Abstract

Accurate fetal birth weight (FBW) estimation shortly before delivery is clinically valuable yet challenging due to its reliance on operator expertise, particularly in low-resource settings. To reduce this reliance, we study near-term birth-weight regression from blind-sweep ultrasound (US) videos acquired within 48 hours prior to delivery, with post-delivery weighing as ground truth. Accordingly, we propose a foundation model–driven key anatomy frame selection framework that enables accurate FBW regression despite the absence of plane constraints in blind sweeps. Our highlights are as follows: (1) We believe this is the first work to estimate FBW using blind-sweep US videos, enabling operator-independent assessment. (2) An Anatomy-Guided Frame Selection module equipped with a vision-language foundation model is proposed for keyframe collection in unconstrained sweeps. (3) A Redundancy-Aware Feature Compression module is designed to compress frame features while preserving task-relevant information, alleviating temporal redundancy. Extensively validated on prospectively collected data from 839 patients, our method achieves an MAE of 161.3 g, with 90.23% and 100% of cases falling within 10% and 15% absolute percentage error, outperforming typical Hadlock estimation and strong competitors. Codes are available at https://github.com/ouleoule/BlindSweep-EBW.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1307_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ouleoule/BlindSweep-EBW

Link to the Dataset(s)

N/A

BibTex

@InProceedings{OuLe_Foundation_MICCAI2026,
        author = { Ou, Le AND Zhu, Xiliang AND Liang, Huanwen AND Pan, Wenxiong AND Huang, Yuhao AND Deng, Yuxiang AND Sheng, Xuan AND Yin, Hong AND Xiao, Juhua AND Zhou, Xin AND Ni, Dong},
        title = { { Foundation Model-Driven Key Anatomy Frame Selection for Blind-Sweep Ultrasound Fetal Birth Weight Estimation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces a framework for fetal weight estimation from blind-sweep ultrasound studies. A key aspect of the proposed approach is its hierarchical anatomy-guided frame selection strategy, which is particularly relevant in this setting, where blind sweeps may fail to capture standard planes and are often affected by noisy and redundant data. The method incorporates an anatomy-guided selection module based on FetalCLIP, together with a redundancy-aware feature compression mechanism to reduce overlap among neighboring frames. The proposed framework is compared against conventional clinical estimation methods, including Hadlock and INTERGROWTH-21st, as well as other deep learning approaches, and is reported to outperform all of them. Overall, the paper addresses a clinically meaningful problem and has clear potential relevance for low-resource settings, where access to expert sonographers may be limited. To the best of my knowledge, this is the first work that leverages blind sweep ultrasound acquisitions for this type of estimation task.

  • 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.
    • Clear: The paper is well organized and provides a clear presentation of the motivation, method, and results.

    • Relevant: it focuses on a clinically relevant task using blind-sweep ultrasound studies. This is particularly valuable in low-resource settings, where expert sonographers are scarce.

    • Methodological contribution: To the best of my knowledge, the paper presents a novel framework tailored to the challenges of blind-sweep ultrasound. The method shows improved performance over both traditional clinical estimation approaches and existing deep learning models, and the included ablation studies help clarify the contribution of each component.

  • 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.
    • Limited clarity of the inference pipeline: it is not clear how the inference will be obtained by non-expert sonographers, specifically, what type of anatomical concepts or prompts have to be used in the vision-language model for frame selection.

    • Generalizability to real low-resource settings: All data were acquired using a single mid-to-high-end ultrasound system and by expert sonographers. This should be discussed more explicitly, as the method may not generalize as well to lower-end or portable devices, or to acquisitions performed by less experienced operators, both of which are particularly relevant to the intended application setting.

    -Clarification: The manuscript should clarify why the number of studies exceeds the number of pregnancies, and the clinical acceptance thresholds for PC10 and PC15 should be supported with appropriate references.

