Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Reconstructing 3D coronary arteries is important for coronary artery disease diagnosis, treatment planning and operation navigation. Traditional techniques often require many projections, while reconstruction from sparse-view X-ray projections is a potential way of reducing radiation dose. However, the extreme sparsity of coronary artery volume and ultra-limited number of projections pose significant challenges for efficient and accurate 3D reconstruction. We propose 3DGR-CAR, a 3D Gaussian Representation for Coronary Artery Reconstruction from ultra-sparse X-ray projections. We leverage 3D Gaussian representation to avoid the inefficiency caused by the extreme sparsity of coronary artery data, and propose a Gaussian center predictor to overcome the noisy gaussian initialization from ultra-sparse view projections. The proposed scheme enables fast and accurate 3D coronary arteries reconstruction with only 2 views. Experimental results on two datasets indicate that the proposed approach significantly outperforms other methods in terms of voxel accuracy and visual quality of coronary artery.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1015_paper.pdf

SharedIt Link: https://rdcu.be/dV5Bw

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_2

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1015_supp.pdf

Link to the Code Repository

https://github.com/windrise/3DGR-CAR

Link to the Dataset(s)

https://asoca.grand-challenge.org https://github.com/XiaoweiXu/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT

BibTex

@InProceedings{Fu_3DGRCAR_MICCAI2024,
        author = { Fu, Xueming and Li, Yingtai and Tang, Fenghe and Li, Jun and Zhao, Mingyue and Teng, Gao-Jun and Zhou, S. Kevin},
        title = { { 3DGR-CAR: Coronary artery reconstruction from ultra-sparse 2D X-ray views with a 3D Gaussians representation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {14 -- 24}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This work explores coronary artery reconstruction from ultra-sparse X-rays using a 3D Gaussian representation.

  • 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.

    Reconstruction of 3D targets from 2D projections is an interesting problem with clinical significance.

  • 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.

    Poor organization and weak experiments are the major weakness of this work.

  • 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 does not provide sufficient information for reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    My major comments are:

    1. Paper is tough to read and lacks proper organization. E.g., in the methods section, authors mention about the two stages of the proposed method. But the 3DGR reconstruction stage is never highlighted later.
    2. Can the author provide any insight about the fall in performance while directly implementing [7] for CA reconstruction?
    3. Some parts can be made clearer by using simpler/direct sentences such as U-Net is used for Gaussian center prediction.
    4. In the imageCAS dataset, out of 1000 CCTA images only 20 are used for testing. Is there any specific reason behind using such a small test set?
    5. Authors have only compared their proposed method with FBP and NeRP which significantly weakens the experiment section and reduces the impact of this work.
  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Poor organization and weak experiments are the major reasons behind my decision.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    Author’s have successfully addressed most of my concerns. I’m happy with most of their responses. Even though I think including more comparisons and performing better experiments can further strengthen this work, in its current form this work will still be useful and relevant for the community. Hence, I change my overall opinion to accept.



Review #2

  • Please describe the contribution of the paper

    In the paper “3DGR-CAR: Coronary artery reconstruction from ultra-sparse 2D X-ray views with a 3D Gaussians representation” the authors leverage a novel approach from the computer vision domain - 3D Gaussian splatting - and employ it in a 3D vessel reconstruction task from limited views. The original approach is extended by a prior Gaussian center estimation to enable a good performance for the task at hand.

  • 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.
    • As the essence of MICCAI is to present novel approaches, using Gaussian splatting as a technique for 3D reconstruction is an excellent point for discussions.
    • Also, the extensions proposed for this application in the Gaussian center estimation are technically sound and novel.
    • Experiments were conducted on large public data sets easing reproducibility.
  • 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.
    • For me the clinical motivation for the task at hand is very weak. The authors state that “Accurate 3D reconstruction of coronary arteries greatly assists physicians in the diagnosis and treatment planning of CAD”, but this task is usually performed pre-operatively with a CT and not in the cath lab, where the issue of reconstructing from a limited number of views can actually be formulated. However, in clinical practice it is unlikely that a physician performs an acquisition from a single angle then rotates the C-arm and finally performs another acquisition to then evaluate a 3D representation of the coronaries.
    • Also, it is unclear which angles where between the individual views, which would have been a very important information to judge the performance of the method.
    • Furthermore, the comparison with related work is in my opinion weak. I would have expected some traditional method extending on Unberath et al. [1] as a baseline.
    • The reproducibility of the paper is weak as a lot of details with respect to the network architecture are missing. [1] Unberath, Mathias, et al. “Improving segmentation quality in rotational angiography using epipolar consistency.” Proc MICCAI CVII-STENT, Athens (2016): 1-8.
  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The data used is public, but the concrete implementation is unclear.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    • Consider extending the clinical motivation for this paper to be more convincing
    • Please add details on the angles between the views
    • It’d be appreciated if the concrete implementation details of the employed network could be specified or even the code is shared to further enable reproducibility
  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    A technically interesting paper with poor clinical motivation. For me the advantages outweigh the shortcomings but additional technical details need to be added.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    The authors promised to release their code including models, which is great. However, they stick to the narrative that scans performed with a modality used in the intraoperative suite can be used in a screening scenario, which does not make any sense. Therefore, I am between an accept and a reject for this paper, slightly leaning towards the accept.



