List of Papers Browse by Subject Areas Author List
Abstract
Congenital heart disease (CHD) is the most common birth defect. Many CHD phenotypes are individually rare, which limits labeled data and motivates anomaly detection for multi-modal fetal echocardiography screening. Color flow Doppler (CFD) provides critical hemodynamic cues in fetal echocardiography, yet its color overlays are often spatially sparse and temporally intermittent, and vary substantially across scanners and acquisition settings. As a result, reconstruction-error-based anomaly detection can be poorly calibrated because the reconstruction-error may be dominated by acquisition shift rather than pathology. Our key innovation in this paper is to propose a bi-level optimization scheme that couples a video translator (inner model) with an error estimator (outer model). The inner and outer tasks are interdependent: the error estimator predicts clip-wise reconstruction-error estimation to re-weight the cross-modal translation model training, while the translation errors produced in the inner model are used to supervise the outer error estimator. At inference, we derive a region-discriminated anomaly score by standardizing observed translation errors with clip-conditioned predicted error statistics and aggregating in-/out-region deviations into a participant-level score. The framework was validated on a large-scale, multi-site clinical cohort of 440 participants, covering six CHD types. Results show that the proposed method outperforms standard reconstruction-based baselines, achieving AUROC_all=0.723 and Spec@Sens75%=0.497 with all improvements being statistically significant (p < 0.01). Code will be released at [anonymous website].
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3455_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/QianyeYang/XCalibAD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanQia_Anomaly_MICCAI2026,
author = { Yang, Qianye AND Yang, Yingyu AND Wang, Yipei AND Hu, Yipeng AND Peng, Can AND D’Alberti, Elena AND Salovic, Bojana AND Iriondo-Coysh, Amaya AND Tsai-Goodman, Beverly AND Carvalho, Julene AND da Silva Costa, Fabricio AND Savvidou, Makrina AND Patey, Olga AND Papageorghiou, Aris T. AND Noble, J. Alison},
title = { { Anomaly Detection in Fetal Echocardiography via Cross-Modal Translation and Region-Discriminated Error Calibration } },
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 an anomaly detection method for fetal cardiac ultrasound, specifically using color flow Doppler imaging. They split this task into two: i) predicting flow from B-mode images, and ii) estimating the error/difficulty of the first task. An anomaly score is then computed as a reconstruction error, normalized by the estimated difficulty.
- 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.
- Strong improvement over baselines
- Important and challenging task with very limited existing solutions
- Well designed and described pipeline. Clever use of Doppler information in multiple ways. E. g. as a diagnostically relevant feature that can be estimated from B-mode images, promoting learning features related to cardiac function/flow, also using it to crop out an ROI, etc.
- 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.Is fetal cardiac Doppler imaging routine?: Although Doppler imaging provides more information, it can also be more difficult to acquire. From my limited knowledge, Doppler imaging is most accurate when the beam wavefront is aligned with the direction of blood flow. This adds further constraints to an already challenging imaging task (even acquiring the correct cross-section can be difficult), especially if you need to train a workforce large enough to screen an entire population. If there are any limitations in this regard, could you add some discussion on them? Or provide support to argue that fetal cardiac Doppler is routine.
2.Evaluation baselines/ablation: Most of the benchmark methods are some form of ablation rather than an actual competing state-of-the-art method. Could consider comparing with (foundation) embedding-based or self-supervised methods. Also, regarding the ablation, (Recon_B->D vs. Proposed-Direct), it would be interesting to know how much of this improvement comes from the training (masked losses and sample re-weighting) vs. the anomaly scoring (normalizing scores by the estimated difficulty of the prediction). This is by far the largest jump in AUPRC, so it could be useful to dissect it further.
3.Reconstruction error and its estimation: The annotations in Fig 3 highlight prediction errors that are directly related to flow dysfunction. However, those examples also exhibit gross errors outside of the annotated region (e. g. in case 2, frame 3, the red flow in the upper ventricle is largely absent in the prediction). This indicates that, even in this cropped ROI, US reconstruction is challenging and can produce false positives (as supported by the poor performance of Recon_B->D). It would be helpful to include a random healthy case and error estimates for all examples, as well as an example where error estimates are high. It could also be interesting to see the accuracy of the error estimator (which I suppose is ideally inaccurate on unseen classes?).