  • 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 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

    The paper addresses a clinically relevant problem and presents a thoughtful approach for fetal weight estimation from blind-sweep ultrasound, a setting of particular importance for low-resource environments. The method is well motivated, clearly presented, and strengthened by comparisons with clinical baselines and deep learning models. That said, several aspects would benefit from further clarification. In particular, the inference workflow is not yet fully transparent, especially regarding how the anatomy-guided frame selection operates in practice. The practical deployment perspective would also be strengthened by a broader discussion of how the method may generalize beyond expert-acquired studies and beyond the specific ultrasound system used in this work. In addition, some details should be clarified, such as the distinction between the number of studies and the number of pregnancies, and references should be provided for the clinical acceptance thresholds used in the evaluation.

  • 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 addresses a clinically relevant problem with a method that is well designed to tackle the practical challenges of fetal weight estimation from blind-sweep ultrasound, particularly in low-resource settings. A notable strength is the anatomy-guided, redundancy-aware framework, which is well-motivated for handling noisy and non-standard blind-sweep studies. The manuscript is clearly structured, the method shows improved estimation performance compared against both clinical baselines and existing deep learning methods, and includes an ablation study that supports the contribution of its main components. Although some aspects of the inference workflow could be described more clearly, particularly the anatomical concepts used for frame selection and how the method would be implemented in practice, these limitations are relatively minor compared with the overall strengths of the work.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The authors have addressed concerns about generalizability and clarified the identified ambiguities in detail.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a framework for estimating fetal birth weight (FBW) directly from blind-sweep ultrasound videos acquired within 48 hours before delivery, without requiring standard-plane acquisition or manual biometric measurements. The key technical contribution is a hierarchical frame selection and regression pipeline composed of:

    (1) an Anatomy-Guided Frame Selection (AFS) module using a fetal vision-language foundation model to identify anatomically informative frames from unconstrained blind sweeps, (2) a Redundancy-Aware Feature Compression (RAFC) module to remove temporally redundant frames while preserving task-relevant information, and (3) a Prior-Guided Attention Regressor (PGAR) for final birth-weight prediction.

    The practical contribution is the demonstration, on a prospective dataset of 870 exams from 839 pregnancies, that blind-sweep ultrasound may support accurate, more operator-independent near-term FBW estimation, achieving performance that is reported to surpass Hadlock-4 and several learned baselines.

  • 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.organisation and structure of the paper is clean and clear.

    2.The method is conceptually well aligned with the problem structure. Blind-sweep videos are noisy, long, and highly redundant, with only sparse informative content. The two-stage selection strategy matches this challenge well: AFS tries to identify anatomy-bearing frames using cross-modal similarity to anatomical concepts, and RAFC then removes redundancy and promotes diversity/task relevance before regression. This is well intentioned design choice, and much more thoughtful than naively averaging all frame features or applying an off-the-shelf video regressor.

    3.The paper shows promising empirical results. The dataset is prospective, relatively large for this type of ultrasound work, and the test results are strong: MAE 161.34 g, MAPE 4.89%, PC10 90.23%, and PC15 100%. The comparison against both clinical Hadlock-4 and several learned baselines is useful, and the ablation study supports the importance of both AFS and RAFC.

  • 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 study assumes availability of blind-sweep ultrasound within 48 hours prior to delivery. While this is plausible in scheduled or high-risk clinical settings (e.g., induction or planned cesarean section), is it routine for all pregnancies, like spontaneous labour? The manuscript would benefit from clarifying the clinical scenarios in which this approach is intended to be used and discussing how this assumption impacts generalizability.

    Does this introduce selection bias – dataset likely includes more planned or high-risk deliveries? Not representative of general population/ community screening.

    The authors could expand on how useful is knowing fetal birth weight at such narrow time windows (<48 hours before delivery). What actual interventions can be gained with this? It seems fetal growth restriction diagnosis would be too late, and perhaps only macrosomia cesarean section is the only use case?

    Moreover, it would be helpful to understand whether the reported improvements over Hadlock-based estimation would alter decisions around clinically relevant thresholds. Does improved accuracy at that timepoint actually change clinical decisions? For example: if Hadlock says 3900g and model says 3700g would that incur an actual change in chosen intervention? The manuscript would benefit from a clearer discussion of the downstream clinical impact of improved prediction accuracy.

