List of Papers Browse by Subject Areas Author List
Abstract
Congenital heart defects (CHD) are a leading cause of fetal morbidity and mortality. Early prenatal detection using ultrasound can inform delivery management and postnatal intervention planning. Automated CHD screening from fetal ultrasound can support clinicians, particularly in settings with limited specialist resources. In this work, we propose an anomaly detection framework for CHD screening from four-chamber view (4CV) fetal echocardiography videos, one of the most widely acquired fetal cardiac views. Our method learns a motion representation from normal videos, yielding a one-dimensional latent motion trajectory over time. In normal hearts, this trajectory evolves smoothly and quasi-periodically across the cardiac cycle. In CHD, deviations from normal dynamics introduce increased high-frequency fluctuations. We capture this phenomenon using a cycle-adaptive spectral regularity score, defined as the ratio of high-frequency power to total power within sliding windows spanning a fixed number of cardiac cycles and robustly aggregated over videos of arbitrary length. Evaluated on one in-centre test set and two external-centre cohorts, the proposed framework outperforms reconstruction-based and latent-distance anomaly baselines, demonstrating improved effectiveness and generalizability across sites and across a range of critical CHD conditions. On the in-centre test set, our method achieves performance comparable to clinical expert readers, highlighting its potential to support fetal CHD screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0959_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{YanYin_Latent_MICCAI2026,
author = { Yang, Yingyu AND Yang, Qianye AND Peng, Can AND Salovic, Bojana AND D’Alberti, Elena AND Carvalho, Julene AND Patey, Olga AND Papageorghiou, Aris T. AND Noble, J. Alison},
title = { { Latent Motion Spectral Regularity for Congenital Heart Defect Screening in Fetal Echocardiography } },
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
The paper develops an unsupervised pipeline that aims to detect fetal CHD from ultrasound 4CH video sequences. It first does a representation of the structure and motion (1D) of each image quantifying the amount of fast deformations above 2 times the cardiac cycle frequency, which is estimated from the maximal spectral power.
- 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 unsupervised approach is interesting , specially in the setting of CHD where there is a lot of variability.
-
Outperforms other similar unsupervised methods.
-
- 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.
-
Prediction is lower than other methods in the literature (see https://www. nature. com/articles/s41591-021-01342-5). This should be commented.
-
Not fully convinced on the rationale that abnormal patients have more irregularities, since many CHD have a regular rythm. In fact, and due to fetal shunts, several CHD are fully asymptomatic after birth (for instance CoA). While some pathologies affect motion and make it irregularity, and I understand that it might have some unstability, I think that other pathologies are more structural. I feel that the approach should be better discussed
-
The method does not seem to generalize well to other datasets, where they have very low specitivity
-
There is little information on the performance of the reconstruction and motion estimation performance in the train, validation and test datasets, and whether this loss is similar in pathological or control cases.
-
- 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?
The data origin is not cited, probably due to the anonymization.
- 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
-
There is little evidence that the HR estimation works. It would give much more strength to the paper if the authors did a comparison with the manually computed HR, and the one derived from the authors.
-
As a suggestion of future work for the authors, if they think that they should quantify spectral irregularities, they can see if the LVIO Doppler has information, since it is already a time signal and relatively easy to obtain rather than a good quality 4CH.
-
GA of the populations is missing.
-
It is not clear how were the diagnosis done, whether it involves individuals who were diagnosed at birth, if they were diagnosed during the same scan that is given to the model; or later in the fetal stage.
-
Resampling to 30 Hz is a bit concerning, since the fetal heart has a very high rage (140bpm in 3rd trimester).
-
Rationale is not clear, since 4CH echocardiographic is a technically challenging acquisition for an obstetrician not specialised in fetal echocardiography.
-
- 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 has some methodological contributions. and the fetal application is very interesting and challenigng. However, I think that things should be better explained and clarified.
Even if it has shortcoming that the authors should clarify, I like the idea of using a knowledge driven “anomaly detection” in terms of spectral power, instead of using a classifier that might overfit.
- 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
- Anomaly detection framework for congenital heart disease (CHD) screening using four-chamber (4CH) fetal echocardiography videos.
