List of Papers Browse by Subject Areas Author List
Abstract
Accurate 3D coronary artery reconstruction from two non-simultaneous Digital Subtraction Angiography (DSA) projections is important for guidance in complex cardiac interventions, yet remains unreliable in practice. Cardiac pulsation and respiration introduce non-rigid motion between views, breaking rigid-geometry assumptions and causing cross-view misregistration, which often leads to vessel discontinuities, ghosting artifacts, and topology errors in reconstructed volumes. We propose NR-Align, a motion-aware non-rigid alignment module that reduces cross-view mismatch on dual-view backprojected volumes before downstream 3D reconstruction. To enable controlled training and analysis, we further build a physiology-driven simulation pipeline that synthesizes realistic cardiopulmonary deformations and provides paired 3D vascular ground truth from Coronary Computed Tomography Angiography (CCTA) volumes. Experiments on a synthesized dual-async benchmark constructed from the ImageCAS dataset and on unseen real clinical DSA cases show that NR-Align improves 3D Dice and reprojection agreement, and consistently reduces motion-induced artifacts, yielding more coherent coronary reconstructions. Code will be made publicly available.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3825_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wpbzs/NR-Align
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanPen_NRAlign_MICCAI2026,
author = { Wang, Pengbo AND Meng, Zhu AND Yu, Qiwan AND Su, Fei AND Di, Chunxia},
title = { { NR-Align: Non-rigid Alignment for Non-simultaneous Two-View 3D Coronary Reconstruction in Complex Cardiac Interventions } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the challenge of 3D coronary reconstruction from non-simultaneous DSA views by introducing NR-Align, a 3D dual-head network (Union & Overlap) that corrects non-rigid motion mismatch in backprojected volumes. A key contribution is a physiology-driven SVF simulation pipeline that models respiratory and cardiac motion to bridge the sim-to-real gap. Validated on simulation from ImageCAS and real clinical cases, the method significantly improves 3D Dice and topological consistency over uncorrected baselines.
- 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.
High Clinical Relevance: Addresses the critical but under explored issue of non-rigid misregistration in asynchronous single-plane DSA, moving beyond the unrealistic synchronous assumptions of prior works.
Physiology-Grounded Synthesis (questionable): The SVF-based simulation for respiratory and cardiac motion is technically interesting (to be verified). Ablation studies (Table 4) convincingly demonstrate that physiological constraints are essential for clinical generalization.
Novel Dual-Head Architecture: The NR-Align design elegantly separates union and overlap predictions. This structural constraint is well-justified and yields a significant performance gain (+0.10 Dice) over single-head baselines. Robust Sim-to-Real Validation: Evaluation on real clinical cases with manual 2D masks provides strong evidence of practical utility and successful domain transfer despite the lack of 3D ground truth.
- 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 cardiac phase and respiration phase are used to model the deformation. But what is the evidence / reference to support such a deformation? (Eq (1) and Eq (2)) What are the parameters in the 2 equations: v_resp, v_card, T_r, T_c? It is not clear how this is used in the synthesis, which is critical to understand if the synthesis obtained is reasonable.
Insufficient Baseline Comparison (most important): The module is only compared against uncorrected “Raw” inputs. The absence of comparisons with rigid/affine registration, established SoTA deformable models, or DeepCA’s [3] own motion handling makes it difficult to assess the relative contribution.
Limited Anatomical Scope: Evaluation is restricted to the RCA. Given the distinct motion patterns and complexity of the left coronary system (LAD/LCX), the lack of evidence for generalization narrows the claimed clinical impact.
Small Clinical Cohort: The test set is limited to 10 cases from a single site. Large confidence intervals in Table 2 raise concerns regarding statistical significance and robust sim-to-real transfer.
Significant Residual Error Gap: A large performance gap (~0.1 Dice) persists between NR-Align and the upper bound. The paper lacks a decomposition of whether this stems from the alignment module or the downstream reconstructor.
Lack of Computational Benchmarking: The paper does not provide the inference time for the NR-Align module. In complex cardiac interventions, near real-time feedback is essential, and the latency of 3D U-Net-based processing should be quantified.
- 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.
- 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
Minor comments: 1.The lambda_o and lambda_c weighting coefficients in the final loss (Eq. 6) are not specified in the paper. 2.Fig. 1 is informative but the color scheme and label font size are difficult to parse in print. A higher-contrast version may improve readability. 3.The clinical dataset (10 cases, single hospital) could be described with more detail, which is important context for assessing clinical generalizability. (protocol, views, how to establish the ground truth, used in training?)
- 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?
Refer to the strength and weakness part. The paper addresses a real, underserved clinical problem and presents two technically sound contributions: a physiology-grounded deformation synthesis pipeline and a dual-head non-rigid alignment network. The ablation results are convincing, and preliminary clinical results are encouraging. These factors merit a weak accept. However, the absence of comparison against any existing alignment or registration baseline and other minor weakness make it difficult to contextualize the method’s contribution. Physiology-Grounded Synthesis is also need to be verified. The paper would benefit from a stronger comparative evaluation and broader experimental scope.
