Abstract

Blind sweep ultrasound protocols can expand access to antenatal screening in low- and middle-income countries (LMICs), but their clinical utility depends on whether a sweep contains diagnostically useful views. In routine obstetric imaging, gestational age is estimated from fetal biometry measured on standardised planes (e.g. head and abdominal circumference). In blind sweeps, these planes may appear only briefly and under large appearance variation, making manual review burdensome and automated processing unreliable without quality screening. We propose biometry plane detection as an application-driven quality assessment task: given a blind sweep video, identify frames suitable for femur length, head- and abdominal-circumference biometry. We formulate the problem as visual-query-guided retrieval in long ultrasound videos and introduce a spatio-temporal transformer that matches a canonical query image to candidate frames. We show that we can leverage this query with an ultrasound-specific curriculum to learn effectively under extreme label sparsity. To transfer knowledge from a large freehand dataset to sparsely labelled sweeps, we incorporate scale-robust visual encoding and domain-adversarial training. Experiments on a participant-disjoint blind-sweep test set demonstrate improved detection performance over image-based and video-based baselines, while highlighting the remaining challenge posed by ultra-low prevalence and incomplete frame-level annotation. We hope this work supports adoption of blind-sweep fetal ultrasound by enabling automatic screening and rapid re-acquisition when biometry content is absent. Project page: https://github.com/kwon-j/vqcl.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4097_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/kwon-j/vqcl

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KwoJon_VisualQuery_MICCAI2026,
        author = { Kwon, Jong AND Jiao, Jianbo AND Self, Alice AND Papageorghiou, Aris T. AND Noble, J. Alison},
        title = { { Visual-Query Curriculum Learning for Rare Fetal Biometry Plane Detection in Blind-Sweep Ultrasound Video } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This is an application-driven paper describing a framework for detecting planes in “blind sweep” ultrasound videos. The proposed method poses this as a visual query selection problem.

  • 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 is an important practical application that is not well studied in the literature. The results are compared against several baselines and there is a thorough ablation study of the components.

  • 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.

    There is not a lot of information given on how the sweep data were acquired, which is important because these are not usually gathered. Where were they acquired, what gestational age, what scanner, etc?

    The motivation for formulating this problem using a visual query task is poorly justified in an unconvincing, informal, and imprecise way “Without queries, the model task is optimising too many competing losses at once – the model is trying to learn semantics of this is what general AC looks like vs FL vs HC vs background, at the same time as learning to find the frame temporally out of n frames”. It is not clear to me what these competing losses are, nor how the visual query task avoids this problem.

    The design of the method feels rather over-complicated: a set of disconnected components thrown together with often unclear motivations. However, the results do seem to be strong.

  • 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

    The authors state that the focal loss is applied “between the temporal binary predictions”. What does this mean, between predictions and what? Between the predictions and the labels? If so, how is that a temporal loss?

    This is a multi-class problem. What exactly are the frame-level AUROC (and AUPRC)? Are they macro or micro-averaged over classes? This also hides a lot of important detail about how the miscategorisations break down by label.

  • 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 compelling application and strong results, but this is tempered by unclear motivation for various design decisions and lack of clarity/precision in the technical writing.

  • 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 clarified several important points and it is feasible that they are addressed in the revision. This improves my assessment of the paper. However there are still important concerns regarding dataset size, so overall I still rate this within the weak accept bracket



Review #2

  • Please describe the contribution of the paper

    This paper introduces biometry plane detection in blind-sweep fetal ultrasound as an application-driven quality assessment task and proposes a visual-query-guided spatio-temporal transformer to address it. The core idea is to frame detection as query-guided retrieval in long videos, where a single exemplar image of an HC, AC, or FL plane guides the model to locate matching frames. To enable learning under extreme label sparsity (<0.1% positive frames), the authors introduce an ultrasound-specific curriculum that progressively trains the model from simple pixel-matching (same-video query) to semantic matching (low-quality, cross-patient query). The method further incorporates scale-robust visual encoding via Swin-FPN with heavy zoom augmentation and domain-adversarial training to transfer knowledge from abundant freehand standard plane data to sparsely labeled sweep videos. Experiments on 115 blind-sweep videos show consistent improvements over image-based and video-based baselines, achieving AUROC of 0.973 for HC and 0.979 for AC detection.

