Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Prenatal anomaly classification and localization is of critical importance for fetal health and pregnancy management. Although ultrasound (US) is the primary modality for prenatal screening, accurate diagnosis remains challenging due to the low prevalence and high heterogeneity of anomalies. Existing deep learning methods for prenatal tasks rely on large-scale annotated datasets, which are difficult to obtain in practice. Although few-shot learning alleviates data scarcity, it typically requires fine-tuning for new categories, limiting its practicality in resource-limited clinical settings. To address these challenges, we propose a training-free framework for multi-class prenatal US anomaly classification and localization that operates with only a few reference images per class, representing the first exploration of this setting. Our framework comprises three key components: (1) a memory bank with multi-granular prototypes that explicitly models both class-level semantics and anomaly characteristics; (2) a prototype-driven soft merging mechanism that aggregates discriminative features to detect the anomaly region; and (3) a class-aware refinement strategy that leverages prototype consistency to improve category prediction. Extensively validated on a multi-center prenatal US dataset containing 1,149 cases, with a total of 2,357 images and 9 categories, our proposed method outperforms the competitors.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1851_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/LL-AC/TFF

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiaHua_Prototype_MICCAI2026,
        author = { Liang, Huanwen AND Huang, Yuhao AND Zhu, Xiliang AND Zhang, Yuanji AND Deng, Xuedong AND Gao, Xinru AND Tao, Guowei AND Zhang, Yuhan AND Ni, Dong},
        title = { { Prototype Memory-Guided Training-Free Anomaly Classification and Localization in Prenatal Ultrasound } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a training-free framework for multi-class prenatal ultrasound anomaly detection. In the proposed method, a memory bank is constructed of multiple granularities. First, they use the CLS token to understand the coarse class identity of the disease. Then they extract embeddings of the bounding box where the anomaly lies and store it in their memory bank. They use DINOv3 as the backbone to extract features. Using this memory bank, the authors first shortlist top-M disease classes by comparing the CLS token of the new image with the feature vector of each class in their memory bank. For each of these shortlisted disease classes, the anomaly features are retrieved and slides against the new image to find patches with highest similarity. Top-S highest similarity patches are then taken and they are averaged out to get a centroid. Using this centroid, the authors re-score the entire image. Their intuition behind this is that rescoring with centroid helps aggregate the neighbouring patches that share features and suppress noise. The authors finally draw a bounding box around the patches lying in the top 5% anomaly scores. The proposed method outperforms known training-free methods as well as few shot object detection methods achieving mAP of 60.67.

  • 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 proposes a training-free method, which outperforms FSOD methods by a large margin (TRR-CCM generates mAP of 49.26, which is the second-best result after the authors’ proposed framework, which achieves an mAP of 60.67). The idea of having multi-granular memory bank is reasonable and intuitive. Scaling this method with more data can potentially provide better results and can be employed in a clinical setting.

  • 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 authors do not talk about the value of M and S in top-M and top-S selection for disease class and anomaly embeddings, respectively. The authors do not use a public dataset either. This makes it difficult to reproduce the results. The choice of M and S is extremely important since these are the driving hyperparameters in a training-free setting. I would encourage authors to discuss the methodological or empirical decision of choosing a particular value for M and S.

  • 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 provide sufficient information for 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 provides a training-free framework - which makes it easier to deploy in resource-constrained environments. Moreover, this does not come at a cost of accuracy, the authors achieve a drastic improvement over traditional methods. However, the choice of hyperparameters is not sufficiently discussed. That, along with not using a public dataset makes reproducibility rather limited.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper proposes a training-free framework for multi-class prenatal ultrasound anomaly detection that requires no task-specific fine-tuning and operates with only a few annotated reference images per category.

  • 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 training-free, few-shot approach directly addresses the real-world bottleneck of scarce annotated prenatal anomaly data, making it clinically practical.

    1. Class-level prototypes for coarse-grained recognition and anomaly-specific prototypes for fine-grained localization naturally align with the clinical “classify-then-localize” diagnostic workflow. 3.The multi-center dataset spanning 9 categories across three anatomical regions, combined with five random splits reporting mean and standard deviation, provides good diversity and statistical rigor.
  • 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 proposed method uses a frozen DINOv3 foundation model while competing FSOD methods use weaker backbones (e.g., ResNet-101). Without re-running baselines with the same backbone or ablating the backbone choice, the reported gains cannot be attributed to the proposed framework itself. 2.The 95th-percentile thresholding always highlights top-responding regions regardless of whether an anomaly exists, inevitably producing false detections on normal images (~29% of the dataset). The paper reports no false positive rate or specificity, which is a critical omission for a clinical screening tool. 3.The stated split (1,440 reference + 1,017 test) does not sum to the reported total (2,357), leaving 100 images unexplained. Additionally, it is unclear whether splits are constructed at the patient level; categories with few patients but many images per patient (e.g., single ventricle: 29 patients, 197 images) pose a high risk of data leakage that could inflate results.

