Abstract

Fetal echocardiography screening videos are acquired under continuous probe motion, so imaging planes change over time and diagnostically decisive cues for Tetralogy of Fallot (TOF) may appear only transiently and only in specific views. Conventional video classifiers treat ultrasound clips as generic temporal signals, which often misses flash-by findings and fails under incomplete plane coverage and cross-hospital domain shift. We propose a plane-aware VLM framework for TOF screening that explicitly exploits view evolution as a key variable. First, plane-labeled frames are treated as keyframes and expanded with local neighborhoods to capture short-lived anatomical cues. Second, a vision–language model performs plane-aware checklist reasoning: TOF criteria are decomposed into view-appropriate, visually checkable questions, while cues that are not assessable in the current plane are marked as such to reduce over-interpretation. Third, a plane-wise memory bank stores normal/TOF prototypes and provides retrieval-based evidence via similarity log-likelihood ratios, enabling robust inference when plane labels are missing and improving generalization across hospitals. Case-level diagnosis fuses retrieval evidence with aggregated checklist cues, with plane-coverage modulation to upweight informative but under-represented views. Experiments on multi-center fetal echo screening videos show that the proposed method improves sensitivity and cross-site robustness compared with standard video classification and VLM baselines, while producing concise, auditable findings aligned with clinical screening practice.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2220_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LvXin_PlaneWise_MICCAI2026,
        author = { Lv, Xingguo AND Xu, Kai AND Liang, Bocheng AND Qiu, Zilang AND Wu, Jintao AND Zhao, Lei AND He, Guannan AND Huang, Chaoning AND Liang, Yuanhao AND Liang, Pengchen AND Pu, Bin AND Li, Kenli},
        title = { { Plane-Wise Retrieval Memory and Structural Evidence Fusion for Tetralogy of Fallot Video Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a plane-aware vision-language framework for diagnosing Tetralogy of Fallot (TOF) from fetal echocardiography videos. The method explicitly models the evolution of imaging planes and integrates plane-guided sampling, checklist-based VLM reasoning, and a plane-wise memory bank for retrieval-based evidence. The final decision is obtained through coverage-aware fusion, producing interpretable case-level diagnosis.

  • 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.

    S1.The paper addresses a clinically meaningful problem and aligns well with real-world fetal cardiac screening workflows. The modeling of plane transitions is particularly well motivated, as it reflects how clinicians interpret dynamic echocardiography rather than treating the video as a generic temporal signal.

    S2.The structured checklist-based reasoning is a strong design choice. By constraining the VLM to evaluate only view-dependent findings and explicitly handling “not assessable” cases, the method reduces hallucination and improves interpretability, which is critical for clinical adoption.

    S3.The integration of a plane-wise memory bank is well justified. It provides complementary evidence to the reasoning module and improves robustness in cases of incomplete plane coverage or missing annotations.

    S4.The experimental validation is strong, particularly the multi-center setup with external evaluation across hospitals with different acquisition parameters. Consistent performance across different hospitals significantly strengthens the practical relevance and credibility of the method.

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

    W1.The method relies on plane labels (or their estimation), and it remains unclear how robust the approach is when these labels are noisy or inaccurate. In practice, plane recognition itself can be error-prone, which may affect real-world performance. A more detailed quantitative analysis of the fallback path (using inferred plane posteriors) would strengthen the paper.

    W2.The checklist design introduces useful domain knowledge but may limit generalizability. Extending the framework to other cardiac conditions (e. g. , hypoplastic left heart syndrome, transposition of great arteries) would likely require manual redesign of the checklist. The authors should discuss the extensibility of this design.

    W3.The interaction between checklist reasoning and retrieval-based evidence is not fully analyzed. More detailed ablations or case-level analysis would help clarify their respective contributions and potential redundancy.

  • 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?

    This is a well-designed and clinically relevant work that integrates domain knowledge with modern VLM techniques in a principled manner. The plane-aware design is well motivated by clinical practice, and the multi-center evaluation with consistent cross-hospital performance provides strong evidence of generalizability. While there are some limitations regarding plane label robustness and checklist extensibility, they do not outweigh the overall impact and strong experimental validation.

  • 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.

