List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yu Zhang, Jun Ma, Jing Li
Abstract
In recent decades, coronary artery disease (CAD) is the leading cause of death worldwide. Therefore, automatic diagnostic methods are strongly necessary with the progressively increasing number of CAD patients. However, it is difficult for physicians to recognize the lesion from Coronary CT Angiography (CCTA) scans as the coronary plaques have complicated appearance and patterns. Previous studies are mostly based on the single image patch around a lesion, which are often limited by the field of view of the local sample patch. To address this problem, in this paper we propose a novel vessel-wise object detection method. Different with previous approaches, we directly input the whole curved planar reformation (CPR) volume along the coronary artery centerline into our deep learning network, and then predict the plaque type and stenosis degree simultaneously. This enables the network to learn the dependencies between distant locations. In addition, two cascade modules are used to decompose the challenging problem into two simpler tasks and this also yields better interpretability. We evaluated our method on a dataset of 1031 CCTA images. The experimental results demonstrated the efficacy of our presented approach.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16437-8_20
SharedIt: https://rdcu.be/cVRs6
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The paper discuss a two stage cascade neural network to detect and classify abnormal regions in a coronary artery. The first neural network is used for detection and the second neural network is used for classification of the abnormal regions into plaque type and degree of stenosis. The authors discuss their results and compare them with other existing 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.
Main Strengths are:
- Well written and explained
- Good and convincing diagrams
- Comparison with existing methods.
- Experiments
- Large dataset
- 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.
The paper presents a good application of the Faster R-CNN method for abnormality detection but in doing so lacks to present any new innovation/method for coronary CTA. The paper also lacks details in centerline extraction and vessel segmentation methods The paper does not provide any details on how straight line MPR views were extracted from cardiac CTA. The authors should have omitted that the data was collected from 6 centers in China in order to not break anonymity.
- 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
I know that the centerline extraction and straight MPR generation tools sometimes comes with the softwares associated with the scanner. But they should be clearly specified in the paper. If the authors are using some other tools for centerline extraction and straight line MPR generation, they should specify that too. If they make the code available that would be very helpful for reproducibility too.
- 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/2022/en/REVIEWER-GUIDELINES.html
I think the paper is well written and explained. It was consistent and diagrams were appropriately explained. The authors have compared their work to previous published works too. However, it is lacking some details as mentioned earlier and I think including those details would make it an interesting paper. Some other work in the literature for example (Performance of a Deep Neural Network Algorithm Based on a Small Medical Image Dataset: Incremental Impact of 3D-to-2D Reformation Combined with Novel Data Augmentation, Photometric Conversion, or Transfer Learning, Automatic stenosis recognition from coronary angiography using convolutional neural networks) are missing. I would encourage the authors to include them as they see fit.
- 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?
I think the paper is well written, but it lacks any new technique or ideas. The results are comparable to the previous methods and I do not see any statistically significant difference in the accuracy values.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #2
- Please describe the contribution of the paper
This paper presents a deep learning based method for coronary artery lesion detection and analysis. It incorporates Faster R-CNN and a multi-task network for lesion detection, plaque classification and stenosis degree regression.
- 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 main strengths of this paper include introducing Faster R-CNN framework for lesion detection. Then a multi-task network is employed for plaque classification and stenosis. In the plaque classification branch, two FC layers are used to determine the presence of calcified and non-calcified plaque respectively. Experiments have been conducted using a large clinical dataset consisting of 1031 CCTA images.
- 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.
The first stage of the method seems to be a simple application of Faster RCNN to the lesion detection problem without modifications for medical images or lesion detection target. The second stage of the multitasking network is an intuitive solution to accomplish lesion classification and stenosis regression at the same time. Therefore the method may not be innovative enough. Please see detailed comments below.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 authors did not release all the details about the proposed two modules, such as network architecture and feature map channels. The dataset used in the experiments are collected clinically and will not be made public.