4.Non-diagnostic clips: In plain B-mode imaging, there are often many non-diagnostic clips included in US videos, which make reconstruction more difficult (by increasing the variance of normal). Is this an issue here as well, or does the Doppler aspect restrict clips to only diagnostically relevant planes? Are there erroneously acquired Doppler clips that are rare (harder to reconstruct), but not pathological?
- 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?
The authors present a clever design and demonstrate its merits on a very challenging task. Unfortunately, benchmark evaluation is a bit weak, consisting mostly of an ablation study.
- 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
Review #2
- Please describe the contribution of the paper
The method performs anomaly detection in fetal ultrasound using B-mode ultrasound (showing just anatomy) to color flow Doppler (blood flow) translation. This in itself is the main contribution, as no other work has sought to leverage both the B-mode AND the CFG modalities. It does this by exploiting the fact that the translation process is difficult, often resulting in error due to the different presentation of the CFD modality across different acquisition settings. With that in mind, they train 2 models; an inner model to translate the B-mode Doppler to CFG Doppler, and an outer model to predict the expected error of the inner model. Both models are entirely trained on ultrasound clips from healthy patients. At test time anomalies are identified by applying both models, measuring the translation error and comparing it to the predicted translation error, with a higher than expected error indicating a higher likelihood of anomaly. The outer error-prediction model is required (as oppose to using the translation error directly) as certain acquisition settings are inherently noisier or more ambiguous, so normalising using the predicted error accounts for this.
- 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.Innovation: The primary strength of this work is the innovation to adapt the modality translation task for anomaly detection. These two data modalities have not been employed for OOD before, let alone in this specific manner.
- S2.Appropriate setting for anomaly detection: There is a high degree of class imbalance in identifying cogenital heart disease in fetal ultrasound, both when grouping all anomalous classes together (8-9 / 1000 live birth) and even more so when you consider that there are many types of CHD. This makes supervised methods unsuitable, but unsupervised anomaly detection fits perfectly.
- S3.Results: The method achieves strong results, which are demonstrated on a test dataset covering 6 different CHD types. Through resampling the test set they show these results to be statistically significantly better than any simpler translation-based approach.
- S4.Data: as well as the various CHD types, the data covers multiple scanner types and multiple clinical sites, making the reported results more robust. The train / test split is correctly done at a patient level rather than clip level.
- S5.CHD type ablation: The ablation across different CHD types provides helpful insight, relating the different performance levels to the influence of the CHD on the hemodynamics. This helps highlight both the strength and potential blind spots of the method.
- S6.Clarity: the paper is well written and clearly laid out
- 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.Paired-image requirement: the main weakness is the requirement for paired B-mode and CFD mode ultrasound at inference time, which limits the variety of scanners/settings it may be used.
- W2.Lack of localisation: as the predicted vs expected error is only measured on aggregate across the whole clip, the localisation ability is limited to the size of the sliding window used (24 frames in this case). This limits the interpretability of the model. Of course the predicted and actual CFD scan can be manually compared, but this doesn’t clearly indicated the expected vs unexpected translation errors. This is done in the paper, but only qualitatively on 2 case studies.
- W3.Lack of non-translation-based baselines: The compared baselines are all modality-translation-based, it would be good to also compare against standard anomaly detection methods trained on the B-mode clips only.
- W4.Data availability: although the authors state the code will be provided upon acceptance, there is no mention of the data (train or test) being made available.
- W5.Data views: The conclusion mentions “Future work will expand view coverage”, but you don’t mention which views or acquisition planes are included in the current dataset. Could this be added?
- W6.No speed benchmark: There is no mention of inference time or similar, so cannot tell whether this could be used for real-time feedback (or provide a result within a ultrasound consultation), could this be added?
Some more minor weaknesses / suggestions:
- You don’t indicate how many repeats are used to calculate the mean/std of the metrics reported
- It would be good to have a ‘random’ classifier baseline in table 1, as the test set imbalance actually favours the abnormal class, which is unusual amongst anomaly detection methods where the test set is generally in-balanced in favour of healthy images. This leads to a baseline AUPRC above 0.5, which is unusual and readers may not expect without it being made clear to them.
- In section 2.3 I found the definition of
xright before reminder ofDa bit confusing, as it somewhat impliesxis drawn from the fullDduring training rather than eitherD_innerorD_outer. You do later specify the use ofD_outer, but I think it would be helpful to rephrase the part wherexis defined to ensure no confusion.