  • 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.

    Addresses a clinically important and underexplored problem: automated quality assessment of blind-sweep fetal ultrasound for LMIC settings with clear real-world impact.

    The visual-query-guided formulation is a creative way to enable video transformers to learn under extreme label sparsity (<0.1% positive frames).

    The curriculum learning progression (same-video query → augmented query → high-quality standard plane → low-quality biometry plane) is well-motivated and bridges pixel-matching to semantic matching.

    The method is complementary to existing end-to-end GA estimation pipelines, adding practical value as a preprocessing quality gate.

  • 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.

    Extremely small dataset: only 115 patients with 248 annotated biometry frames. FL has only 53 frames. No external validation is provided, making generalization claims unsupported.

    Misleading metric interpretation: AUROC appears strong (0.97+) but AUPRC is very low (0.088–0.135). The paper acknowledges this but does not adequately discuss the clinical implications, at such low precision, the tool would generate many false alarms in practice.

    Flawed human-expert comparison: the dataset has annotation missingness (sonographers picked only 1–2 planes per patient, not all usable planes). Many model “false positives” are clinically usable (confirmed by GA estimates within 14 days), making the F1 comparison unreliable.

    Missing key baselines: no comparison with recent visual-query methods like MCAT (AAAI 2025), and no comparison with specialized fetal ultrasound plane detection methods (e.g., SonoNet variants, anatomy-aware contrastive learning).

    Coarse ablation of curriculum learning: the paper shows curriculum is essential for convergence (0.550 without it), but does not ablate individual stages. It is unclear which stages matter most or whether simpler alternatives could work.

  • 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?

    The paper tackles a meaningful problem and the core idea of visual-query-guided retrieval under extreme sparsity is interesting. However, the empirical evidence is not yet strong enough for acceptance. The lack of external validation, very low AUPRC, and flawed human-expert comparison make it difficult to assess real-world effectiveness. The absence of comparison with directly related recent work (e.g., MCAT) further weakens the novelty claim.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    I appreciate the authors’ efforts in addressing previous concerns during the rebuttal phase. While the proposed visual-query-guided method for fetal ultrasound quality assessment is interesting and addresses a relevant clinical problem, I remain concerned that the manuscript is not yet mature enough for publication in its current form. There is significant room for improvement, particularly regarding experimental validation and presentation clarity. Consequently, I believe the paper is not ready for acceptance at this stage and would benefit from further strengthening before being submitted to a comprehensive journal such as TMI.



Review #3

  • Please describe the contribution of the paper

    The paper formulates biometry plane detection in blind-sweep fetal ultrasound as a visual-query-guided retrieval problem. The proposed approach leverages a spatio-temporal transformer that matches a canonical query image to candidate frames in long ultrasound videos. To address extreme label sparsity (<0.1% positive frames), the authors introduce a curriculum learning strategy combined with domain adaptation and contrastive learning. The method demonstrates improved detection performance over image-based and video-based baselines, highlighting the importance of inductive bias in ultra-sparse regimes.

  • 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.

    First, the problem formulation is novel and well motivated. Recasting biometry plane detection as a retrieval task using visual queries is conceptually strong and aligns well with how clinicians recognize anatomical structures.

    Second, the paper explicitly addresses extreme class imbalance, which is central to blind-sweep ultrasound. The use of focal loss, clip sampling, and balanced batching is appropriate.

    Third, the curriculum learning strategy is well designed, progressing from trivial matching to semantically meaningful retrieval across patients and quality levels.

    Fourth, the integration of multiple datasets (sweep, freehand, and standard planes) is a strong idea that enables transfer learning in a low-data regime.

    Finally, the experimental results demonstrate consistent improvements over several baselines, particularly highlighting the limitations of standard video models without strong inductive biases.

