thereby reducing privacy-compliance and annotation-cost barriers. In addition, the paper goes beyond reporting measurement errors and evaluates diagnostic performance on ventriculomegaly diagnosis and germinal matrix hemorrhage grading against three radiologists with different levels of experience, which supports its clinical relevance.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1) The core assumption is insufficiently validated, and the potential impact of registration on individualized structural characteristics is not analyzed. The paper assumes that the regression branch is more robust to domain shift than the heatmap branch, but does not provide direct experimental evidence for this claim. More importantly, the atlas registration process may alter individualized brain structural characteristics through nonlinear deformation, potentially distorting morphological measures such as gray matter volume or cortical thickness. The paper does not analyze the effect of registration error on biometry accuracy, nor does it discuss whether degraded registration quality in pathological brains, such as ventriculomegaly or germinal matrix hemorrhage, may introduce systematic bias.

    2) The technical boundary with related work is not sufficiently clear, and the novelty requires further clarification. Li et al. (MIDL 2025) has already presented a highly similar anatomy-guided test-time adaptation method for automated fetal brain MRI morphometry that also uses atlas registration to constrain heatmap prediction. Although BioTTA differs in its TTA mechanism and self-supervised loss design, the use of atlas priors is highly overlapping, and the paper does not adequately clarify the technical distinction and contribution difference. In addition, the entropy minimization loss and boundary gradient loss are largely direct applications of the TENT paradigm and common edge-based constraints, respectively, which limits their novelty.

    3) The diagnostic evaluation remains limited in scope, and the potential for normative reference construction is not explored. The paper validates only ventriculomegaly diagnosis and germinal matrix hemorrhage grading, without evaluation on other common fetal brain abnormalities such as agenesis of the corpus callosum, cerebellar maldevelopment, or neural tube defects. More importantly, although the adapted model could in principle support automated biometry on unlabeled data from new centers and potentially enable gestational-age-specific normative reference construction, the paper does not discuss this direction or evaluate sensitivity for abnormality detection.

    4) The experimental validation and statistical rigor are still limited. The ablation study is conducted only on BioCNN, without corresponding ablations on BioFormer or analysis of interactions among the loss terms. The reported improvements are not accompanied by statistical significance testing, so some smaller gains may not be significant. The fairness of some comparisons is also unclear, for example when DGMLD is used outside its intended metric scope. In addition, the target-domain sample size is limited, with only 50 cases each in FeTA and LFC.

  • 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.

    (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 addresses an important and clinically meaningful problem in fetal brain MRI biometry, namely the degradation caused by domain shift in cross-center deployment. The motivation is strong, as current manual biometry is time-consuming and variable across observers, while existing adaptation methods for this task remain largely unexplored. The main contribution is a source-free test-time adaptation framework that is well matched to realistic clinical deployment. In particular, the sparse weighted differentiable landmark detector and the length consistency loss constitute the central methodological contribution by using task-specific geometric structure to transfer robust information from the regression branch to the heatmap branch. The experimental evidence is also relatively strong, with multi-center validation, two target datasets, two backbone families, ablation analysis, and diagnostic evaluation against radiologists, and the reported performance is comparable to or better than supervised fine-tuning with a small number of target labels. At the same time, there are nontrivial limitations, including insufficient validation of the core robustness assumption, unclear novelty boundaries relative to closely related prior work, limited disease coverage, and incomplete experimental and statistical analysis. Overall, I find the problem important, the core methodological idea meaningful, and the empirical results convincing enough to support a weak accept recommendation despite these concerns.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

# R1: LFC metrics: Only 3 cerebellar measurement labels are available in the LFC public dataset, which limits our evaluation scope.

BioTTA on OoD-GMH: The OoD-GMH dataset (SS/TSE sequences, NesVoR recon) exhibits a larger domain gap from the source domain in acquisition parameters and reconstruction methods, resulting in more pronounced distribution shift. Nevertheless, BioTTA’s diagnostic results remain comparable to senior radiologists, validating its cross-domain generalization and clinical utility.

# R2: Regarding the comparison framework:

(ii) TENT is a source-free TTA baseline that updates BN parameters by minimizing entropy. Our method adopts the same update strategy while designing three self-supervised losses: length consistency, entropy minimization, and boundary gradient. Consequently, the ablation setting with only entropy minimization is equivalent to TENT, which alone actually increases error. The next revision will explicitly note that TENT is already included in the ablation framework.

(iii)The method of Karani et al. targets segmentation tasks and is not applicable to landmark heatmap regression-based biometric measurement. Unsupervised TTA methods for brain biometry are extremely scarce. We tested DGMLD [8], a TTA method specifically designed for fetal cerebellar measurement that exploits the cerebellum’s regular geometric contour for boundary pre-training. However, DGMLD’s direct TTA version performed much worse than the source model, likely due to large imaging parameter differences, so we adopted its supervised fine-tuned version.

