List of Papers Browse by Subject Areas Author List
Abstract
The cardiothoracic diameter ratio (CTR) biometric in four-chamber ultrasound plane is often measured for diagnosing congenital heart disease. However, due to the commonly existing artifacts like acoustic shadowing, manual measurement can be time-consuming and labor-intensive task, and may results in high measurements variability. Presently, one of the most popular approaches is segmentation-based methods, which utilize deep learning networks to segment the cardiac and thoracic regions. Then, the metric is calculated through an ellipse fitting scheme. This is inefficient, and requires additional post-processing. To tackle the above problems, in this paper, we therefore present an one-stage ellipse detection network, namely EllipseDet, which detects the cardiac and thoracic regions in ellipse, and then automatically calculates the CTR biometric in four-chamber view. In particular, we formulate the network that detects the center of each object as points and regresses the ellipses’ parameters simultaneously. Besides, we propose a novel ellipse feature alignment module and Ellipse-IoU loss to further regulate the regression procedure. We have evaluated EllipseDet on a clinical echocardiogram dataset and the experimental results show that our proposed framework outperforms several state-of-the-art methods. As an open science, source code, images dataset and pre-trained weights are available at https://github.com/szuboy/FOCUS-dataset.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5236_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/szuboy/FOCUS-dataset
Link to the Dataset(s)
https://zenodo.org/records/14597550
BibTex
@InProceedings{ZhaHon_You_MICCAI2025,
author = { Zhang, Hongyuan and Xie, Haoyu and Ye, Tingting and Wu, Songxiong},
title = { { You Can Detect It: Fetal Biometric Estimation Using Ellipse Detection } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a one-stage deep learning framework for cardiothoracic ratio (CTR) estimation, termed EllipseDet. The method simultaneously detects object centers as key points and regresses ellipse parameters to delineate anatomical structures. To enhance regression accuracy, the authors introduce an Ellipse Feature Alignment module, which leverages self-attention via key-query-value (K-Q-V) projections to capture spatial correlations. Additionally, they propose an Ellipse-IoU loss to better align predicted ellipses with ground truth. The approach is evaluated on an in-house clinical echocardiogram dataset, and the authors express intent to release the dataset publicly.
- 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.
- Use of self-attention through key-query-value (K-Q-V) projections to capture spatial correlations and incorporate this into the deep learning model.
- Implement an Ellipse-IoU loss to ensure that the ellipses are more accurately predicted in relation to the ground truth.
- The authors have an in-house clinical echocardiogram dataset and plan to make it publicly available. This would be a valuable contribution to the community.
- 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.
- Limited validation on only 300 images may not fully demonstrate generalizability. Details on train/val/test split criteria are lacking.
- The claim of being the first one-stage CTR detection method is questionable, as similar approaches have been explored in prior work (e.g., Chen et al.).
Chen, Jiancong, et al. “EllipseNet: anchor-free ellipse detection for automatic cardiac biometrics in fetal echocardiography.” Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VII 24. Springer International Publishing, 2021.
- 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.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Validation is limited to 300 images, raising concerns about generalizability. Details on the data split strategy are missing.
The performance comparison in Table 1 shows unusually large gains of EllipseDet over EllipseNet. It is unclear whether this is due to differences in dataset difficulty or model design. Evaluating EllipseDet on EllipseNet’s dataset would help clarify this.
The rationale behind EllipseDet’s performance advantage, compared to EllipseNet, remains unclear. An ablation study or deeper analysis of each component’s contribution is needed.
The method is not validated on public or large-scale datasets, limiting the evidence for its broader applicability.
Hyperparameter choices in Equation (8) are not explained. Additionally, the evaluation metric P_{CTR} should be properly cited if it is not introduced by the authors.
Writing issues: …..scheme. This is wasteful, inefficient, and…… wasteful word could be changed. “Backbone Network” The ResNet50—>the ResNet50?
Chen, Jiancong, et al. “EllipseNet: anchor-free ellipse detection for automatic cardiac biometrics in fetal echocardiography.” Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VII 24. Springer International Publishing, 2021.
- 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.
While the authors have addressed some previous concerns, I still have two major reservations that preclude acceptance at this stage:
Lack of Clear Technical Novelty: The proposed method is described as the first one-stage approach for the task; however, this claim does not convincingly differentiate it from prior work, particularly EllipseNet (Chen et al., MICCAI 2021), which also adopts a one-stage, anchor-free design for ellipse detection in fetal echocardiography. The methodological overlap raises concerns about novelty, and the authors have not provided sufficient evidence of substantive advancement beyond this prior work.