    After reading the authors’ rebuttal, I maintain my positive assessment and recommend acceptance. The rebuttal adequately addresses the main concerns raised by all reviewers. The Acc inconsistency was transparently acknowledged as a calculation error, with corrected values that are consistent with the reported sensitivity and specificity. The clarification of the experimental protocol (patient-level splits, zero-shot external inference) resolves concerns about evaluation fairness. The explanation of plane labels as sparse anatomical annotations with a principled fallback mechanism is convincing. The External-3 sensitivity issue is reasonably attributed to domain shift, as all baselines degrade on this cohort. The clinical motivation remains strong, the multi-center evaluation with consistent cross-hospital performance is a significant strength, and the plane-aware design is well-grounded in clinical practice. While some reproducibility details (e.g., noisy-OR aggregation, coverage-weight notation) remain to be clarified in the camera-ready version, these are minor issues that do not affect the core contributions.



Review #2

  • Please describe the contribution of the paper

    In my view, the main contribution of this paper is a structured multimodal framework for the automatic diagnosis of Tetralogy of Fallot from fetal echocardiography videos. A notable aspect of the work is that the authors do not treat the problem as a purely end-to-end video classification task. Instead, they decompose it into clinically motivated components, including plane-aware keyframe selection, checklist-based reasoning over principal anatomical signs, and a memory-retrieval module that provides additional case-level support.

  • 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.Evaluation on internal and external datasets. A major strength of the paper is that the proposed method is evaluated not only on an internal dataset, but also on multiple external datasets. This is particularly important in medical imaging, where performance often degrades across institutions because of differences in acquisition protocols, operators, and patient populations. Evaluating across different centers strengthens the practical relevance of the study and is an important step toward demonstrating robustness.

    2.Clinically structured reasoning pipeline. Another positive aspect is that the method is designed around clinically meaningful intermediate steps rather than relying solely on direct end-to-end classification. The use of plane-aware processing and checklist-style reasoning makes the approach more structured and potentially more interpretable than a generic black-box classifier. This formulation is well aligned with the clinical nature of the task.

    3.Ablation study of the main components. The paper combines several components, including structured prompting and memory retrieval, and provides an ablation study to assess the contribution of each module. This is useful because it helps the reader understand which parts of the system are driving performance improvements, rather than presenting only a final aggregate result.

  • 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.Insufficient detail on the train/validation/test protocol. The paper does not describe the internal train/validation/test split with enough detail. It is unclear how cases were partitioned, whether the split was performed at the patient level, and how model development decisions were separated from final evaluation. These details are essential in a medical imaging paper to assess the reliability of the results and to rule out potential bias or leakage.

    2.The method is not described clearly enough for reproducibility. Several important parts of the pipeline remain insufficiently specified. In particular, the integration of the additional information sources is not explained clearly enough. From the current presentation, plane coverage appears to be incorporated at the fusion stage, while the memory bank seems to operate through a separate retrieval branch whose output is fused later at the case level. However, the manuscript does not explain this process with sufficient clarity. More generally, the exact inputs to the different modules, especially the memory/retrieval branch, are not described in enough detail to make the approach easy to reproduce or fully interpret.

    3.Concerning inconsistency among reported evaluation metrics. There appears to be a substantial mismatch between the reported accuracy and the reported sensitivity/specificity, especially for External 3.For a binary classification task, accuracy should satisfy the relation:

    Acc = pi * Sen + (1 - pi) * Spe

    where pi is the prevalence of the positive class.

    For External 3, the paper reports:

    Positive cases: 53 / 103, so pi = 53 / 103 = 0.5146 Sensitivity = 53.5% Specificity = 100.0%

    Using these values gives:

    Acc = (53 / 103) * 0.535 + (50 / 103) * 1.000 = 0.7607

    which corresponds to approximately 76.1%, whereas the paper reports 61.7%. This discrepancy is too large to be explained by rounding and should be clarified. Smaller inconsistencies also appear in External 1 (computed 88.51%, paper 86.2%) and External 2 (computed 89.24%, paper 87.3%). Because this issue directly affects confidence in the reported empirical results, it should be addressed explicitly.