  • 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.

    (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 addresses a clinically meaningful problem with an intuitive framework, but concerns regarding the fairness of backbone comparison, handling of normal cases, and dataset partitioning consistency need to be clarified before stronger endorsement can be given.

  • 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 my concerns well. It would be even better if the dataset could be made publicly available in the future.



Review #3

  • Please describe the contribution of the paper

    The main contribution of the paper is a training-free multi-class prenatal ultrasound anomaly detection framework that operates with only a small number of reference images per class. The method builds a multi-granular prototype memory bank using DINOv3 features, combining class-wise prototypes and anomaly-related prototypes, then performs candidate retrieval, anomaly-aware soft merging, and class-aware refinement to localize anomalies and predict their categories. The paper further validates the approach on a multi-center prenatal US dataset with 9 categories.

  • 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 paper addresses a highly important clinical problem. Prenatal ultrasound anomaly analysis is a low-incidence, high-heterogeneity setting in which large supervised datasets are difficult to build, so the few-reference training-free direction is practically relevant. Second, the dataset is a real strength. A multi-center, 9-category, 1,149-case prenatal US benchmark spanning brain, heart, and abdomen is meaningful and clinically diverse. Third, the method is internally coherent. The separation between coarse class semantics and fine anomaly prototypes, followed by soft merging and class-aware refinement, forms a clear and plausible pipeline. Fourth, the empirical results and ablations are encouraging. The paper reports 60.67 mAP at 32-shot, and the ablation indicates systematic gains from HR upsampling, ASM, and CR, with the full system improving performance by 15.49 mAP over the base variant.

  • 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 the paper is that the task framing is somewhat ambiguous. Although the work is presented as “training-free anomaly detection,” the actual setup is closer to a closed-set few-shot localization/classification problem: the method is given a fixed set of 9 predefined categories, each supported by annotated reference images with bounding boxes, and at test time it predicts one of these known categories by selecting top candidate classes and then choosing the most similar prototype. This differs from anomaly detection in the more standard open-set sense, where the emphasis is on detecting unknown deviations rather than recognizing one of several predefined abnormal categories. For this reason, the claim of being the first training-free “anomaly detection” method appears somewhat overstated.

    Second, the term “training-free” should be qualified more carefully. While the method does not optimize model parameters at test time, it still relies on curated reference images with both class labels and bounding-box annotations, and the memory bank is built directly from these annotations. The paper itself states that the framework relies on “a few annotated reference images.” Thus, this is not annotation-free or label-free in a strict sense, but rather training-free inference with annotated references.

    Third, the methodological novelty appears limited. The use of DINOv3 features, cosine similarity, class-wise prototypes, anomaly-related prototypes, candidate retrieval, and refinement is sensible and effective, but it is also conceptually close to prior few-shot detection and training-free prototype-based medical vision methods. The paper itself compares against DeFRCN, DiGeo, SMILe-FSOD, TRR-CCM, ProtoSAM, and MAUP, which reinforces the impression that the contribution is primarily a well-engineered adaptation/system integration for prenatal ultrasound rather than a fundamentally new formulation.

    Fourth, the dataset and split protocol are under-specified. The submission uses a private multi-center dataset with 1,149 cases and 2,357 images, and describes five reference splits plus the remaining 1,017 test images, but it does not clearly state whether the split is strictly case-level or merely image-level, whether multiple images from the same case can appear across reference and test sets, or how center-wise generalization is handled. In ultrasound, near-duplicate images from the same case can materially inflate performance, so this issue matters. The class distribution is also imbalanced, with categories such as single ventricle being relatively small.

    Fifth, the few-shot claim is not fully substantiated. Although the paper mentions 4-, 8-, 16-, and 32-shot settings, the main quantitative table is centered on the 32-shot regime, and 32 reference images per category may not be especially low-shot in many practical contexts. Since the paper emphasizes the few-reference setting in the title and abstract, the empirical case would be much stronger if it demonstrated more extreme low-shot robustness. Based on the visible results, the evidence for truly scarce-reference performance remains somewhat limited.