    More broadly, the assumption that blind-sweep ultrasound would be routinely performed within 48 hours prior to delivery may limit real-world applicability, particularly outside controlled or resource-equipped settings.

    The paper does not fully disentangle whether performance gains come from the proposed architecture or from the strength of the pretrained foundation model and the dataset/protocol itself. Since AFS uses FetalCLIP, one would like stronger controls: for example, selection using a non-medical encoder, random anatomical prompts, simpler confidence heuristics, or explicit standard-plane detectors. That would help isolate the actual value of the proposed anatomy-guided selection mechanism

    The baseline comparison may not be fully fair. Several baselines cited, such as EchoNet-Dynamic and ViFi-CLIP, are generic video architectures rather than for blind-sweep fetal ultrasound. Adapting them with a regression head is reasonable as a sanity check, but strong conclusions should rely more on task-matched baselines. Similarly, SelectGA is a gestational-age method repurposed for FBW; which is much nicer due to it’s fetal ultrasoun domain specificness but not a definitive state-of-the-art comparator for birth-weight estimation. The authors compare to Hadlock-4, which is important clinically, and the biggest strength of the results section, but a more convincing evaluation would include additional strong fetal-ultrasound-specific alternatives.

    Whilst the equations are useful and mostly clear, the authors could make it easier e.g. define the dimensions of s_imp and big H in equation 6, (I assume KxD where D is feature dimension). Also lower case h in equation 7 was not defined, although I assume it’s the single frame feature vector of H?.

    Additionally, the implementation of the sliding window in Eq. (7) is ambiguous: it is unclear whether it is applied over the original temporal sequence or over the selected top-K frames, which affects the interpretation of “temporal” redundancy.

    Lack of some details about anatomical text prompts, whether prompts are fixed or tuned, and how many and which anatomy classes were used, and how six loops per exams are fused. How K, L window size affected.

  • 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 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

    I encourage the authors to improve reproducibility by providing more detail about the anatomical text prompts, hyperparameter choices, how six loops per exam are fused, and exact train/validation/test handling at the patient and exam levels. These additions would make the paper substantially stronger.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The work is interesting and useful - I’d be happy to change to accept once these small issues are addressed that make the paper more readable and complete in my opinion. I’m very experienced (4 years of PhD research) with blind sweep fetal ultrasound videos which the paper investigates, and specifically biometry planes and estimating gestational age. I’m also experienced with the techniques the authors used.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 rebuttal addresses the main concerns I raised sufficiently, and I recommend acceptance.

    My initial concerns were mainly about the clinical scenario, possible selection bias, baseline fairness, reproducibility details, and whether the performance gains were clearly attributable to the proposed anatomy-guided selection pipeline rather than to the pretrained foundation model and acquisition protocol. After rebuttal, I think the paper remains somewhat limited in scope, but the authors have provided enough clarification that these concerns no longer justify rejection.

    First, the authors clarified the intended clinical scenario more convincingly. The target use case is pre-delivery FBW assessment in term pregnancies, rather than general fetal growth screening. They also clarify that delivery time was unknown at scanning and that women approaching delivery underwent blind-sweep ultrasound, with inclusion determined by whether delivery occurred within the predefined 48-hour window. This helps address my concern that the method may only apply to planned deliveries. The reported cesarean-to-vaginal ratio also suggests that the cohort is not exclusively composed of planned or highly selected delivery cases. I still think the clinical impact of improved prediction within 48 hours of delivery should be discussed more explicitly in the paper, especially in terms of which decisions would actually change, but the rebuttal makes the intended setting clearer.

    Also, the authors addressed some reproducibility and implementation ambiguities that I raised. They clarify that the anatomical prompts were fixed and taken from the public FetalCLIP repository, that the split was pregnancy-level, that the number of exams exceeds pregnancies because some women had up to two eligible scans, and that six loops per exam are processed independently before concatenating 6L features for a single FBW prediction. They also clarify the dimensions and definitions in the equations, including H, s_imp, h_i, and the fact that the sliding window is applied after the AFS candidate sequence. These are important details and should be added to the final version.

