List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yuehui Qiu, Zihan Li, Yining Wang, Pei Dong, Dijia Wu, Xinnian Yang, Qingqi Hong, Dinggang Shen
Abstract
Accurate extraction of coronary arteries from coronary computed tomography angiography (CCTA) is a prerequisite for the computer-aided diagnosis of coronary artery disease (CAD). Deep learning-based methods can achieve automatic segmentation of vasculatures, but few of them focus on the connectivity and completeness of the coronary tree. In this paper, we propose CorSegRec, a topology-preserving scheme for extracting fully-connected coronary artery, which integrates image segmentation, centerline reconnection, and geometry reconstruction. First, we employ a new centerline enhanced loss in the segmentation process. Second, for the broken vessel segments, we propose a regularized walk algorithm, by integrating distance, probabilities predicted by centerline classifier, and cosine similarity to reconnect centerlines. Third, we apply level-set segmentation and implicit modeling techniques to reconstruct the geometric model of the missing vessels. Experiment results on two datasets demonstrate that the proposed method outperforms other methods with better volumetric scores and higher vascular connectivity. Code will be available at https://github.com/YH-Qiu/CorSegRec.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43898-1_64
SharedIt: https://rdcu.be/dnwBY
Link to the code repository
https://github.com/YH-Qiu/CorSegRec
Link to the dataset(s)
ASOCA: https://asoca.grand-challenge.org/
PDSCA: Not Available
Reviews
Review #3
- Please describe the contribution of the paper
A topology-preserving scheme is proposed to accurately segment fully-connected coronary arteries from coronary computed tomography angiography (CCTA) scans. A new centerline enhanced loss comprising Normalized Skeleton Distance Transform (NSDT) Soft-ClDice and Soft-Dice, is proposed and applied in the segmentation method based on nnU-net. To address the issue of broken vessel segments, a regularized walk algorithm is proposed, which considers distance, centerline classifier probabilities, and cosine similarity to reconnect centerlines. In the coronary reconstruction stage, cross-sectional profiles are constructed perpendicular to the stitched centerlines, and vessel contours are extracted using level-set segmentation. The reconstructed vessels are merged with the original vessel branches to ultimately obtain a fully-connected coronary artery tree. The proposed method is validated on two datasets, comprising 60 CCTA images from the MICCAI 2020 Automated Segmentation of Coronary Arteries
- Please list the main strengths of the paper; you should write about 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 proposed segmentation method effectively focuses on the connectivity and completeness of the coronary tree.
- The NSDT Soft-ClDice loss is novel and tailored to this segmentation task. It increases the weights of vessel skeletons and reduces the effect of various vessel diameters, resulting in improved accuracy.
- The DPC walk algorithm, a locally optimal decision based on distance (D), centerline probability (P), and cosine similarity (C), is an innovative approach to reconnecting broken vascular centerlines and improves the connectivity of the segmented coronary arteries.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
In the abstract, the metrics of volumetric scores and vascular connectivity are mentioned but need to be clearly defined in the manuscript. Further clarification is required to determine if they relate to DICE and HD scores.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Code and models are not available now.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- The manuscript should clarify the metrics of volumetric scores and vascular connectivity and their relationship to established evaluation metrics, such as DICE and HD.
- The description of Equation 13 should be expanded upon to provide a clearer understanding of why RecAcc can indicate the success rate of DPC Walk reconnected vessels.
- An ablation study for the NSDT Soft-ClDice Loss should be conducted to further investigate its effectiveness compared to other methods.
- In Fig 5, the vascular reconstruction (blue) overlaps with some segmentation results (red), and its radius is inconsistent with the red part. Is there any simple method to optimize it?
- Comparing the inference time of the proposed method with other state-of-the-art methods would provide further insights into its performance.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed segmentation method for coronary arteries effectively preserves the topology of coronary arteries. It enhances the segmentation loss of nnU-net, reconnects the broken vessels, and reconstructs vessels based on stitched centerlines, ultimately improving the overall accuracy. While some details remain vague, the proposed method is still effective and innovative.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
The authors propose a coronary artery segmentation methods, focusing on improved postprocessing and maintaining physiologically correct topology of coronary vessels.
- Please list the main strengths of the paper; you should write about 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.
This work is very interesting, and addresses a major issue with voxel-level segmentation methods (e.g. CNNs). Since coronary arteries are thin and distributed over a large region, traditional CNNs can struggle to maintain vessel connectivity, particularly in areas with disease or artefacts. This work addresses the issues of disconnected segments by attempting to either remove them if they are a false positive or reconnect it to the main coronary tree. The results are well validated and seem to indicate the proposed method leads to better segmentations.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
In Table 2 the proposed method seems to have much higher standard deviation compared to other methods. It might be interesting to explore and discuss why these predictions are more inconsistent and if this could cause problems.
- Please rate the clarity and organization of this paper
Excellent
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The paper includes sufficient details on the methods used, parameters, and underlying datasets used for training and evaluation.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
I am somewhat curious what leads to this method having 20x higher standard deviation in table 2 compared to the baselines. It might be interesting to explore cases where the model fails and whether this would be relevant to those who want to use the model.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work is a good step towards making coronary artery segmentation models better able to handle disease, artefacts and other issues causing disconnected segmentations. The results are convincing and well validated.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #1
- Please describe the contribution of the paper
The authors propose a novel architecture, CorSegRec, a topology preserving scheme for extracting fully-connected coronary artery, which integrates image segmentation, centerline reconnection, and geometry reconstruction.