    Sixth, the evaluation remains somewhat narrow. The main metric is mAP@0.5, and while the paper includes qualitative examples and ablations, it does not report more comprehensive evaluations such as AP across multiple IoU thresholds, per-center generalization, inference time, calibration, or threshold sensitivity. This is especially relevant because the method converts the anomaly map into a final localization via a fixed 95th-percentile threshold, yet the sensitivity of this choice is not analyzed. In addition, ProtoSAM and MAUP are segmentation-oriented training-free baselines, whereas the proposed method is more naturally a box-level localization/classification framework, so the fairness of these comparisons could be better justified.

  • 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.

    (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?

    I find the paper valuable because it addresses a clinically important prenatal ultrasound problem in a low-annotation, training-free setting, and the multi-center dataset is a meaningful asset. The reported 32-shot results are clearly stronger than both FSOD and training-free baselines. My main concerns are that the methodological novelty is closer to a careful integration of existing prototype/foundation-model ideas than a fundamentally new formulation, and the “few-shot” claim would be more convincing with more extreme low-shot experiments. For these reasons, I place the paper slightly above threshold as a Weak Accept.

  • 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.

    I maintain my final decision as Accept. The rebuttal substantially addresses my main experimental concerns by clarifying the patient-level split and adding lower-reference results, threshold sensitivity, balanced accuracy, mAP@0.5:0.95, inference time, and hyperparameter details. My remaining concerns are that the task is closer to anomaly classification/localization than to conventional anomaly detection, and that the novelty is more about strong system integration than a fundamentally new formulation. Nevertheless, given the clinical importance of the problem, the meaningful multi-center prenatal ultrasound setting, and the practical value of the few-reference training-free approach, I now view the paper as slightly above the acceptance threshold.



Review #4

  • Please describe the contribution of the paper

    This paper proposes a training-free framework for multi-class prenatal anomaly detection in ultrasound, built around a multi-granular prototype memory bank, an anomaly-aware soft merging mechanism, and a class-aware refinement strategy. The method operates with only a few reference images per class and requires no fine-tuning, leveraging DINOv3 as a frozen backbone. Evaluated on a private multi-center dataset of 1,149 cases across 9 categories, the proposed approach outperforms both few-shot and training-free 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 training-free setting is well-motivated and practically relevant, particularly given the low incidence of prenatal anomalies and the difficulty of assembling large annotated datasets in clinical settings.

    • Statistical significance (p < 0.05) is reported in Table 1.
    • The paper is clearly written and well-structured, with a logical progression from motivation through methodology to experimental validation.
  • 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 method is evaluated using a single backbone (DINOv3, Siméoni et al., 2025, under review at TMLR), which is a reasonable choice given recent evidence of its strong medical imaging performance. However, neither the motivating benchmark (Liu et al., 2025) nor the DINOv3 reference directly addresses anomaly detection in ultrasound specifically. Investigating at least one alternative backbone would help clarify whether the performance gains stem from the proposed framework or the underlying feature extractor.

    • The ablation study reports standard deviations across splits, which is appreciated. Extending it to include additional metrics beyond mAP, such as localisation accuracy or numerical per-category breakdowns, would provide a more complete and convincing picture of each component’s individual contribution, same for the general comparison in Tab. 1.

    • The dataset is entirely proprietary (as far as the reviewer understood), with no public benchmark evaluation included. While the authors mention that code will be released upon acceptance, the absence of a publicly available dataset makes it difficult for the community to reproduce or build upon this work, and evaluation on at least one public benchmark would be a valuable addition.

    • The main quantitative comparison in Table 1 is conducted only under the 32-shot setting, while results across other shot settings (4, 8, 16) are presented only in aggregate in Fig. 3 without per-method or per-category breakdowns. Including a fuller comparison across shot settings would give a more transparent picture of relative performance under lower-data regimes.

    • The 95th percentile threshold used for anomaly localisation appears to be fixed, and it would be helpful to include a brief sensitivity analysis or some practical guidance on how this parameter should be set across different categories and clinical scenarios.

    • The baseline comparisons would benefit from including more recent (ultrasound) foundation models as backbones and ultrasound-specific anomaly detection methods to more fully situate the contribution within the current landscape.
  • 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

    • Evaluating with at least two alternative (medical/US) backbones would help disentangle the contribution of the proposed framework from the underlying feature extractor.