    4.Unclear baseline setting in Table 2.It is not sufficiently clear whether the compared VLM baselines in Table 2 are evaluated in a zero-shot setting or after fine-tuning. This distinction is important for interpreting the results fairly, especially for strong open-source models that can benefit substantially from lightweight adaptation. Without this clarification, it is difficult to determine whether the proposed method is being compared against appropriate baselines.

    5.Unclear construction of the memory bank. The origin and construction of the memory bank are not described clearly enough. The paper states that the memory bank stores class-specific prototypes, but it is unclear how these prototypes are selected, whether they are built strictly from training data, and how leakage from validation or test data is avoided. Since this component may have a significant effect on performance, these details are important for methodological soundness.

    6.Limited clarity in the positioning of methodological novelty. The proposed combination of structured prompting and retrieval-based support is reasonable and potentially useful for this application. However, related ideas have already appeared in prior work, including Xia et al., MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Models (ICLR 2025) on retrieval-augmented medical VLMs and Zhang et al., Multimodal Chain-of-Thought Reasoning in Language Models (TMLR 2024) on structured multimodal reasoning. While these papers are not direct matches to the present setting, they are relevant for contextualizing the contribution, and the manuscript would benefit from a clearer explanation of what is genuinely new methodologically versus what is primarily a domain-specific adaptation and integration of existing ideas.

    Xia, P., Zhu, K., Li, H., Wang, T., Shi, W., Wang, S., Zhang, L., Zou, J., and Yao, H. MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Models. ICLR 2025.Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., and Smola, A. Multimodal Chain-of-Thought Reasoning in Language Models. Transactions on Machine Learning Research, 2024. 7.Presentation issues in Figure 1.Figure 1 appears to contain corrupted or improperly rendered symbols, which reduces clarity.

  • 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 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.

    (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 relevant clinical problem and presents a structured multimodal framework that is more clinically motivated than a simple end-to-end classification approach. The evaluation on internal and external datasets is also a positive aspect.

    However, my recommendation is mainly limited by three issues. First, there is a substantial inconsistency among the reported evaluation metrics, particularly for External 3, and smaller discrepancies also appear in other external datasets. This directly affects confidence in the correctness of the empirical results. Second, the method is not described with enough detail to support reproducibility or careful technical assessment, especially regarding the train/validation/test protocol, the integration of plane coverage and memory retrieval, and the construction of the memory bank. Third, the manuscript does not position its methodological contribution clearly enough relative to prior work on structured prompting and retrieval-augmented multimodal reasoning. Taken together, these concerns place the paper slightly below the acceptance threshold in its current form.

  • 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.

    My main concerns have been sufficiently clarified by the author response. The paper addresses an important and clinically relevant problem: Tetralogy of Fallot screening from fetal echocardiography videos, where diagnostic findings are transient, plane-dependent, and vulnerable to incomplete view coverage. The proposed framework is well motivated, combining plane-guided keyframe sampling, plane-aware checklist reasoning, and plane-wise retrieval memory to produce case-level diagnostic evidence.

    I previously had concerns regarding metric consistency, the train/validation/test protocol, baseline settings, memory-bank construction, and reproducibility. The rebuttal satisfactorily addresses the most important of these issues. In particular, the reported accuracy inconsistency appears to be a calculation/reporting error limited to the Acc column, while the primary metrics and case-level predictions remain unchanged. The authors also clarify that all trainable methods use the same internal training/validation protocol, that external cohorts are held out for zero-shot evaluation, that all splits are patient-level, and that the memory bank is built only from internal training data.

    Some limitations remain. The novelty over structured prompting and retrieval-augmented VLMs is moderate, and the sensitivity drop on the hardest external cohort should be discussed more carefully. The final version should correct the metric table, explicitly describe the split protocol and memory-bank construction, and better explain threshold calibration for high-sensitivity deployment.

    Overall, however, the paper presents a clinically meaningful and coherent framework, evaluates it across multiple external cohorts, and provides useful ablation evidence. I therefore recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a plane-aware VLM framework for prenatal TOF screening in fetal echocardiography videos. The core contributions are: (1) keyframe-centric sampling anchored on plane labels with local temporal neighborhoods; (2) a structured checklist reasoning protocol that decomposes TOF hallmarks into plane-routed queries; (3) a plane-wise memory bank providing log-likelihood ratio retrieval evidence for cross-site robustness.