- 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.
Most concerns are addressed. However the experiments are lacking of comparison with more other works such as [5] (it has code public available), and demonstrating the result on LCA.
Review #2
- Please describe the contribution of the paper
The paper at hand proposes a 3D coronary artery reconstruction from two DSA views using a novel “non-rigid mismatch compensation module” to address inconsistencies across the two views. It proposes a simulation pipeline with realistic cardiopulmonary deformations to train the approach and evaluate the task at hand.
- 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.
Interesting method to enable a course segmentation of the vasculature. A lot of effort went into making the simulated data as realistic as possible.
- 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.
Really small test data size. The approach depends on a prior segmentation algorithm and is highly dependent on the cross-view consistency of such a segmentation. The impact of the parameterization/design of the simulation pipeline itself was not really evaluated. The results on the clinical test data look quite poor. For me it is not obvious why one view has such a good alignment while the other one is completely off. Maybe it is due to different heart/pulmonary phases between the two views. This fact, however, questions the overall validity of the second view metrics on the clinical data.
- 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.
- 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
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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The approach looks interesting and there was substantial effort spend on creating a realisticly simulated dataset for training. However, the results on the clinical data look very far from being applicable in practice and the evaluation metrics there do not seem reasonable to me.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I stand with my reject as the authors basically confirmed that the second view metrics are not valid as there was no phase matching applied. Sure somehow the metric improved compared to raw, but given the small data set size this is not necessarily conclusive and the visual impression of the second view projection this is till very far from clinical application as stated before.
Review #3
- Please describe the contribution of the paper
The paper under review addresses a significant challenge in interventional cardiology, namely reconstructing a 3D model of coronary arteries from two 2D DSA projections acquired at different times. The authors propose a method for 3D reconstruction that includes a model for cardio-pulmonary movement. The method is then validated extensively on both simulated and clinical data.
- 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 authors developed a simulation pipeline that models heart (torsion) and lung (translation) motions. This allows them to create a large dataset of “mismatched” views with known 3D ground truths for training. NR-Align Module: A deep learning module that takes a mismatched back projected volume and “rectifies” it. The model, though trained on synthetic data, was successfully tested on real clinical DSA cases, showing it can generalize to actual hospital settings. Unlike many previous methods that assume rigid motion or simultaneous acquisition, this paper directly tackles the non-simultaneous nature of single-plane clinical systems. Quantitative results show a huge improvement over raw reconstruction, particularly as deformation increases
- 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.
How would the model perform on the LCA that has a more complex 3D shape? How realistic is the respiration model? As is, it combines a translation and a rotation component. What are the inference times for the proposed method?
- 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.
- 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
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.
(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 paper proposes a novel framework for 3D artery reconstruction, the authors include a simulation framework as well as a novel alignment module. The results show the benefits of the proposed method and a successful sim to real deployment is demonstrated.
- 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.
The authors responded to concerns related to the deformantion pattern choice. Still think it would be interesting to validate on other arteries as well. The paper is an interesting one and shows that sim to real translation results are promissing in a clinically relevant problem.
Author Feedback
We thank the AC and reviewers for the constructive feedback. We clarify the main concerns below.
[AC1, R2] Clinical performance and asymmetric view alignment. In real-world clinical practice, this remains an extremely challenging clinical task. In non-simultaneous single-plane DSA, two views are acquired at different time points while the coronary tree undergoes cardiac/respiratory motion, especially during C-arm switching. Thus, the projections may correspond to different physiological states, explaining the asymmetry noted by R2.A single reconstructed 3D vessel state cannot perfectly satisfy two projections acquired under different cardiopulmonary phases. Our goal is to reduce this insufficiently addressed non-rigid mismatch, rather than assume synchronized views; we therefore did not manually select phase-matched frames. Experimentally, NR-Align improves all clinical metrics over Raw: View-1 Dice 0.7638 to 0.8151, View-2 Ot(1) 0.2944 to 0.3526, and View-2 Ot(2) 0.3920 to 0.4322. [AC2, R1, R2, R3] Physiological validity of the SVF synthesis model. The deformation model is physiologically motivated, not unconstrained random deformation. Prior coronary imaging studies identify cardiac and respiratory motion as major motion factors, motivating our decomposition into cardiac and respiratory phases. In Eq. (1)-(2), T_r/T_c denote respiratory/cardiac periods; v_resp models respiration-dominant displacement, while v_card models localized cardiac-induced non-rigid torsion. The SVF exponential map follows established diffeomorphic modeling [16,19,20], enabling smooth, topology-preserving motion synthesis. Compared with unconstrained SVF, the physiology-aware model transfers better to real clinical data (View-1 Dice 0.7412 to 0.8151), showing that the gain comes from physiologically constrained cardiopulmonary modeling.