    • Reporting additional metrics not only in the ablation study but in Tab. 1, and providing numerical per-category results would give a clearer picture of each component’s contribution.

    • Evaluation on at least one publicly available benchmark is strongly encouraged to enable reproducibility and fair comparison with future work.

    • A sensitivity analysis on the 95th percentile threshold, and guidance on how this parameter should be set in practice, would strengthen confidence in the localisation strategy.

    • Please revise Tab. 1, the formatting is suboptimal.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the paper addresses a clinically relevant problem and the core idea is interesting, the evaluation does not yet meet MICCAI standards. The reliance on a single evaluated backbone without sensitivity analysis, the absence of a public benchmark (dataset), and the sparse comparisons with mAP only and only ~1 US-specific baseline model collectively limit confidence in the reported results. The authors are encouraged to address these issues and resubmit to, e.g., a MICCAI workshop or journal.

  • 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.

    I maintain my Reject recommendation. The two core issues remain unresolved in the rebuttal: the ~11 mAP gain cannot be attributed to the framework rather than the feature extractor without a unified-backbone comparison in Table 1, and a single private dataset with no public benchmark is insufficient to establish generalisability for a methodological contribution at MICCAI. The rebuttal’s new results cannot substitute for what is in the submission.



Author Feedback

We thank all the reviewers (R) for reviewing and recognizing our work. We provide explanations to address the comments.

Q1.Fairness of comparison(Meta R, R2, R3) (1) Without changing the core idea of each method, we unified all replaceable backbones with that used in our framework to ensure a fair comparison. (2) To the best of our knowledge, training-free methods closely related to our task remain limited, but segmentation-oriented methods share similarities with our framework. Specifically, the generated masks correspond to anomaly regions and can be used for bounding box generation.

Q2.Performance in lower-reference setting(Meta R, R3, R4) Due to space limitations, we present the category-wise results in the 32-reference setting to analyze category bias, while Fig. 3 shows performance in different settings. The performance in lower-reference (4, 8, 16) settings is 29.34±1.83, 37.05±1.13, and 52.02±0.44, respectively. We will improve Fig. 3 in the final version to better illustrate the performance in lower-reference settings.

Q3.Dataset split and experimental details(Meta R, R1, R2, R3) Sorry for the confusion. (1) The inconsistency in the dataset description is due to a writing error; the test set contains 917 images. (2) The dataset is split at the patient level. We randomly sample patients without replacement; once the accumulated number of images exceeds 32, we randomly select 32 images to form a fold and remove these patients from the pool. We repeat this process to obtain 5 folds as reference sets, and use the remaining patients as the test set. (3) For hyperparameters, we set M = 2 (since M=1 disables class refinement). S varies from 1 to 32 in ablation studies, with performance variation within 3.68%. Due to space limits, results are not reported. We use S = 16 in the paper. We will correct the errors and add details in the final version.

Q4.Validation of the fixed thresholding strategy(Meta R, R2, R3, R4) (1) Following [25], we use a 95th-percentile threshold to generate anomaly regions. Sensitivity analysis across thresholds of 90-95 yields a standard deviation of 4.42%. (2) Due to the thresholding strategy, normal images also yield anomaly regions, which are refined to the normal class by the CR module. The false positive rate in different settings is 20.01±5.56, 16.05±4.53, 11.85±2.49, and 6.90±0.78, respectively.

Q5.More evaluation metrics(Meta R, R3, R4) Our framework supports both classification and localization analysis. Since localization metrics are built upon classification and IoU, localization performance can also reflect classification results. Due to space limitations, we mainly report AP. The Balanced Acc. across the 9 categories under different reference settings is 61.95±5.08, 66.96±4.95, 75.45±3.21, and 83.19±2.61, respectively. Moreover, mAP@0.5:0.95 is ~20% lower than mAP@0.5.The inference time is ~3 seconds per image.

Q6.Public dataset and reproducibility(R1, R4) To the best of our knowledge, public datasets for prenatal US anomalies with localization annotations are still limited. We will discuss with our partner hospitals the possibility of releasing a dataset to facilitate reproducibility.