    The baseline issue is still the main remaining weakness. I remain unconvinced that generic video models such as EchoNet-Dynamic and ViFi-CLIP are strong task-matched comparators for fetal blind-sweep FBW estimation, and I would be cautious about interpreting the margin over these methods. However, the authors’ rebuttal reasonably reframes these baselines as feasibility and transferability checks rather than definitive state-of-the-art FBW comparators. The clinically most important comparison is still against Hadlock-4, and the method improves over that reference while using a weaker, less operator-dependent acquisition setting. The lack of BabyNet/BabyNet++ or paired standard-plane DL comparisons is unfortunate, but the authors explain that paired standard-scan videos were not collected. I consider this a limitation rather than a fatal flaw, provided it is clearly acknowledged.

    Overall, I still view the methodological novelty as moderate: the paper combines foundation-model-based frame selection, redundancy-aware compression, and attention-based regression rather than introducing a fundamentally new learning paradigm. However, the problem formulation is novel and clinically relevant, and the pipeline is well matched to the structure of blind-sweep ultrasound data. The prospective dataset, strong empirical performance, clinical Hadlock comparison, and ablation results make the contribution worthwhile for MICCAI. My remaining concerns are mainly about scope, clarity, and future validation, not about the validity of the central contribution.

    After rebuttal, I would move from a weak accept to accept, assuming the authors incorporate the rebuttal clarifications into the final manuscript, temper claims about generalization, and clearly explain the limitations of the baseline comparisons and clinical use case.



Review #3

  • Please describe the contribution of the paper

    The paper addresses Fetal Birth Weight (FBW) estimation within 48 hours of delivery using “blind-sweep” ultrasound videos. The primary contribution is a workflow that eliminates the need for expert-acquired standard planes, potentially democratizing prenatal assessment. Key methodological components include: Anatomy-Guided Frame Selection - A foundation model-based approach for identifying keyframes from unconstrained video sweeps. Redundancy-Aware Feature Compression - A mechanism to aggregate and compress information across temporally similar frames. The authors demonstrate that this automated pipeline significantly outperforms traditional clinical formulas and existing deep-learning baselines.

  • 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 method achieves performance within the threshold of clinical acceptance, which is a significant milestone for automated FBW estimation.

    By utilizing blind sweeps rather than manually captured standard planes, the method addresses a major bottleneck in obstetric care, particularly in low-resource settings.

    The framework demonstrates substantial improvements over the clinically used Hadlock formula and other baseline models.

    The analysis in Figure 2 successfully validates that the Anatomy-Guided Selection focuses on relevant fetal biometry (e.g., abdomen, head), providing confidence in the model’s decision-making process.

  • 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.

    From a technical perspective, the paper largely integrates existing architectures and foundation models. The individual components (frame selection and feature compression) do not represent a significant departure from established computer vision techniques.

    The study would benefit from a more rigorous comparison against “gold-standard” inputs. Specifically, the authors compare against the Hadlock formula but do not compare against a deep learning model trained on the manually selected standard planes provided by the senior doctor or videos acquired during standard fetal US acquisition. This would clarify whether the “blind sweep” approach is simply a convenient alternative or a superior approach. The Authors should consider comparing their method with previous works in this field and discuss them e.g.:

    • Płotka, Szymon, et al. “BabyNet: residual transformer module for birth weight prediction on fetal ultrasound video.” International conference on medical image computing and computer-assisted intervention. Cham: Springer Nature Switzerland, 2022.- Płotka, Szymon, et al. “BabyNet++: fetal birth weight prediction using biometry multimodal data acquired less than 24 hours before delivery.” Computers in Biology and Medicine 167 (2023): 107602.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Have you compared the performance of a vision model trained on expert-selected standard planes against your blind-sweep method?

    Are there existing methods for automatic standard plane detection from blind sweeps that could be integrated into a FBW estimation pipeline for comparison?

    Figure 1 Labeling: There appears to be a discrepancy in the diagram labeling; the “tunable” label is applied to the Adapter, but the “frozen” label is not utilized, leaving the status of other modules ambiguous.