Review #3

  • Please describe the contribution of the paper

    In the paper authors propose a novel model 3DGR-CAR for coronary artery reconstruction from 2D X-ray views using 3D Gaussian representation. The method is built out of two modules: gaussian center predictor (GCP) and gaussian optimization. The GCP module is a U-Net network that regresses initial 3D gaussian centers from single 2D view. The second module uses already trained GCP to obtain initial locations for gaussian centers from multiple 2D views, and optimizes gaussian representation. Authors compare with previous approaches for this task on ImageCAS and ASOCA dataset, and report better results especially in ultra-sparse regime - up to four 2D images.

  • 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) Paper is well-written and easy to follow

    2) Using U-Net to perform gaussian center prediction is an original way of initialising gaussian optimisation. Authors compare to FBP initialisation as previously proposed by Li et al. [7], and show that they’re proposed method is performing better.

    3) Proposed method achieves significantly better reconstruction quality for ultra-sparse view regime. As reported in Fig. 3 and 4, the method achieves comparable results at 2-4 views to FBP and NeRP needing 16. That’s a very promising result, both in terms of reducing radiation dose and allowing for fast 3D reconstruction.

    4) Benchmarked on public datasets ImageCAS and ASOCA.

    5) Data cost of the method is pretty low, since X-rays views for GCP learning can be synthetised based on raw CTA volumes only - no hand-annotations are required.

  • 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) Novelty of the work is a bit limited. Using 3D gaussian splatting based on FBP has been proposed before [7]. Rendering the main contribution of the work to be GCP module which is a simple U-Net architecture regressing gaussian centers for initialisation.

    2) I miss 3DGR-FBP in Figures 3 and 4. I think this is the main work which the authors improve upon and thus it should be included in this benchmark.

    3) I lack comparison on real paired X-ray and CTA examples. Authors employ digitally reconstructed radiographs (DRR) to obtain 2D X-ray views from CTA volumes.

    4) I miss information about how much time does the method need to perform the 3D reconstruction. Also in comparison with other included methods. This is an important factor in utilising such a method in clinical setting.

  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Authors don’t provide code, however method seems straight-forward to implement based on the provided details.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    1) I would like authors to include 3DGR-FBP in Fig. 3 and 4, and elaborate on how it performs in comparison to 3DGR-GCP on limited views.

    2) How long does it take for 3DGR-GCP to perform reconstruction? How does it compare to other included methods?

    3) It could be interesting to quantify noise in center initialisation (e.g. how many centers are inside of 3D arteries) and compare it against the quality of obtained reconstruction.

    4) Can we expect the model trained on DDR obtained 2D X-ray views to achieve good results on real 2D X-ray? It would be relevant to know that the model can be trained on synthetic data only and perform well on real data.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I find this work to have a limited novelty since the only original module is the U-Net like GCP module for gaussian center initialisation. However, I lean towards accepting this work since I find it to be relevant step towards enhancing 3D reconstruction from 2D views exhibiting large improvement in ultra-sparse view regime. The method is also fairly-well benchmarked - though I lack inclusion of 3DGR-FBP in some experiments.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    I thank authors for responding to my and other reviewers comments. Authors will add 3DGR-FBP in Fig. 3 and 4 where it was previously missing - that will make the evaluation more comprehensive. Provided percentages of prior gaussian centers close to artery centerline seem to back-up the claim of 3DGR-GCP being more suitable for working in ultra-sparse regime. Overall, I keep my previous score of 4 (weak accept), I find this work relevant though having a limited novelty as I mentioned before.