Q7.Novelty(R3) Our study proposes a clinically oriented training-free framework that can be easily deployed in resource-constrained environments. Built on memory retrieval and class refinement, we introduce an anomaly-aware soft merging mechanism to reduce feature-level ambiguity and enhance region-level attention. The training-free, few-reference design addresses the scarcity of prenatal anomaly data, making the method practically valuable for clinical applications.

Q8.Task definition(R3) In our paper, “anomaly” denotes malformations, and the task focuses on anomaly classification and localization rather than conventional anomaly detection (AD). The definition of training-free follows [25,11,23]. We will further clarify the task framing in the final version.



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.

    Reviewers agreed that this paper addresses a clinically important problem and presents a practical training-free framework for prenatal ultrasound analysis with promising empirical performance. However, the methodological contribution appears somewhat limited, with parts of the framework viewed more as a careful integration of existing ideas rather than a fundamentally new formulation. In addition, reviewers raised concerns regarding the experimental design, including the reliance on a single DINOv3 backbone and potentially unfair comparisons with competing methods implemented using substantially weaker architectures (e.g., ResNet-based backbones), which makes attribution of the reported gains less clear.

    The rebuttal should address the following points, including but not limited to: 1.Fairness of backbone comparison. Only a single backbone (DINOv3) is evaluated for the proposed method, while competing methods rely on weaker architectures. It remains unclear to what extent the reported gains are attributable to the strong feature extractor versus the proposed framework itself.

    2.Performance in lower-reference settings. The paper emphasizes a few-reference setting in the title and abstract, yet the quantitative evaluation is primarily centered on the 32-shot setting. More comprehensive results under more extreme low-shot regimes would strengthen this claim.

    3.Dataset split and experimental details. The reported sample counts appear inconsistent and should be clarified. In particular, it is important to specify whether the split is conducted at the case/patient level, since near-duplicate images across reference and test sets could inflate performance. Key hyperparameter choices should also be better justified.

    4.Validation of the fixed thresholding strategy. Several reviewers raised concerns regarding the fixed 95th-percentile threshold used for anomaly localization, including the lack of sensitivity analysis, limited practical justification, and missing evaluation of false-positive behavior on normal cases (e.g., specificity / false positive rate).

    5.More evaluation metrics. The experimental evaluation could be strengthened with more comprehensive metrics and analyses, beyond mAP@0.5 alone, to better establish robustness and clinical applicability.

  • 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 rebuttal addressed many main concerns and resulted in an overall positive post-rebuttal balance. In particular, the response clarified the patient-level split, lower-reference performance, threshold sensitivity, false positive behavior, additional evaluation metrics, and key implementation details, and most reviewers considered their concerns sufficiently resolved. One reviewer still raises a valid concern that the reported gains cannot yet be cleanly attributed to the proposed framework rather than the backbone, and that the lack of a public benchmark limits the strength of the generalizability claim. At the same time, the other reviewers also considered the paper from an application perspective, rather than only from a strict methodological standpoint, and found the paper is clinically meaningful and useful. Taking these perspectives together, I will support the paper is slightly above the acceptance threshold.



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 paper addresses an important and clinically meaningful problem: prenatal ultrasound anomaly localization under scarce annotations. The proposed training-free prototype-memory framework is practical, coherent, and provides a useful methodological contribution for data-scarce prenatal ultrasound settings. Beyond the method itself, the multi-center dataset, patient-level split clarification, and consistent empirical gains provide a reasonable level of clinical/application validation for a MICCAI submission. The rebuttal clarifies important concerns, including dataset-count inconsistency, hyperparameter choices, lower-reference performance, threshold sensitivity, and false-positive behavior. The ablation study further supports the contributions of HR upsampling, anomaly-aware soft merging, and class-aware refinement. While novelty is moderate and backbone attribution could be clearer, I 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.

    Accept

  • Please justify your recommendation.

    Three reviewers recommend acceptance. The framework addresses a clinically meaningful problem in a practically motivated setting, the multi-center dataset spanning nine categories is a genuine asset, and the rebuttal adequately addressed the main concerns around dataset split integrity, hyperparameter transparency, and threshold sensitivity. One limitation noted by R4 should be acknowledged in the final version. The primary comparison in Table 1 is conducted against FSOD baselines using weaker backbones such as ResNet-101, while the proposed method uses DINOv3.The rebuttal states that backbones were unified where possible, but this is not clearly reflected in the submitted paper, making it difficult to attribute the reported gains unambiguously to the proposed framework rather than the underlying feature extractor. The final version should make the backbone comparison explicit.



back to top