- There are two stages to the architecture where stage 1 trains a latent motion trajectory (LMT) model on normal fetal cardiac videos to get a 1D scalar signal that tracks the cardiac phase-related signal. During inference, the video is scored using a spectral regularity measure computed from this 1D signal. The spectral regularity measures the proportion of high frequency power in the LMT assuming that normal cases produce smooth and quasi-periodic motion while CHD/anomalous cases have high frequency content.
- The LMT model is an autoencoder and the paper evaluates it with two decoder heads - one that uses reconstruction and another that uses registration.
- THe method is tested on one in-center and two external datasets and outperforms baselines and performance comparable to clinical experts.
- 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 significant problem of detecting a rare condition (congenital heart disease). The design premise of the method is that “structural defects and functional irregularities are often interleaved: altered anatomy can induce abnormal contraction patterns, suggesting that motion provides complementary cues for detecting abnormality.” It is intuitive and grounded in physiological mechanisms of the disease. The problem is best formulated as an anomaly detection one given the abundance of normal cases.
- The data used in training and evaluation come from multiple centers and captured on a diverse set of ultrasound machines. This is crucial to establishing the robustness of any clinical workflow and addressing issues of domain shift (as seen in the results of this method).
- The inference time evaluation of a spectral regularity measure provide an interpretable framework for clinicians to assess the model predictions.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- There is significant performance degradation on external test sites drop from 92.72 to 75.56 (Set 2) and 67.25 (Set 3) AUROC. It is clearly acknowledged. The paper would benefit from a more in depth analysis of this drop.
- The test-time protocol uses manual cropping of the heart ROIs while the training uses an automatic tool. This can be a confounding factor as well as is a major practical constraint. THe paper could report test performance with the use of the automatic cropper as well.
- The 80th percentile threshold on normal validation data is a reasonable heuristic, but its behavior across sites is not analyzed. Does the score distribution of normals shift across sites? A calibration analysis would be informative and clinically important.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The model is trained and evaluated on fetal echocardiography videos. While fetal ultrasound is a common procedure during pregnancy, fetal echocardiography still remains a specialized modality and the broader clinical problem would be better addressed by having a method that works on the more routine imaging.
- 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 significant problem of detecting a rare condition. The design premise of the method is sound and grounded in physiology. The method is evaluated on multi-center, diverse acquisition protocols. The contribution is meaningful but the experimental analysis needs strengthening to fully support the claims (especially the performance degradation on external sites).
- 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 #3
- Please describe the contribution of the paper
The authors present a framework for detecting abnormality in fetal US 4CH clips. The method is based on an auto-encoder model encoding a 1D latent motion trajectory and the definition of an abnormality score based on spectral analysis of that latent.
- 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 is well written, the scope and the problem neatly exposed. The method is clear and addresses a clear clinical need. The evaluation is fair and nicely explained.
The reviewer appreciates the indenpendance of the test set vs. the internal training set. The methodology to evaluate the results is also appreciated, with blinded clinical experts comparison.
- 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 code base is not provided
-
On the reconstruction head, it looks like the loss used combines an L2 between the output anchor and the input, and an L2 between the output at time t and the input. The authors could have explained more why the first term was needed, as it seems counter intuitive to impose that term. The registration loss makes more sense mathematically.
-
The poor specificity results (on test set 2 and 3) are somewhat surprising and concerning. The authors could have discussed that a bit further.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Introduction: some related methods (TVAE, DISCOVR) are referenced later in the manuscript instead of at first appearance
Methods:
Normalised cross correlation is used as acronym without being defined
The self-supervised AE model imposes the latent feature to be a linear expression of a learnt video-wise fixed latent point (anchor) and a learnt global vector in the latent dimension (trajectory). The coefficient (alpha_t) of this global vector becomes the 1D signal that is later analysed for its spectrum to detect abnormality.
It is quite neat to formulate this as a pure self-supervised auto-encoder reconstruction (or registration) problem. It forces the latent space to artificially focus (1) a static representation of the 4CH view, and (2) on cyclic motion to explain the variability across the US clip.
On the reconstruction head, it looks like the loss used combines an L2 between the output anchor and the input, and an L2 between the output at time t and the input. The authors could have explained more why the first term was needed, as it seems counter intuitive to impose that term. The registration loss makes more sense mathematically.