Limited Validation: The evaluation is based on only 300 2D samples, which is relatively small for drawing generalizable conclusions, especially in the context of fetal echocardiography, where inter-subject and inter-frame variability is high. A more comprehensive validation—ideally including multiple patients, views, and acquisition settings—is necessary to establish the robustness and clinical relevance of the proposed approach.
Review #2
- Please describe the contribution of the paper
The paper proposes an ellipse representation for fetal biometrics, and introduces an ellipse feature alignment module and intersection over union loss to supervise the training process. Additionally, the authors have gathered and organized a fetal biometrics dataset, which they have promised to release to the research community.
Existing methods to solve the same problem are segmentation-based, landmark-based, and box-based. Segmentation approaches are intermediary steps that accumulate errors during estimation, landmark approaches are affected by speckle noise and artifacts, and bounding boxes do not represent certain biomedical objects well. Therefore, EllipseDet is proposed to regress the ellipses’ parameters in a single-stage for fetal cardiac biometrics measurement in four chamber scans.
- 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 model uses both a backbone network, and a feature pyramid network which does indeed allow for generating multi-scale feature maps. The authors correctly employ heatmap-offset aggregation to resolve the discretization error from output stride after the feature alignment module. The feature alignment module EFAM reintegrates the decoupled location and regression features to provide shared offsets, which allows for dynamic weighting of global rotated ellipse feature learning. The authors have compared their model with SOTA approaches till 2024, which is reasonable and competitive are obtained. Qualitatively as well, the final results are more promising than the other approaches in recent literature. The ablations are well defined, and show the quality of each component.
- 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.
- It is only tested on one private (will be made public) dataset.
- The rationale for using an “ellipse” is not entirely clear. It definitely works in the context of this study, but it is not explained how or if this shape can apply to all fetal cardiac and thoracic shapes across different gestational weeks.
- Similarly, the EFAM module appears to contribute to the performance, but some more elucidation can be provided for how the aligning actually happens when location and regression features are concatenated. How does this relate to the final biometric parameters? Does the aligning use ground-truth annotations?
- How does the predicted offset, heatmap and size relate to the final estimated CTR, as in what aspect of the cardiothoracic object are these parameters capturing? Is the difference between predicted and actual CTR also calculated when scoring? If not, being a multi-task model, why not, how does it affect the performance?
- How are the confidence scores used across the FPN levels to assemble the predictions?
- Figure 3. could use a more intuitive explanation in the caption. What is a reader supposed to look out for, without reading the entire accompanying section? On that note, the architecture overview and Fig. 2 can be improved, to explain the steps (a, b) and how it relates to the entire training process more clearly. The pipeline is not well depicted, and would be difficult to understand without significant effort in reading. The text size in the figure, especially for b), can be increased.
- Minor suggestion, Table 1, please order methods by either year of publication, or attained scores in the most relevant metrics. Right now, the scores do not follow a trend, which makes it harder to observe the improvement proposed.
- 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.
- 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/2025/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 — could be 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 has potential, but issues in presentation obfuscate the contribution of the paper. The approach appears to be novel, and technical fidelity is noticed, with promising experimental results. But the results are just based on one single dataset.
- 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.
Address most of my comments
Review #3
- Please describe the contribution of the paper
The authors propose a one-stage ellipse fitting network to estimate the cardiothoracic diameter ratio in cardiac ultrasound images.
- 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.
To fit the ellipse, the authors use an Ellipse Feature Alignment Module to align positional and size information, and compute the loss based on the intersection-over-union of ellipses.
- 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.
Ellipse fitting has been explored in works like EllipseNet (MICCAI 2022); the authors should clarify the novelty and specific challenges addressed by their method. Additionally, the first two contributions appear repetitive—please explain their distinctions.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
1.The content of (b) and the regression branch in Fig. 2 seems repetitive. What is the intended emphasis? Also, should Floc (not Fcls) in the regression head? 2.The alignment module and loss are claimed as key innovations, but the use of DCN is unclear; the authors should further explain the challenges these components address. 3.For fair comparison, please specify the annotations used by each method (e.g., keypoints, masks, ellipses, rotated boxes). Also, the widely validated FPN ablation may be unnecessary. 4.N=100 performs better, but N=36 is used; why? Also, testing only two values may not be enough to find the optimal setting.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The experiments are complete and the paper is well structured.
- 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.
The rebuttal satisfactorily addresses previous concerns. The proposed one-stage ellipse fitting method and the introduction of a new public dataset represent a valuable contribution, and I support acceptance.