    To improve readability, please ensure that the mathematical symbols introduced in the text (e.g., in the feature compression section) are explicitly labeled or visualized within the corresponding figures.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a highly performant and clinically relevant application of Fetal US analysis. While the methodological novelty is somewhat incremental, relying on the combination of existing foundation model techniques, the successful application to FBW estimation from blind sweeps is a valuable contribution to the MICCAI community. The improvement over the Hadlock formula and the demonstration of clinical feasibility outweigh the moderate technical novelty.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    While the technical novelty remains limited, the successful application to FBW estimation from blind sweeps is a valuable contribution to the MICCAI community. I believe that referring to the earlier works related to the FBW estimation would be beneficial for the reader.



Review #4

  • Please describe the contribution of the paper

    The paper proposes a foundation model-driven framework for fetal birth weight estimation from blind-sweep ultrasound videos, eliminating the need for standard plane acquisition and expert operators. The core contributions are: (1) an AFS module that uses FetalCLIP to select anatomically informative frames via image-text similarity; (2) a RAFC module that removes temporally redundant frames using a composite score of task importance and local diversity; and (3) a Prior-Guided Attention Regressor that aggregates the refined frame features for final weight prediction.

  • 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 blind-sweep paradigm for FBW estimation is genuinely novel and addresses a real unmet need in low-resource obstetric care, where skilled sonographers are scarce. -The dataset is prospectively collected, large (839 patients, 5220 videos), multi-sonographer (10 operators), and uses post-delivery weight as ground truth -Leveraging FetalCLIP for anatomy-guided frame selection is a principled choice that avoids the need for explicit plane detection labels, making the approach training-efficient and scalable.

  • 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.

    -All 839 patients were scanned with a GE LOGIQ P9 system at one institution, raising concerns about generalizability across different ultrasound machines, probe types, and acquisition protocols. -The work is clinically innovative but in terms of pure methodology the contributions are more incremental than genuinely novel. The methodological merit lies primarily in the intelligent integration of these components into a coherent pipeline for a specific problem, rather than in the introduction of novel operators -Several baselines are methodologically misaligned with the proposed setting: EchoNet is a cardiac video model with no fetal relevance, ViFi-CLIP lacks any domain-specific pretraining, and USFM is evaluated on blind-sweep data despite being pretrained exclusively on standard planes (conditions that structurally disadvantage these competitors and artificially inflate the perceived margin of improvement)

  • 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 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?

    The paper addresses a clinically meaningful problem with a well-structured pipeline. However, the methodological novelty is incremental, the comparisons include structurally disadvantaged baselines, demographic bias is unaddressed, and single-center single-device validation limits generalizability claims.

  • 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 rebuttal satisfactorily addresses the main concerns. The clinical scenario is well-motivated, residual limitations (single-center validation, demographic bias) are acknowledged and appropriately deferred. I revise my score