The approach is evaluated on two datasets: ASOCA and PDSCA, against other methods. The authors report superior results to the other methods.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) The Centerline stitching based on the DPC walk is an interesting approach to reconstructing the vessel topology. Authors report a much lower HD, than in reference works, which proves that false positive vessels are robustly filtered out.
2) the proposed approach is clearly described
3) the method is evaluated on two publicly available datasets
4) the method achieves strong results
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1) The disconnections in the CA segmentations are most often caused by stenosis or plaques. The qualitative comparison, however, does not showcase these cases - the reconstructed lumen is non-stenotic. Evaluation of pathological cases is important for proving the clinical feasibility of the approach.
2) The work lacks an ablation study of the proposed NSDT Soft-ClDice loss function. It is only evaluated together with the whole framework.
3) The reconstruction stage lacks evaluation on stenotic segments with plaques where the disconnections often happens.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Given the model was trained on publicly available data, it should be feasible to reproduce given results. However, since the method utilizes uncommon techniques, one would probably require an considerable amount of time to do it. It would be much easier, if the authors share the code.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
1) I’d like to see an ablation study of the proposed NSDT Soft-ClDice loss function. Its contribution to the whole framework’s performance is not clear to me.
2) I’d like to see evaluation, and discussion of the method performance in the presence of pathological parts of CA.
3) Table 2: row for clDice has wrong HD (mm) - based on PLF[20] it should be 6.29 pm 0.18
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is interesting and seems to be novel. However, I miss a discussion on the pathological parts of CA which are the most common causes of segmentation discontinuities.
The authors claim strong results on two publicly available datasets, which makes the work relevant in my eyes.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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 a combination of deep learning and conventional image segmentation approaches to extract fully connected coronary artery trees from CT images. The work goes beyond many recent works that merely perform voxel classification, and also incorporates the topology of the coronary arteries. Reviewers are unanimous in their positive evaluation of this work.
Strengths
- The work not only considers voxel segmentation, but actually considers extraction of the coronary tree as a contiguous structure.
- The paper proposes a new loss for small artery segmentation.
- The method is clearly described.
- Performance is excellent on two public datasets.
Weaknesses
- The evaluation is limited in the sense that it only looks at typical distance and overlap metrics, but does not consider stenosis or diseased areas.
Author Feedback
We’d like to thank all the reviewers for their careful reading and constructive comments. Our responses are as follows:
Response about evaluation of method performance in pathological parts of coronary arteries (to Meta-Reviewer and Reviewer #1) : In two datasets we used, we were provided with their voxel-wise labels but no information about the location of stenosis or plaques in each case. We agree that the evaluation of disconnected lesion segments is important for proving CorSegRec’s clinical feasibility. And we will try to validate this using other CCTA datasets which contain specific lesion information.
Response about ablation study on NSDT Soft-ClDice loss (to Reviewer #1 and Reviewer #3): By assigning the weights of masks and skeletons with distance transform values, the network will be more sensitive to the accuracy of centerlines, especially true positive centerline points. By balancing the weights of thick and thin vessels, the network will pay the same attention to vessels with different radii. The values of Dice metric for Soft-Dice+NSDT Soft-ClDice loss made 0.49% and 0.1% improvements compared with Soft-Dice loss on ASOCA and PDSCA. We will make more evaluation and attempts to investigate and improve its effectiveness.
Next is our response to the reviewer’s different comments.
Responses to Reviewer #1: Thank you very much for your valuable and constructive comments. Some responses are put above. We will correct the HD error for clDice on PDSCA.
Responses to Reviewer #2: Thank you very much for your suggestions and approval of our method. Q1: In Table 2 the proposed method seems to have much higher standard deviation compared to other methods. A1: We calculated HD metric using the evaluation framework of ASOCA challenge. And the HDs for other methods using PDSCA were presented by Zhang, X. et al. (“Progressive Deep Segmentation of Coronary Artery via Hierarchical Topology Learning, MICCAI 2022” cite [20] in original paper). We are not sure if it is caused by the calculation details of HD metric or other reasons.
Q2: It might be interesting to explore and discuss why these predictions are more inconsistent and if this could cause problems. A2: We’ll make more exploration about the disconnected and failure occasions of CorSegRec and analysis possible reasons for better utilization.
Responses to Reviewer #3: Thank you very much for your careful reading and detailed comments. Q1: The manuscript should clarify the metrics of volumetric scores and vascular connectivity and their relationship to established evaluation metrics, such as DICE and HD. A1: We’ll clarify the relationship between evaluation metrics and “volumetric scores and vascular connectivity” we mentioned in the abstract. Q2: The description of Equation 13 should be expanded upon to provide a clearer understanding of why RecAcc can indicate the success rate of DPC Walk reconnected vessels. A2: We’ll try to expand our description of RecAcc to make its indication more understandable. Q3: In Fig 5, the vascular reconstruction (blue) overlaps with some segmentation results (red), and its radius is inconsistent with the red part. Is there any simple method to optimize it? A3: In Vascular Reconstruction Stage, we reconstruct the vessel model of the stitched centerlines. And the predictions of network near the broken area are always incomplete, which causes the radius difference. If we do reconstruction along the merged branch (the candidate centerline, the stitched centerline and the broken centerline), we can get a smooth and consistent segmentation. Q4: Comparing the inference time of the proposed method with other state-of-the-art methods would provide further insights into its performance. A4: We’ll consider this and try to make more comprehensive evaluation of our method’s performance.