  • 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 clinical motivation is well-grounded and the plane-aware design aligns closely with ISUOG guidelines. The zero-shot multi-center evaluation paradigm is rigorous. The ablation study systematically validates each module’s contribution, and the comparison encompasses 12 baselines. The structured reasoning traces further support clinical auditability.

  • 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.
    • Severe sensitivity degradation on External 3 is insufficiently discussed. The paper claims “we preserve balanced sensitivity (97.5%) and specificity (93.7%)”, yet this characterization applies only to the internal set. On External 3, sensitivity drops to 53.5% with specificity at 100%, implying that nearly half of TOF cases would be missed.
    • The source of plane labels is unspecified, undermining feasibility assessment. The method critically relies on plane labels $v_t$ for keyframe sampling and reasoning routing, yet the paper does not clarify whether these labels are obtained through expert annotation or an automated plane recognition model. If the former, how are labels acquired at deployment? If the latter, how do labeling errors propagate to downstream performance? Although Section 2.4 mentions a degraded inference mode for missing labels, neither its operational details nor its standalone performance are reported.
    • Critical implementation details of memory-augmented retrieval are missing, limiting reproducibility. Specifically, the paper does not describe how prototypes $m_{v,k}^{(y)}$ are constructed from the training set (e.g., whether clustering is used and of what type), the number of prototypes $K_y$ per plane-class pair, whether prototypes are updated during training, how the temperature $\tau$ is selected, or the explicit formulation of the noisy-OR aggregation $g_f$ in Eq. 5.- A potential contradiction exists between Eq. 3 and the corresponding text. As formulated, planes with smaller $c(v)$ receive lower weights in Eq. 5, yet Section 2.2 states that “evidence from under-represented but highly informative planes can be upweighted if present.” These two statements are semantically opposite; the authors should clarify the actual implementation.
    • Training of the VLM and visual encoder are not described separately. The Implementation section provides only a consolidated training description, without distinguishing the objectives, optimizers, and hyperparameters of Qwen3-VL-8B-Instruct (LoRA fine-tuning) from those of the visual encoder $\phi(\cdot)$ (contrastive loss training). Whether the two pipelines are jointly or sequentially optimized is likewise unspecified.
  • 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 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

    In the first paragraph of the Introduction, abbreviations such as LVOT, RVOT, and 3VT appear before their full forms are introduced.

  • 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 methodological direction is sound, clinical alignment is strong, and the cross-domain evaluation is persuasive. However, several issues limit the current submission: the 53.5% sensitivity on External 3 represents a serious deficiency for a screening application and is not adequately addressed; the absence of any explanation for how plane labels are obtained undermines practical feasibility; the training pipelines of the VLM and visual encoder are conflated; and the lack of an explicit definition for $g_f$ and related implementation details hinders reproducibility. Several of these concerns are amenable to clarification in the rebuttal.

  • 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.

    After considering the rebuttal, I maintain my reject recommendation. The authors clarified the corrected accuracy values and the general internal/external evaluation protocol, which I appreciate. However, several major concerns remain insufficiently addressed. In particular, the rebuttal still does not clearly specify how plane labels are obtained at deployment, nor does it report performance under missing or noisy plane labels, despite the method relying critically on them. Key implementation details of the plane-wise memory bank remain absent, including prototype construction, prototype number, temperature selection, update strategy, and the explicit noisy-OR aggregation. The contradiction between the coverage-weight formulation and the claimed upweighting of under-represented planes is acknowledged only as a notation issue but not technically resolved. The relationship between VLM LoRA fine-tuning and the visual encoder/contrastive retrieval training also remains unclear. Finally, the severe External-3 sensitivity drop is explained as domain shift, but no operating-point or calibration analysis is provided to support screening feasibility. Overall, the rebuttal improves clarity on some points but does not sufficiently resolve my concerns about reproducibility, methodological verifiability, and clinical robustness.



Author Feedback

We thank the reviewers and meta-reviewer for the constructive comments. We group the main concerns below.