Author Feedback
We thank all reviewers for their valuable comments and positive feedback, such as the novel technical fidelity (R1), complete experiments (R1&R3), and promising results (R1).
Q: Rationale for Using the Ellipse (R1) A: The ellipse shape approximates the fetal thorax and heart contour in the four-chamber view during the mid-trimester, where these structures typically appear oval and regularly positioned. Its applicability to earlier or later gestational stages may require shape adaptation, which could be explored in future work.
Q: Relationship to CTR (R1) A: CTR is determined mainly by ellipse size. We do not compute the difference between the predicted and actual CTR in current work. It is interesting to incorporate CTR in our multi-task optimization loss to further enhance fetal biometric estimation.
Q: Comparison with EllipseNet (R2, R3) A: EllipseNet is still bounding box based, making the feature non-alignment and IoU computation expensive. Our EllipseDet is the first single-stage framework with ellipse representation, and two specifically designed new modules: Ellipse Feature Alignment Module (EFAM) and Ellipse-IoU, making the optimization and learning direct and efficient, as demonstrated by the ablation studies in Table 2.
Q: EFAM Module (R1, R3) A: Since the thorax and cardiac are always rotated, conventional networks suffer from rotation variations. Instead, EFAM which particularly handles rotations is an effective manner for fetal biometric estimation. We also introduce a deformable convolutional network layer to calibrate this rotation bias, and thus better predict the elliptical parameters without ground-truth annotations, as shown in Sec. 2.2 and visualized in Fig. 2(b).
Q: FPN Assembled (R1) A: In inference, we only assemble ellipse from predictions with the highest confidence scores across all FPN levels. We will add more details in the revision.
Q: Figure Enhancement (R1) A: Thanks for this good suggestion. We will add more intuitive explanations to enhance the readability of the figure.
Q: Not Repetitive in Fig. 2 (R3) A: The content of Fig. 2(b) is not a repetition of the regression branch. It is a visualization of the alignment of thorax and cardiac features with our EFAM.
Q: Ellipse-IoU Loss (R3) A: We use Ellipse-IoU loss to calculate the IoU between two ellipses and directly optimize the ellipse parameters. Ablation studies in Table 2 show that Ellipse-IoU brings 2%-3% improvement.
Q: Annotations of Each Method (R3) A: We have pointed the annotations of each method in Sec. 3.2. We will add the annotation format in in Fig. 4 and Table 2.
Q: Necessity for FPN ablation (R3) A: Compared to baseline EllipseNet, the FPN module is a new component of EllipseDet. Therefore, we conduct this ablation study to analyze the contribution of this component.
Q: N=100 vs. N=36 (R3) A: We set N=100 and 36 to test the sensitivity of our method, and the performance difference is less than 1%, although they may not be optimal.
Q: Dataset and Evaluation (R1, R2, R3) A: The dataset release at EllipseNet (2086 images) is larger than our open dataset (300 images), but it is not publicly accessible, therefore, we can’t evaluate EllipseDet on EllipseNet’s dataset. Instead, we contribute the first publicly available fetal cardiac biometrics dataset, providing a foundation for the medical community. As shown in Table 1, our method EllipseDet outperforms EllipseNet (77.87 vs. 93.18, 81.36 vs. 94.87, 45.15 vs. 94.41).
Q: Data Split Strategy (R2) A: As described in Sec. 3.1, we randomly split 300 images of 217 cases from 2 centers into training: validation and testing (200:50:50) at the patient level. The dataset will include guidelines, images, and annotations.
Q: Language Enhancement (R1, R2, R3). A: Thanks. We will carefully revise typos to enhance the paper, such as ordering the methods in Table 1, adding the references to the hyperparameter settings in Eq. 8 and P_{CTR} metrics, and correcting F_{cls}.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper receives divergent comments from three reviewers: R1 and R3 favor the paper and R2 disfavors the paper, all post-rebuttal.
After checking the paper, the reviewing comments and the rebuttal carefully, the AC tends to agree with R2 that the paper lacks the novelty and validation necessary for a MICCAI paper.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
While Reviewer #2 maintains a negative stance due to concerns about the paper’s novelty, I find the authors’ rebuttal to have comprehensively addressed these points. They clarified the methodological distinctions from prior work and demonstrated how their approach introduces a significant advancement. The difficulty in obtaining data makes it difficult for them to expand the scale of testing. Given the appropriate response to critiques—alongside strong alignment with the other reviewers’ post-rebuttal acceptance recommendations—I recommend acceptance.