Author Feedback

We appreciate the reviewers’ comments and insightful suggestions, especially some thought-provoking proposals (R3,R4). We thank for the acknowledgement of the interest(R1), novelty(R3,R4), organization(R3,R4), and impressive experimental results(R4) of our paper. Our responses are as follows:

3DGS-FBP CLARITY R1-Q2: we clarify that for 3DGR-FBP, the Gaussian Centers (GC) derived from FBP introduce increasing noisy initialization as the number of views decreases, and much of this noisy GC can’t be eliminated during the optimization process, leading to a decline in 3DGR-FBP performance. In contrast, for 3DGR-GCP, the GCP module’s prediction of a rough coronary sparse point cloud significantly reduces the introduction of non-coronary geometric GC, enhancing reconstruction quality even with fewer views. R4-Q1&Q3: We calculated the proportion of GC close to 3D artery centerlines in the ImageCAS test set. We found that the ratio of 3DGR-GCP based on a single view is 21.78%. For 3DGR-FBP with 2,4,8,16 views, the percentages are 17.73%, 21.52%, 28.18%, and 34.72%. This confirms the earlier explanation. As the number of views increases, the reconstruction quality of 3DGR-GCP ranges from significantly better than 3DGR-FBP to nearly identical. R4-Q1& R3-Q3: We will add 3DGR-FBP to Fig. 3 and 4 and release the code, data and checkpoints.

EXPERIMENT R1-Q5: Lacking of arteries data with precise 3D labels, we focus on methods that require no external annotated data for training. Traditional methods such as FBP cope poorly with sparse view problems. Because Implicit Neural Representations use a continuous 3D representation to model scenes and are trained through view-consistency, they achieve good 3D reconstruction even with limited views and serves as good baselines. Therefore, we conducted extensive experiments on the FBP and NeRP methods. Additionally, we tried NAF[1], which improved results over NeRP when views exceeded two, but still conform to our conclusions. For voxel DSC metrics, NAF: 0.29-0.66-0.79-0.89; Ours: 0.70-0.89-0.92-0.92. This reinforces our confidence that 3DGS delivers SOTA results in reconstruction from ultra-sparse views without requiring 3D labels. [1] NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction, MICCAI 2022. R3-Q4: We extended FBP with orthogonality constraints for views consistency, and the results still showed artifacts similar to FBP. As suggested by the idea in the work of Unberath et al., using epipolar consistency helps improve segmentation quality, which we believe opens the path for us to add regularizations between projections to the 3DGS and GCP module, enhancing reconstruction performance.

CLINIC R3-Q1: Our study is essentially an exploration of the limits of reconstructing 3D arteries by a minimal number of X-ray views. In future clinical practice, autonomous scanning might reduce physician intervention, and our method holds great promise.

DETAILS R1-Q1&Q3: We will revise some indirect expression such as Sec-2.1’s subheading to ‘3DGR Reconstruction’, Sec-2.2’s to ‘Gaussian Center Predictor Training’ and complex sentences. R1-Q4: Considering that only GCP requires training, it is reasonable to use most of the ImageCAS dataset for training GCP and to select 20 random samples for validating the comparative methods. Since none of the comparison methods require external training data, and the results on the 40-sample ASOCA dataset are also similar, this corroborates that we did not selectively choose 20 samples for testing. R3-Q2: For 2-views, the interval is π/2. 4-views–π/4, 8-views–π/8, and so on. R4-Q2: The time cost is labeled as ‘Train’ in Fig. 1. We will include “Time cost” details in caption. R4-Q4: Quite promising. Artifacts and noise present in real contrasted X-ray (from other tissue), might impact the reconstruction results. Notably, 3DGS focuses on the instance itself, avoiding the issue of distribution shifts between real and synthetic data.



Meta-Review

Meta-review #1

  • 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’

    Creative paper, an interesting addition to MICCAI. The use of Gaussian splitting for medical image computing is (still) new. It would be good to still address one remark of the reviewer in the rebuttal, namely that intraoperative imaging has no role in screening, with which I concur.

  • 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).

    Creative paper, an interesting addition to MICCAI. The use of Gaussian splitting for medical image computing is (still) new. It would be good to still address one remark of the reviewer in the rebuttal, namely that intraoperative imaging has no role in screening, with which I concur.



Meta-review #2

  • 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’

    All reviewers agree to accept this paper.

  • 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).

    All reviewers agree to accept this paper.



back to top