(iv) To comply with the double-blind policy, we cannot discuss the relationship with the reference [10]. Technically, [10] is equivalent to our atlas-constraint only ablation setting, and our results confirm that the proposed additional losses yield significant improvements over this baseline, which will be explicitly clarified in the next revision.

Regarding Table 2:

(1) BioTTA’s quantitative advantage over other methods is demonstrated in Table 1; (2) Table 2 validates clinical usability by comparing with radiologists; (3) We appreciate this suggestion and will supplement other methods’ diagnostic performance in the next revision. “Preliminary experiments confirm that the source model achieves an accuracy of 0.921 and F1 of 0.924 for VM Diagnosis, and an accuracy of 0.908 and F1 of 0.912 for GMH Grading on the OoD-GMH dataset, both inferior to BioTTA (Acc 0.932, F1 0.934 and Acc 0.919, F1 0.922, respectively).

# R3: Regression branch robustness: In segmentation domain adaptation, auxiliary prediction of global quantities such as region proportions has been shown to be effective and domain-invariant (Bateson et al., 2020). Inspired by this, we introduce a length regression branch to model more global structural relationships among landmarks (Song et al., 2024). Empirically, this branch provides stable structural guidance for the heatmap branch and improves robustness under domain shift.

Atlas registration constraint: We warp atlas landmarks to the test image space with a 4 mm radius soft constraint to prevent outlier heatmap maxima, which only provides a search region without strictly confining predictions or relying on fine structural details.

MIDL short paper: To comply with the double-blind policy, we cannot discuss the relationship with that paper. Our ablation study demonstrates that the additional losses significantly improve performance over using the atlas alone.

Space limitations: (1) Standard developmental curves with gestational-age-specific percentiles have been constructed but omitted due to space; they will be publicly released with the codebase. (2) Larger-scale validation with broader pathological coverage and more rigorous ablation experiments is planned for future work. (3) OoD annotation is highly time-intensive, requiring experts to manually label anatomical landmarks; we currently annotated 50 cases and will expand in subsequent studies.




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 find merit in this paper - particularly regarding the importance of the application and novelty (sparse weighted differentiable landmark detector and length consistency loss, task-specific geometric structure) as well as the cross-center aspect and general performed validation. The source-free test time framework is adapted well to clinical need.

    There are however major weaknesses / points to be addressed: more in-depth discussion of the results (incomplete experimental and statistical analysis), insufficient validation of the core robustness assumption, unclear novelty boundaries relative to closely related prior work (BioTTA, missing groups in table 2)

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The authors have addressed some of the concerns and there was feedback and discussion with the reviewers. While one reviewer continues to recommend acceptance and mentioned that some concerns have been addressed, they specify themselves that the criticism re lack of depth, more critical analysis were not appropriately addressed and that the acceptance recommendation is mainly based on the topic being highly relevant and the methodology being interesting. The other reviewer engaged intensively with the rebuttal and maintained (even stronger then before) the recommendation reject arguing similarly that the main concern about methodological clarity and a lack of proof about the main contribution, supervision free TTA.

    In summary, significant strengths remain

    • important application with possible future impact
    • interesting methodological contribution,

    Identified WEAKNESSES remain

    • lack of proper comparison to the state of the art
    • insufficient methodological clarity with regard to the exact contributions
    • lack of critical evaluation/discussion of the results



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 reviews are mixed, with two reviewers recommending acceptance and one reviewer maintaining rejection after rebuttal. I agree that Reviewer 2 raises a substantive concern: for a paper positioned as source-free unsupervised test-time adaptation, the comparison framework should more clearly separate no-adaptation, domain generalization, supervised adaptation, and true source-free TTA baselines. The rebuttal only partially resolves this, and the authors should explicitly clarify the TENT-equivalent ablation, the relationship to the atlas-only prior, and the role of Table 2 as clinical-usability evidence rather than direct evidence for the TTA component.

    However, I do not consider this issue sufficient for rejection. The paper addresses an important clinical deployment problem in fetal brain biometry, proposes a task-specific and technically meaningful framework, including differentiable landmark extraction and length-consistency-based adaptation, and provides multi-center validation with two backbones, ablations, and clinically oriented diagnostic evaluation. Reviewers 1 and 3 judge the work to be relevant and above the acceptance threshold, and Reviewer 2 also acknowledges the merit, clinical relevance, and meaningful technical design, despite disagreeing on the sufficiency of the validation.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    After rebuttal, R2 remains a reject, finding the TENT explanation inconsistent with the reported ablation numbers and the radiologist comparison incapable of isolating the TTA-specific gain. Because of these unresolved fundamental issues, the recommendation is therefore Reject.



back to top