[AC3, R1] Baseline comparison. We understand R1’s concern as whether NR-Align improves over simpler motion assumptions such as rigid, affine, or generic deformable modeling. The experiments address this at two levels. First, Raw applies the existing DeepCA pipeline directly to the same non-simultaneous input. Since DeepCA’s motion modeling mainly relies on rigid approximations, Raw provides the relevant rigid-approximation comparison. With only NR-Align added, overall 3D Dice improves from 0.3872 to 0.5295 (+36.8%). Second, our synthesis ablation compares unconstrained SVF, i.e., random non-rigid deformation without cardiac/respiratory phase constraints, with physiology-aware SVF. The physiology-aware model improves clinical View-1 Dice from 0.7412 to 0.8151, showing that the gain comes from physiologically parameterized modeling.
[AC4, R1, R3] LCA generalization and runtime. RCA was selected because it is more affected by cardiac/respiratory motion and is therefore a strict target for non-rigid mismatch correction, not because NR-Align is RCA-specific. NR-Align uses generic backprojected union/overlap cues and does not depend on RCA-specific labels or fixed topology templates; thus, its mechanism can extend to LCA. The inference pipeline is feed-forward without case-specific optimization and the average inference time from loading the input to producing the 3D reconstruction is about 87.8 ms/case.
[R1, R2] Additional concerns. The 10 clinical cases are only an independent real-world evaluation set for sim-to-real transfer, not the training scale of the method; training and quantitative evaluation use the large-scale ImageCAS dataset (1000 cases). Clinical evaluation uses manual vessel masks, so the conclusions do not depend on a specific segmenter. The Upper Bound represents ideal synchronization and serves as a reference; the Raw-to-NR-Align improvement isolates the alignment gain, while the remaining gap reflects residual mismatch and downstream limits.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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 work tackles non-rigid motion mismatch in 3D coronary reconstruction from non-simultaneous DSA views, with a physiology-driven SVF simulation pipeline and a dual-head alignment network. The problem is clinically important and underexplored, and the sim-to-real transfer is encouraging.
R2 observed that the clinical test results show one view well-aligned but the other completely off, and found the overall clinical performance unconvincing for practical use. This is a concrete observation that the authors need to explain.
R1 noted that the method is only compared against uncorrected raw input with no rigid, affine, or deformable registration baselines, making the contribution hard to contextualize. R1 also questioned the evidence supporting the SVF deformation model in Eq. 1 and Eq. 2, noting that the parameters are not explained. Both R1 and R3 asked about generalization to the left coronary system, which has more complex geometry than the RCA tested here.
For the rebuttal, the authors should mainly focus on address the following issues:
1.explain the asymmetric view performance on clinical data that R2 flagged; 2.justify the SVF deformation model with references or evidence, 3.discuss why no registration baselines were included, 4.comment on LCA generalization and inference time. The clinical results question from R2 is most important since it speaks directly to whether the method works in practice.
- 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.
Please understand this is a difficult decision as the study is valuable, innovative and clinically important, and I understand the challenges behind the data collection and evaluation processes. However, after a careful read of the manuscript, reviews and rebuttal, I believe The paper still falls below the bar, for the reasons below:
1.For a paper whose whole contribution is non-rigid alignment, it is never compared against any real deformable registration, classical or learned. Relabeling Raw as the rigid floor does not fill that gap, so the reported gain cannot be attributed to the proposed design rather than to adding any deformation. That was a standard experiment for the submitted paper, not barred new data, and R1 named it the most important weakness.
2.The clinical evidence is too thin to carry the claim: 10 cases (really too small cohort for meaningful significance analysis and solid conclusion), single site, large confidence intervals, and the second of only two views stays near 0.35 to 0.43 overlap after correction. The reconstruction collapses toward one view, so the two-view premise is only half demonstrated. This is the soundness objection R2 defended at score-2 after rebuttal.
3.Even on simulation ground truth the absolute reconstructi on quality is low, at 3D Dice 0.53.
The problem is genuinely underexplored and the idea is sensible, but the submitted record does not show the method works. I really hope can see a more completed work done and published in a high impact conf or journal.
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.
Although some concerns remain regarding limited comparisons with additional public-code baselines, insufficient validation on LCA or other arteries, and the reliability of second-view metrics without phase matching, the rebuttal addresses most major issues and the paper presents a promising sim-to-real translation approach for a clinically relevant problem. Therefore, the AC recommends acceptance.
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.
The paper proposes a method to reconstruct 3D coronary images from 2D DSA views. The reviewers agree that the paper is interesting. There are some remaining issues after the rebuttal but 2 out of 3 reviewers are satisfied with the responses and recommend acceptance. I feel too that the authors addressed the main concerns.