- 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
- When evaluating OOD on datasets with multiple clinical sites or scanner, it can be particularly useful to hold one site (/ scanner) out from training so you can measure how much distribution shift affects the anomaly detection performance. This is an important aspect of out-of-distribution detection, as we are primarly concerned with pathological anomalies but unknown scanners or differences in patient distribution can lead to false positives. This is obviously outside the scope of changes for this work, but would be good to apply for future studies (maybe use the same dataset, but hold out one of the four sites fully for eval).
- One possible idea for inference when only B-mode images are available could be to have 2 models trained on the translation task with distinct train sets (or some other way to make their learnt features different), then at inference you could measure the agreement vs disagreement of the two models, with the hypothesis that models behave more consistently for in-distribution data but differ on out-of-distribution samples. Addressing the inference data requirement using some method (possbily like this), and improving the localisation ability would be very exciting future work and I encourage you to keep pursuing it.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is exactly the sort of paper we want to be presented at MICCAI. It presents an innovative method, tackling a clinically relevant problem, which is evaluated in a clinically plausible setting (multiple scanner types, clinical sites, anomaly types) and it performs well. The only thing holding my rating back from being a Strong Accept is the lack of localisation ability/evaluation, if this were present the paper would be truly outstanding. As it stands, there is no doubt in my mind that the paper should be accepted.
- 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 #3
- Please describe the contribution of the paper
This paper proposes an anomaly detection framework for fetal echocardiography using paired B-mode ultrasound and color flow Doppler (CFD) videos. The core idea is to learn a B-mode-to-CFD translation model on healthy data and to detect anomalies by comparing the observed reconstruction error against clip-conditioned expected error statistics predicted by a second model, yielding a calibrated, region-discriminated anomaly score. The proposed method outperforms the included reconstruction-based 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 paper addresses a clinically important problem, namely CHD screening in fetal echocardiography, where anomaly detection is well motivated by the rarity of many CHD phenotypes and the difficulty of exhaustive supervised labeling
- The idea of calibrating the reconstruction error with clip-conditioned predicted error statistics is a meaningful contribution
- The distinction between in-mask and out-of-mask errors is sensible and gives the anomaly score a more interpretable structure than using a single global residual
- The bi-level training setup is technically nontrivial and goes beyond a standard reconstruction-based anomaly detector by coupling translation and error estimation
- The dataset is relatively large and multi-site, and the inclusion of several CHD phenotypes improves the practical relevance of the study
- The qualitative case studies are helpful and support the claim that the method can produce clinically interpretable failure patterns in abnormal cases
- 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 manuscript does not directly validate the central assumption that CFD is sufficiently predictable from B-mode in healthy cases. While anomaly detection performance is reported, standalone reconstruction quality or translation fidelity metrics are not provided
- The comparison set is limited. The paper mainly compares against reconstruction-based baselines and one direct-prediction ablation, but does not include a broader set of stronger or more recent anomaly/OOD detection alternatives
- Hyperparameters are said to be selected empirically in preliminary experiments, which weakens reproducibility and makes it harder to assess whether the configuration was fairly chosen
- Although the method uses a CFD-derived binary mask and includes a Dice-based loss term, the paper does not report any segmentation-related performance or even numerical values of the mask-related losses, making it difficult to assess whether the learned support region is reliable
- The proposed anomaly score fundamentally depends on comparing observed reconstruction error to predicted clip-conditioned error statistics, yet the paper does not report numerical summaries of these underlying quantities or directly analyze how they relate to each other
- The qualitative figures show reconstructed Doppler patterns, but they do not visualize the predicted uncertainty/error statistics together with the observed reconstruction error, even though this comparison is the basis of the anomaly detection formulation
- The paper does not provide a plot or calibration analysis showing how predicted uncertainty/error statistics correlate with observed reconstruction error across normal and CHD samples. Such a visualization would make it much easier for readers to judge whether the outer model is actually capturing clip difficulty as intended
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically important and difficult problem, namely CHD anomaly detection in fetal echocardiography, where supervised learning is limited by the rarity of many CHD subtypes. Methodologically, it proposes a reasonably novel combination of cross-modal B-mode to CFD translation, clip-conditioned error calibration, and bi-level optimization, rather than a standard raw reconstruction-error pipeline. It validates the method on a relatively large multi-site cohort and shows statistically significant improvements over the included baselines. However, the baseline set is limited, and the paper does not analyze its central calibration mechanism as thoroughly as it should. Still, the clinical relevance, technical idea, and consistent empirical gains make it a worthwhile 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
To all reviewers: We sincerely thank the reviewers for their constructive comments and early acceptance. We will improve clarity and discuss limitations in the camera-ready version, while leaving additional ablations, baselines, and analyses for the journal extension, due to limited space.