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

    (A) Dataset clarity: The composition and usage of the three datasets (sweep, freehand, and standard planes) are insufficiently described. It is unclear how data splits are performed, whether there is subject overlap, and how domain differences are handled during evaluation.

    (B) Annotation noise: The authors acknowledge that the dataset contains missing annotations, meaning some model ‘false positives’ may actually be valid biometry planes. This undermines the validity of reported metrics such as AUPRC and F1-score, yet no correction or sensitivity analysis is performed.

    (C) Query dependency: The model requires a manually selected query image at inference time. This introduces bias and limits full automation, which is critical for real-world deployment in low-resource settings.

    (D) Curriculum learning reliance: The model fails to converge without curriculum learning, suggesting instability and heavy dependence on training heuristics. This weakens the methodological robustness.

    (E) Evaluation limitations: Metrics such as sensitivity and recall are not adequately reported, despite extreme class imbalance. AUPRC is used but is highly unstable in this regime, as acknowledged by the authors.

    (F) Result reporting: The Results section includes subjective interpretations and informal language (e.g., ‘we believe’, ‘surprisingly’).

    (G) Model complexity: The architecture is complex relative to the extremely small number of positive samples, raising concerns about overfitting and generalization.

    (H) Domain adaptation: While domain-adversarial training is included, there is no clear evaluation of its effectiveness or analysis of domain shift.

    (I) Visualization: The paper relies heavily on tables and lacks informative visualizations such as detection timelines, attention maps, or similarity heatmaps.

    (J) Reproducibility: Key implementation details are missing, including curriculum scheduling, query selection strategy, and precise data sampling procedures.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The submission 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

    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?

    The paper proposes a strong and novel idea with clear clinical relevance. However, weaknesses in dataset clarity, evaluation methodology, and reliance on training heuristics prevent it from reaching a higher level of confidence. With improvements, the work has the potential to make a meaningful contribution.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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

We thank the AC and reviewers. We agree the current paper should be positioned as a proof-of-concept on a difficult real-clinical blind-sweep setting, not as deployment-ready validation.

R1/R2/R3 (data acquisition, splits, dataset composition). We will clarify the three datasets and their roles: blind sweeps are used for final evaluation; freehand videos provide auxiliary video supervision; standard-plane images are used only for encoder pretraining. We will explicitly state the blind-sweep split is patient-disjoint (test: 23 participants, 57 annotated planes, ~80k frames) and add acquisition details omitted for space (GE Voluson E8, 18-28 weeks, sonographer-acquired after routine obstetric scanning). We will also soften any generalization claims and state that external validation remains future work.

R2/R3 (missing labels, AUPRC, human comparison). We agree that strict frame-level AUPRC/F1 underestimate performance because positives are incomplete: sonographers marked one/few usable planes rather than all plausible ones. For this reason, the current “human performance” row should be interpreted only as inter-reader agreement, and we will rename/de-emphasize it. We report AUROC to assess model’s ability to rank planes vs background and AUPRC transparently shows the difficulty of the ultra-low-prevalence task.; Table 1 also fixes sensitivity at 0.75 and reports TNR at that operating point. The intended clinical decision, however, is sweep-level screening (“does this sweep likely contain usable biometry content?”), not frame-by-frame alarming. We will therefore add sweep-level recall/specificity and report class prevalence alongside AUPRC (random baseline = prevalence) so readers can interpret the AUROC-AUPRC gap appropriately. We will also add a sensitivity analysis counting second-reader positives as valid.

R1/R3 (why query, why curriculum, why this complexity). We agree the current wording around “competing losses” is imprecise and will revise it. The key motivation is to simplify the learning problem under extreme sparsity: without a query, a temporal model must jointly learn (i) heterogeneous anatomy appearance across AC/HC/FL/background and (ii) temporal localization from very few positives. The query-based curriculum first gives an easier similarity/localization signal, then progressively harder cross-patient queries to encourage anatomy-level semantics. The first stage uses the same positive frame as query, so the model can learn temporal localization from simple pixel matching, before learning diverse anatomy appearance. This is supported by our ablations: removing the query drops mean AUROC/AUPRC from 0.976/0.111 to 0.891/0.045, removing domain discrimination to 0.959/0.068, removing query-similarity to 0.956/0.081, and removing curriculum prevents convergence. At inference, queries are not manually chosen per sweep: we use a fixed bank of three canonical AC/HC/FL exemplars, so the method remains fully automatic.

