List of Papers Browse by Subject Areas Author List
Abstract
X-ray coronary angiography is the clinical gold standard for coronary artery disease during real-time cardiac interventions, but provides only 2D projections of inherently 3D vessels. Existing learning-based 2D-to-3D reconstruction methods typically require wide angular coverage or multiple views, assumptions that are rarely satisfied in routine practice where only two projections with narrow angular separation are available. To address these challenges, we propose NeCA++, a multi-stage self-supervised neural radiance field (NeRF) framework tailored to clinically realistic acquisition constraints.
The framework decomposes reconstruction into two stages that progressively refine spatial support and representation capacity. In the first stage, a coarse 3D representation of the vasculature is reconstructed, restricting the subsequent optimisation to regions with a higher likelihood of vessel presence, termed an active region. Afterward reconstruction is restricted to this region while higher-resolution representations are progressively activated to recover fine vascular details. This multi-stage strategy focuses learning on anatomically plausible regions, mitigates gradient dilution under extreme sparsity, and stabilises global topology before recovering fine vascular branches. Furthermore, two vessel-specific regularisations are introduced: a ray-aligned constraint to reduce projection-induced ambiguity, and a bimodal density penalty to enable early vessel-background separation. Extensive experiments across three datasets (ImageCAS, ASOCA, and Synthetic RCA) and four angular configurations demonstrate consistent superiority over state-of-the-art baselines, particularly under clinically realistic narrow-angle settings, while achieving reconstruction within 58 seconds per case. The code is available at https://github.com/MultiMeDIA-Oxford/NeCAPP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5185_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MultiMeDIA-Oxford/NeCAPP
Link to the Dataset(s)
ImageCAS dataset: https://www.kaggle.com/datasets/xiaoweixumedicalai/imagecas/
BibTex
@InProceedings{MenDey_Multistage_MICCAI2026,
author = { Meng, Deyu AND Lashgari, Mojtaba AND Wang, Yiying AND Banerjee, Abhirup},
title = { { Multi-stage NeRF for Efficient 3D Coronary Artery Reconstruction from Two Narrow-Angle Angiographic Projections } },
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
The authors propose an extension to prior works for the reconstruction of 3D coronary trees from 2D projections. Existing worlds [14,15,24] assume orthogonal angle separation, which are not as common in routine clinical practice. The authors propose a method for narrow angular separation. The technical contributions are a coarse to fine approach that first identifies the vascular region and established topology before refine high resolution details in a second step. The authors propose additional regularisation losses that mitigate issues caused by the narrow angular planes.
- 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 paper reads very well. Clear and concise, but also precise technical detail where needed. Elegant figures. Overall top presentation.
The clinical motivation for a method which uses narrow-angle angiographic projects is solid, and the technical contributions are a direct response to this. All contributions are validated with experiments, and the improvement in performance trend when acquisition is changed from orthogonal planes is good evidence that the method fits its purpose.
The ray aligned regularisation loss to reduce ambiguity in the gradients is a super interesting idea and drives most of the performance gains.
The improvement in efficiency as a result of the multi-stage training is significant.
- 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.
One of the key arguments of the work is that it is solving a problem in XA which is more aligned to real world clinical settings. However, the authors do not use real XA data. Rather they use 2D X-ray projections generated from 3D CCTA volumes. I understand there is a large overlap in the physics and acquisition of CCTA and XA, and the authors try to reproduce X-ray acquisition protocols, but there are also dissimilarities such as the appearance of contrast and tissue textures, different motion artifacts, calibration error. Will the losses and regularisations proposed be equally effective on real XA images as they are on CCTA-derived X-ray images? This is not evidenced.
Perhaps I am missing something here but 0.99 dice similarity on both ImageCAS and ASOCA seems to good to be true? The dice similarity for SOTA segmentation methods on these datasets is ~0.85 dice. Perhaps this is due to the strong ~6x downsampling on the ground truths to 80x80x80? This is extremely coarse - I dont think its possible to resolve a 50% stenosis at this resolution. The authors should report the dice by upsampling their predictions rather than downscaling the ground truth.
The qualitative results reveal that the predicted reconstructions are very low quality. I am skeptical whether these are good enough to be clinically useful either to analyse stenosis or other downstream tasks. For example none of the predictions have anatomically plausible surfaces and contain many topological errors. At 80x80x80 resolution, I don’t think any clinically relevant biomarkers can be extracted. It begs the questions: what can this method actually be used for?
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
“X-ray coronary angiography is the clinical gold standard for CAD assessment” - I don’t believe this statement to be true anymore. CCTA is now the preferred first-line test in many guidelines (ESC, NICE).
“. . . they typically rely on additional calibration steps” - in real world data XA data, calibration is a necessary step due to noise in the acquisition parameters.
“For example, DeepCA [23] employs a generative adversarial network (GAN) [8] for 3D CA reconstruction. However, such supervised approaches rely on a substantial amount of paired training data. ” - Do you mean that lumen segmentations are required to train the model?