1.Metric consistency (R2). We thank R2 for identifying the Acc inconsistency in Table 2.This was caused by an incorrect calculation of the Acc column. We have double-checked the case-level predictions, confusion matrices, and metric scripts for all cohorts. The corrected Acc values for External 1/2/3 are 88.5%, 89.2%, and 76.1%, respectively. Other metrics, including F1, Sen, Spe, Pre, and AUC, were computed correctly from the same case-level predictions and remain unchanged. Therefore, the main comparisons and the SOTA claim based on the primary metrics, especially F1 and AUC, are unaffected. We will correct Table 2, explicitly state the metric computation protocol, and release all code for reproduction.

2.Baseline protocol and reproducibility (R2, R3). We agree that the experimental protocol was under-specified. All trainable baselines and our method are adapted on the same internal training split, tuned only on the internal validation split, and then evaluated on the internal test set and the three external centers under zero-shot external inference. The external cohorts are never used for training, model selection, threshold tuning, or memory construction. All splits are performed at the patient level.

3.Plane labels and memory bank (R1, R2, R3). Plane labels are sparse standard-view annotations used to improve anatomical organization, not disease labels. Their main role during training is to build a plane-wise memory bank, enabling the model to learn plane-conditioned normal/TOF prototypes and better capture view-specific TOF evidence. They also serve as a routing signal for checklist reasoning: only clinically assessable TOF cues are queried in each plane, which reduces unsupported cross-plane speculation by the VLM. At inference, available plane labels guide keyframe sampling and checklist routing. When labels are missing or unreliable, the model falls back to plane posterior estimation from the same prototype memory and marginalizes retrieval evidence over candidate planes. The memory bank is built only from the internal training split and frozen for inference.

4.External-3 sensitivity (R3). External 3 is the most difficult cohort, with the largest acquisition mismatch, including higher FPS, longer average video duration, and more heterogeneous scanning dynamics. All baselines perform relatively poorly on this set, indicating a strong domain shift rather than a method-specific failure. Our method still achieves the best External-3 F1/AUC among the compared methods. The 100% specificity means that all normal cases in External 3 are correctly classified, suggesting that normal cases remain separable under the plane-wise memory despite the domain shift. The reduced sensitivity indicates that some TOF cases are conservatively classified as normal under the validation-selected threshold. We will revise the claim and discuss threshold calibration for high-sensitivity screening deployment.

5.Novelty and module interaction (R1, R2, R3). Our method is not a direct application of generic CoT or medical RAG. Generic CoT permits free-form reasoning, whereas our checklist is plane-valid: each TOF cue is queried only in clinically assessable views, and “not assessable” is treated as neutral evidence rather than negative evidence. Generic RAG retrieves external context, whereas our memory is a plane-conditioned visual prototype bank that produces class-wise log-likelihood evidence for case-level diagnosis. The key novelty is using view evolution as a diagnostic state variable and integrating plane-guided sampling, view-valid checklist reasoning, plane-wise retrieval memory, and coverage-aware fusion into a unified fetal echo screening framework. We will clarify retrieval fusion, add the noisy-OR aggregation, fix the coverage-weight notation, and repair the corrupted symbols in Fig. 1.




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 manuscript was reviewed by three expert reviewers. While acknowledging merits in the proposed work, the reviewers also raised some concerns about methodological novelty and validation. Such concerns should be addressed to strengthen the value of the work.

  • 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 reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns methodological applications in MIC.



Meta-review #2

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

    Reject

  • Please justify your recommendation.

    After rebuttal, reviewer most concerns related to reproducibility, experimental details have been clarified.

    The reliance on plane labels remains a limitation which would benefit from further analsysis, with reviewers suggesting further experiments measure robustness to plane labelling errors.

    Specifically: 1) plane-labels could explain significant portion of the improvement over baselines; 2) methodology suggests that memory bank uses posterior estimation when labels are not available, but there is no quantification of performance in scenarios where labels are available vs. when they are not. This is significant as baselines also do not have access to plane label information.

    As it stands, I believe this paper proposes an interesting and clinically relevant methodological contribution, but would benefit from further ablation experiments to fully validate it.



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.

    This is a clinically well-motivated and technically coherent paper. TOF diagnosis genuinely depends on transient, plane-specific fetal echo cues, and the plane-aware checklist plus retrieval memory directly addresses that. Overall I like this paper, but the authors should also pay attention to the high-sensitivity prompting strategy which may bias the model toward screening-oriented positives. Looking forward to seeing their future work!



back to top