- 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/2022/en/REVIEWER-GUIDELINES.html
(1) For vessel lesion detection, using the curved planar reformation image stacked from centerline seems to be a common preprocessing method. However, I think it is necessary to explain why this transformation is used instead of the original 3D volume of patch. (2) The proposed detection module seems to be a simple application of Faster RCNN to the lesion detection problem without modifications for medical images or lesion detection target. (3) Ablation experiments are needed to demonstrate the superiority of the multi-task network in the second stage, i.e., is it more effective than performing plaque classification and stenosis degree regression separately? (4) Why is it necessary to refine the localization of the region proposals in the second stage? (5) What does the word “vessel-wise” in the title mean specifically? (6) More details are needed for the ground-truth of stenosis degree. What is the doctor’s annotation? What are the advantages of the proposed regression method between [0,1] compared with performing classification directly? (7) The term “polar coordinates” appear for the first time in session “Ablation Experiments”, but is treated as a major novelty, which needs more explanation in the “Method”.
- 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
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method may not be innovative enough. The first stage of the method seems to be a simple application of Faster RCNN to the lesion detection problem without modifications for medical images or lesion detection target. More details and explanation need to be added.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
(1) This paper proposed a novel vessel-wise detection architecture inspired by Faster R-CNN To the best of our knowledge, it is the first application of this type of method on coronary artery lesion detection. (2) This paper proposed a multi-head analysis module that can predict not only plaque types but also the exact stenosis degree, rather than just significant stenosis classification. (3) This paper achieved an outstanding performance on a dataset consisting of 1031 CCTA images with 7961 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.
The whole curved planar reformation (CPR) volume along the coronary artery centerline is used for the coronary plaques analysis. This is an application innovation
- 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.
This paper is generally biased towards innovation in application, and the method and ideas are borrowed from Faster R-CNN.
- Please rate the clarity and organization of this paper
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
The dataset is nice and hopefully made public.
- 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/2022/en/REVIEWER-GUIDELINES.html
The author mentioned arbitrary CPR length is acceptable. Did the experiments in this paper apply different length CPR inputs?
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper tackles detection of plaque-type and stenosis degree simultaneously by inputting the whole curved planar reformation (CPR) volume along the coronary artery centerline instead of slices. A commonly used Faster R-CNN and a novel two binary digits code loss are applied. The method is verified on a private CCTA dataset and the results demonstrated the efficacy of the approach.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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 introduces a multi-task deep learning network for lesion detection, plaque classification, and stenosis degree regression. using Coronary CT Angiography (CCTA) scans. The pipeline is trained and tested on a local cohort of 1031 CCTA images with 7961 vessels. This paper documented an outstanding performance against baseline and other literature methods The work is an interesting application with innovative idea. The main strength in the paper is in 1) using CPR volume along the coronary artery centerline for the coronary plaques analysis, and 2) the introduction of a multi-head module to predict plaque types and the exact stenosis degree. However, the paper needs improvement, and some issues are needed to be accounted for, which are listed below as well as by the reviewers’ comments. The method lacks technical details necessary to better review and understand the approach results. For example, centerline extraction (R1), modules architecture and feature map channels (R2), network parameters selection and settings. Also, the authors need to discuss the efficacy of simultaneous plaque classification and stenosis degree regression compared with being done separately. How the disagreement between radiologist for GT generation is resolved? The reported results lack statistical analysis to show the significance. It is interesting to detail how the authors adopted the methods in [12] and [13], were the methods optimized or defaults setting were used? The performance of such methods are very close (Table 2) even better for some metrics. The authors need to add in the caption of Fig. 3 (or on the figure itself) how good or bad the performance is along with stenosis degree. Preferably on good and on bad example of methods. A brief discussion is missing to show the signifance and the methods limitation.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
3
Author Feedback