R2 (baselines/related work). We agree MCAT/STAN-LOC are relevant and should be discussed. Our reason for not including them is that they localize temporally extended clips via start/end boundary prediction, whereas our targets are often single/few-frame off-axis planes with incomplete frame labels; consistent with this mismatch, our own start/end KL-loss ablation failed to converge (Table 2). STANLOC’s query-selection module is not directly applicable: fetal-heart scans are region-constrained sweeps where query appearance can be inferred, whereas our sweeps give no indication of what AC/HC/FL planes should look like before the plane appears. We will add this discussion explicitly and better clarify that the VGG16 baseline uses anatomy-aware contrastive pretraining from Fu et al.

R1 technical clarifications. We will correct “temporal binary predictions” to “per-frame binary predictions vs labels”, state explicitly that AUROC/AUPRC are one-vs-rest per anatomy, and replace subjective wording in the Results.




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.

    This is borderline. Two weak accept and one weak reject, with agreement that the problem is important and that the query-based formulation is interesting, but also clear concerns about whether the current evidence is strong enough.

    The biggest issue is evaluation. The dataset is very small, labels are sparse and incomplete, and there is no external validation. This makes it hard to judge generalization or real-world usefulness. The low AUPRC despite high AUROC is particularly important and needs a much clearer clinical interpretation.

    There are also unresolved concerns about dataset description, annotation missingness, and how this affects both the reported metrics and the human comparison. Right now, those results are difficult to interpret reliably.

    Re method, the idea is promising, but the paper needs to better justify the visual-query formulation, explain the role of the curriculum more precisely, and clarify the overall design, which currently feels more complicated than necessary to some reviewers.

    So I would send this to rebuttal. The authors should focus on clarifying data acquisition and splits, addressing missing labels and metric reliability, explaining the AUROC versus AUPRC gap, and justifying the key design choices and missing baselines.

  • 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 main concerns are the small dataset, incomplete frame-level annotation, lack of external validation, and low AUPRC despite high AUROC. The rebuttal clarifies that the blind-sweep evaluation is patient-disjoint, that standard-plane and freehand data are used only for auxiliary training/pretraining, and that the task should be interpreted as proof-of-concept rather than deployment-ready validation. The authors also clarify that the low AUPRC is partly a consequence of extreme class imbalance and incomplete positive labels, and they appropriately propose to de-emphasize the human-comparison row and add sweep-level metrics and prevalence context.

    The camera-ready version should clearly describe dataset composition, acquisition, splits, and annotation missingness; temper all generalization claims; report per-class AUROC/AUPRC with prevalence and sweep-level recall/specificity; clarify the fixed query-bank inference setup; and discuss why missing labels limit interpretation of frame-level F1 and AUPRC.



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.

    All reviewers agree that this work proposes an interesting solution to an important problem. After rebuttal, the main concern of R2 remains that the empirical evidence is not yet strong enough, while R1 and R3 weigh the interesting method and application more. I would follow their opinion and recommend acceptance.



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.

    The reviewers raised several key conerns before and post-rebuttal, including:

    1.Extremely small dataset:. No external validation is provided, making generalization claims unsupported.

    2.Misleading metric interpretation: AUROC appears strong (0.97+) but AUPRC is very low (0.088–0.135). The paper acknowledges this but does not adequately discuss the clinical implications, at such low precision, the tool would generate many false alarms in practice.

    3.Flawed human-expert comparison: the dataset has annotation missingness. Many model “false positives” are clinically usable (confirmed by GA estimates within 14 days), making the F1 comparison unreliable.

    4.Missing key baselines: no comparison with recent visual-query methods, and no comparison with specialized fetal ultrasound plane detection methods



back to top