Results:
The reviewer appreciates the indenpendance of the test set vs. the internal training set. The methodology to evaluate the results is also appreciated, with blinded clinical experts comparison. The higher performance of the registration-based reconstruction loss makes sense: it is better posed mathematically.
The poor specificity results (on test set 2 and 3) are somewhat surprising and concerning. The authors could have discussed that a bit further.
- 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?
The key factor for this score are:
- neat and clear method, self-supervised, mathematically sound (with registration loss)
- fair evaluation with blinded experts, independent test sets
- clear delivery, well wriiiten
- 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
Author Feedback
We thank the reviewers and meta-reviewer for their constructive feedback. We provide the following clarifications regarding the main points raised. Method rationale (R1) Our spectral score is not intended to detect rhythm disorders. The hypothesis is instead that a model trained only on normal 4CH videos learns a normal latent motion representation, where frame-wise variation can be compactly described along a shared 1D motion direction. For abnormal hearts, the projection of each frame onto this normal motion axis may be less stable because of altered anatomy, abnormal contraction patterns, or out-of-distribution appearance. These effects can introduce local fluctuations in the extracted latent trajectory, even when the underlying rhythm is regular. The spectral regularity score is therefore used as a compact measure of deviation from learned normal latent dynamics, not as a direct measure of arrhythmia. Relation to supervised CHD screening methods (R1) We agree that supervised image-based methods have reported strong performance in the literature (the paper listed by R1). Our work addresses a different setting: anomaly detection from fetal 4CH videos, without using labelled CHD cases for training or threshold selection. This is motivated by the rarity and heterogeneity of CHD and by the abundance of normal fetal cardiac videos in screening practice. The aim is not to replace fully supervised multi-view diagnostic systems, but to explore whether motion-derived normality modelling can provide a complementary screening cue. Dataset clarification (R1) All data included in this study were acquired during the second trimester. We resampled videos to 30 Hz because some retrospective test videos were stored at 30 Hz. Using a common frame rate reduces confounding from different temporal sampling rates. At 30 Hz, fetal heart rates of approximately 120–180 bpm still yield around 10–15 frames per cycle, which is sufficient for the low-dimensional spectral summary used here. Multi-site evaluation and external-set specificity (R1–R3). The training and validation data were acquired at the same hospital, using GE Voluson machines, while the external sets include Fujifilm, Canon, and Samsung systems (no machine overlap). We agree that the drop in specificity on external sites is important. A likely explanation is that vendor/site-dependent appearance changes shift the anomaly-score distributions of both normal and CHD videos. Because the operating threshold was fixed using only the normal validation set, external normal cases may receive higher scores, reducing specificity. Future deployment would therefore benefit from site/vendor calibration or additional normal-only training data from the target domain. Importantly, under the same domain gap and thresholding protocol, LMT-Reg remained the strongest deep learning method on both external sets and maintained relatively high sensitivity, which is desirable for screening. Manual cropping at test time (R2). For the test sets, we used manual heart ROI cropping to reduce confounding from fetal/probe motion and standardise comparison across vendors. The automatic cropper tool was itself developed mainly on GE data, so applying it to external vendors could introduce an additional, uncontrolled source of domain shift. We agree that end-to-end evaluation using automatic cropping is important for practical deployment and will be included in future work. Reconstruction loss (R3). The first L2 term in the reconstruction loss follows the original LMT formulation and encourages the decoded anchor to represent the mean/static appearance of the video. Additional analyses and reporting. We agree that additional analyses would strengthen the work, including comparison of estimated heart-rate frequency with manual measurements, reconstruction/registration errors for normal versus CHD cases, and site-wise calibration. These will be addressed in follow-up 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.
This paper proposes an unsupervised anomaly detection framework for congenital heart defect screening in fetal four-chamber echocardiography. All reviewers find the paper well written and the self-supervised, physiologically grounded formulation original, interpretable, and clinically meaningful. They appreciate the multi-center evaluation with independent external test sets and blinded expert comparisons. The main concerns are the specificity drop on external sites and the need for deeper discussion of generalization, calibration, along with clarifications on the loss formulation and missing experimental details. Overall, the reviewers view these as points for refinement rather than fundamental flaws.