We sincerely thank all the reviewers (including meta-reviewer) for investing their precious time into reviews and appreciate them expressing their interest in our work. Three reviewers all affirmed the contributions of our work, i.e., multi-task method on coronary lesion detection, large dataset. The most important is that we are the first to apply object detection method on coronary lesion detection. Due to the limitation of pages, reviewers might have some misunderstandings and we will explain them below. All the reviewers concerned about the innovation of the paper, so we reaffirm it here. 1.We think it is a pioneering work that bring object detection method from natural image processing to coronary lesion analysis. 2.Previous works all used point-wise methods to predict each point along the centerline (N to N). On the contrary, we take the global situation into consideration and input the whole CPR image into the network at once (1 to N). (vessel-wise, R2). It is enlightened by the practice of how physicians analyze lesions. 3.We simplify a difficult four type classification task to 2 binary classification tasks. When physicians want to fully observe a vessel they will look from different angles. Besides, the vessel region may only occupy a small region of the cross-section view along the centerline, so we transformed the CPR image from Cartesian to polar coordinates that can enrich the context from every angle of the vessel (as R2 stated, we really should have described this idea in session “Method”). The core idea of the paper was to develop a new perspective for coronary lesion analysis. So we focused on explaining how the whole idea worked and omitted some important details. We thank the reviewers for indicating that. About session “introduction”, we should have included more other works, like the work R1 stated. About the dataset, every case would be checked by 3 experienced physicians, any disagreement would be determined by votes. They identified plaque type by their trainable skills and calculated quantified stenosis rate by measurement on CCTA scan. In previous deep learning methods, they may just classify significant stenosis or not. However, some elaborate clinical CCTA report needs quantified stenosis rate. (R2) As for the method, we followed the diagnosis process (vessel segmentation, centerline extraction, CPR, lesion analysis) and focused on lesion analysis, so did not state the previous steps (R1). It is why CPR was used instead of the original 3D patch (R2). We should have explained Fig. 1 in more detail, i.e., the number before @ is feature map channels, the conv-layers’ (if not specified) kernel size is 3 with 1 stride and 1 padding (R2). Our dataset contained CPR image with length from 64 to 800 points and Fig. 1 shows an example of 480 points (R3). About the experiments, what we have to say is that we actually took a lot of time to design them. In [13], they have verified the efficacy of simultaneous plaque and stenosis analysis compared with being done separately, so we did not repeat the similar experiments (R2). During reproducing other methods, we kept the default setting unchanged. The stenosis result of our method is significant lower than most of other methods, but the plaque result is close to some of them (R1). This is correct. Our main contribution is to apply the established object detection method to the challenging task of coronary lesion analysis and demonstrate their value for combining 2 binary digits code method with polar transformation for analysis. Investigating other models is an interesting direction for future research, for that the proposed method provides an important basis. However, we should have added stenosis result and bad example in Fig. 3 and discussed the significance and limitation of the method for future work. In the end, we can deeply understand these constructive comments raised by reviewers and we will address these helpful suggestions in the final version.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The authors have cleared most of the comments related to contribution of the work, significance, and GT generation. Their response contains many details that are missing in the methods section and should be added to the paper. Also, they should include statistical analysis in camera ready without substantial changes to the original submission.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
2
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
Overall this paper presented a solid clinical system which might be useful and the quantitative experiments on a large cohort of patient data are also positive aspects. However, for MICCAI submission, even from the most positive Reviewer 1 who listed the following comments (which is largely true), this submission lacks of necessary novelty. The performance success of a) is largely depending on B) and c) which are not discussed in the paper. Using Faster R-CNN to detect Coronary Artery Lesion on straight line MPR views with a large number of training data should be relatively straightforward (thus there is no much of technical contributions).
“a) The paper presents a good application of the Faster R-CNN method for abnormality detection but in doing so lacks to present any new innovation/method for coronary CTA. b) The paper also lacks details in centerline extraction and vessel segmentation methods c) The paper does not provide any details on how straight line MPR views were extracted from cardiac CTA.”
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Reject
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
NR
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The paper presents an object detection framework approach for coronary plaque and stenosis characterization. The CTA volumes are analyzed using regional proposal network (similar to Faster RCNN) followed by a multi-head classification for plaque type, stenosis, etc. The performance is shown to be better than those using transformer architectures (MICCAI 2021 work) which is a bit surprising since transformers are later architectures. In any case, the paper is interesting and the reviewers seem satisfied with the rebuttal.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
7