Author Feedback

We thank all the reviewer (R) and meta-reviewers (MR) for reviewing and recognizing our work. Clarifications are provided to address concerns. Q1: Generalization (R1, R4, MR) Our protocol was designed for low-operator-dependence acquisition: operators received brief training, followed predefined blind-sweep trajectories, and did not view the screen, avoiding real-time standard-plane search. Most scans were performed by junior sonographers with <3 years of experience, and standardized probe orientation/sweep paths reduced operator variability. Although acquired by sonographers, the current protocol design shows good practical generalization potential. We will extend validation across centers, devices, and operator levels in future work. Q2: Clarifications (R1, R2, R3, MR) We will clarify the following ambiguities in the final revision. Prompts: We followed the predefined and fixed prompts provided in the public FetalCLIP repository. Implementation: In Eq. (6), H is the K x D feature matrix, s_imp is the K-dimensional vector, and h_i is the i-th feature in H. The sliding window is applied after AFS candidate sequence, with w=10.For each exam, six loops are processed independently; L features per loop are concatenated into 6L features for one FBW prediction. The hyperparameters were selected through validation-set sensitivity analyses. Data split and figure: The split was pregnancy-level to avoid leakage; exams exceed pregnancies because some women had up to two eligible scans. Fig.1 will explicitly label frozen/trainable modules and corresponding symbols. PC10/PC15: The supporting reference has been cited as [14], “ISUOG Practice Guidelines: ultrasound assessment of fetal biometry and growth” Q3: Comparison (R2, R3, R4, MR) 1) No prior work directly estimates FBW from blind-sweep ultrasound ( US ) videos. We therefore selected baselines to test feasibility under this weaker acquisition setting: Hadlock-4/INTERGROWTH-21st as clinical references, ResNet50+WAA/SelectGA as blind-sweep-related models, and the remainder as Foundation model baseline for transferability evaluation. These baseline groups are the most relevant to our study. We will clarify this rationale and avoid treating generic video models as task-matched FBW comparators. 2) Comparisons with BabyNet/BabyNet++ and standard-plane DL models are valuable. However, paired standard-scan videos required by BabyNet-like methods were not collected. Our study focuses on blind-sweep feasibility and uses sonographers Hadlock-4 as the clinical reference. We will collect paired data for future comparison. Q4: Novelty (R3, R4) To our knowledge, this is the first study to estimate pre-delivery FBW from blind-sweep US videos. Blind sweeps contain sparse fetal anatomy, noisy frames, and strong temporal redundancy, whereas FBW estimation requires reliable anatomical evidence. Our framework addresses this through three modules: AFS selects frames using fetal vision-language anatomical prompts; RAFC suppresses redundant candidates via local diversity, anatomical priors, and task-specific importance; and PGAR injects the selection prior into attention-based regression. We will clarify this anatomy-prior-driven positioning. Q5: Clinical scenario and selection bias (R2) Our target scenario is pre-delivery FBW assessment in term pregnancies, aiming to reduce reliance on expert standard-plane acquisition in low-resource settings. We clarify that delivery time was unknown at scanning; all women approaching delivery underwent blind-sweep US, and cases were included only if delivery fell within the predefined window. The cesarean-to-vaginal ratio was 341:498, indicating that data conforms to a typical clinical distribution. Q6: Ablation (R2) Table 3 validates the contribution of each module: AFS and RAFC both improve performance, and their combination performs best. More fine-grained studies on different backbones, prompt strategies, and selection heuristics will be explored in journal work.




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 paper got a positive feedback across all reviewers ( 1 x accept, 2 x weak accept and 1 x weak reject). However, reviewers raised some issues to be addressed: 1.Generalizability to real low-resource settings 2.Clarification 3.There are no comparison or mention with, perhaps, for the first time used deep learning-based methods for birh weight prediction - BabyNet (MICCAI 2022), BabyNet++ (CMB).

    I would suggest to address the issues raised by reviewers in rebuttal.

  • 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.

    Authors addressed previous concerns. However, as suggested by one reviewer and one meta-reviewer, the paper should mention and discuss, according to the authors’ response, related works to fetal birth weight prediction, BabyNet/BabyNet++, and highlight the difference between them and the proposed method.



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 positive overall, and after rebuttal the paper has support from all reviewers. The work addresses a clinically relevant and practically important problem: fetal birth weight estimation from blind-sweep ultrasound without requiring expert standard-plane acquisition. The proposed anatomy-guided frame selection, redundancy-aware feature compression, and prior-guided regression form a coherent framework tailored to the sparse, noisy, and redundant nature of blind-sweep data.

    The rebuttal satisfactorily clarifies several important points, including the low-operator-dependence acquisition protocol, pregnancy-level data splitting, use of fixed FetalCLIP prompts, fusion of multiple loops, and the clinical scenario for pre-delivery assessment. The authors also appropriately acknowledge remaining limitations, especially single-center, single-device validation and the lack of direct comparison to BabyNet/BabyNet++ or expert standard-plane deep learning pipelines due to unavailable paired standard-scan data.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    Reviewers are positive overall and agree that the paper addresses a clinically meaningful problem with a technically sound anatomy-guided framework for blind-sweep fetal birth weight estimation. The rebuttal clarified the clinical setting, data split, implementation details, and baseline rationale, and adequately addressed the main concerns.



back to top