Reviewer #1 Thank you for recognizing the importance of the task and the design of the pipeline. Q1.We agree that fetal cardiac Doppler is more acquisition-dependent than B-mode imaging, and that Doppler color patterns vary with imaging parameters and ultrasound machines. Therefore, our method is intended for examinations where paired B-mode/CFD clips are already acquired. We will clarify this limitation in the camera-ready version. Q2.We agree that the baseline comparisons are relatively narrow, as this study focuses on comparisons within reconstruction-based approaches. We will investigate fair comparisons with stronger baselines in our journal extension and provide more details on the performance gain under different settings. Q3.We agree that adding a normal case would improve illustration, and that showing the outer model’s performance would be helpful. These details will be added in the future extension. Q4.This is a great point. Video-based fetal cardiac ultrasound data often lacks frame-level annotations indicating whether each frame is diagnostic. In our study, each scan is a short clip of around 10 seconds focused on the heart, which largely reduces irrelevant anatomy/background. Although this cannot guarantee that every frame is diagnostic, the AI model may still identify abnormality/normality from frames considered non-diagnostic by humans. This is an interesting problem worth further investigation.
Reviewer #2 Thank you for the encouraging review and for highlighting the novelty of the idea. Q1.We agree that fetal cardiac Doppler is more acquisition-dependent than B-mode imaging. Our method is intended for examinations where paired B-mode/CFD clips are already acquired, and we will clarify this limitation in the camera-ready version. Q2.We strongly agree that localization would provide important interpretability for this method. We will add this analysis in the journal extension due to the space limit. Q3.We agree and will compare with non-reconstruction-based approaches in the extension work. Q4.Fetal cardiac US data is sensitive, and to our knowledge, no public dataset of this type is currently available, even for image-based data. We will investigate whether public adult data are available and use them in the extension work if possible. Q5.By view coverage, we refer to transverse/long-axis and sagittal/short-axis views. This study uses only transverse views and excludes sagittal views to keep the approach simple. We agree this wording may be misleading, as readers may interpret “view coverage” as standard cardiac views such as 4CH/LVOT/…. We also agree that standard-view analysis is important, but due to the lack of frame-level view annotations, we cannot provide the view composition in this revision. We hope to add these labels in future work, even on a small evaluation set. Q-Minor. Thank you for pointing these out. We will fix them in the camera-ready version.
Reviewer #3 Thank you very much for the useful feedback and suggestions. Q1&Q4.We agree and will curate a separate set to report video-translation performance, perform a formal evaluation, and include it in the extension work. Q2.We agree and will compare with non-reconstruction-based approaches in the extension work. Q3.Agree. Due to GPU resource limitations, it was not possible to test many hyperparameter settings. We will perform a more formal evaluation, such as grid search, in the extension work to improve reproducibility. Q5.Fully agree. We will add this analysis in the journal extension due to the space limit. Q6&Q7.These are excellent points. We appreciate the suggestions and will include these analyses in the extension work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
All reviewers recommend acceptance. They agree that the paper addresses an important and challenging clinical problem in fetal echocardiography and proposes a meaningful anomaly-detection framework based on B-mode to color-flow Doppler translation, error calibration, and bi-level training. Reviewers view the method as more substantial than a standard reconstruction-error pipeline, highlighting its strong empirical performance, clinically plausible motivation, and evaluation in a difficult setting. The remaining concerns mainly relate to analysis and presentation, including clearer discussion of Doppler acquisition, stronger positioning against non-translation-based baselines, better separation of the contributions of calibration and training design, validation of translation fidelity, analysis of predicted versus observed reconstruction error, improved visualization and localization, and clarification of paired-image/inference requirements. These issues should be addressed in revision, but they do not overturn the positive consensus. Given the unanimous acceptance recommendations, clinical relevance, methodological substance, and reported performance, early acceptance is recommended.