Any reason why only the LAD and RCA were used, and not the LCx? It is a main branch and often imaged in XA. From Fig 2 it seems there are multiple branches for each RCA and LAD. Perhaps the data you have used are the left and right coronary trees rather than single branches?
- 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?
This is a technically sound and excellently presented work. I have some doubts as to the clinical usability of this work due to the quality and size of the outputs, and choice of data. This is important as one of the central claims is that they are aligning to clinical protocols. The authors should have an opportunity to defend their choices in the rebuttal.
- 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 low resolution (80³) means it can only resolve vessels of minimum ~2mm diameter (0.35mm x (512px/80px)), which excludes most of the tree, and means no realistic biomarkers can be extracted. It is important to acknowledge that in its current state the algorithm is unusable. However, the authors do make a meaningful step towards a full solution by considering the limitations of narrow-angle angiography. There is good evidence that their contributions reduce this issue. The paper should be accepted based on this despite its usability.
Review #2
- Please describe the contribution of the paper
This paper tackles a highly practical and under‑explored problem in clinical coronary angiography: reconstructing 3D coronary geometry from only two narrow‑angle X‑ray projections, a setting closer to real-world cardiac intervention practice than most prior NeRF‑based approaches. The proposed NeCA++ introduces a well-motivated two‑stage optimization and vessel-specific regularization to address severe sparsity and ambiguity caused by narrow angular separation. The contribution is technically sound and practically relevant, though some methodological aspects and evaluations could be clarified or strengthened.
- 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.
1.Addresses clinically relevant constraints neglected by previous NeRF-style approaches. 2.Multi-stage spatial focusing effectively mitigates gradient dilution and improves optimization stability. 3.Ray-aligned constraint is physically meaningful and directly tackles projection ambiguity—a consistently overlooked issue. 4.Bimodal penalty improves early vessel-background separation, accelerating convergence. 5.Implementation based on multi-resolution hash encoding (InstantNGP) ensures high computational efficiency while keeping self-supervised learning feasible per case. 6.Experimental results are comprehensive across datasets; quantitative gains under narrow-angle conditions are significant.
- 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.
1.The method reconstructs static geometry; however, coronary angiography images are affected by motion. How robust is NeCA++ under minor temporal misalignment between the two projections? Could motion compensation or temporal regularization be added later? 2.How sensitive is reconstruction quality to the binarization threshold and morphological dilation used to define the active region? 3.Is there any quantitative validation that stage‑1 localization does not miss thin branches or small distal arteries? 4.The λ values (7.5 for $L_{ray}$, 2 for $L_{bin}$) are fixed “experimentally.”
Was any grid search or sensitivity study performed? Do these values generalize across datasets and angular settings? 5.How were the number of levels $L$ and feature dimensions per level chosen? Did the authors evaluate if adaptive activation (rather than sequential activation) could further accelerate convergence? 6.Only three baselines are compared. Are there recent NeRF acceleration techniques (e.g., mip‑NeRF 360, KiloNeRF, tensor field NeRF) or other sparse‑view 3D reconstruction approaches that could serve as stronger baselines? 7.The results are purely quantitative. Would subjective assessment by clinicians (e.g., vessel continuity or stenosis identification accuracy) corroborate the improved geometry fidelity? 8.The method still assumes known projection matrices $P_1$, $P_2$. In real practice these may be noisy or partially uncalibrated—how tolerant is NeCA++ to such calibration errors? 9.Does the learned structure exploit prior coronary shapes implicitly? Could the approach reconstruct other vascular regions (e.g., cerebral or peripheral) without major modification?
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
The paper describes the network architecture, multi-stage optimisation procedure, loss definitions, regularisation terms, and experimental configuration in considerable detail. Although no anonymised code or data link is provided, the method and training settings are sufficiently documented to permit re‑implementation by an expert in NeRF‑based reconstruction. Providing source code upon acceptance would further enhance reproducibility.
- 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.
(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?
The paper contributes a clinically meaningful advancement of NeRF‑based 3D angiography reconstruction, introducing principled constraints and an efficient optimization scheme. Nevertheless, several conceptual and practical questions remain regarding robustness, parameter selection, and generalization. Addressing these issues in the rebuttal would strengthen the case for acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This article proposes NeCA++, a multi-stage self-supervised neural implicit framework for reconstructing 3D coronary artery structures from two narrow-angle angiographic projections. The method introduces a coarse-to-fine optimisation strategy with an active-region constraint and vessel-specific regularisations to improve reconstruction stability under severely limited input views.
- 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 article addresses a clinically relevant and highly ill-posed problem reconstructing 3D coronary structures from two narrow-angle projections and proposes a structured solution. The multi-stage coarse-to-fine optimisation with active-region restriction is well-motivated and improves both efficiency and stability. The ray-aligned constraint and bimodal density regularisation explicitly target projection ambiguity and vessel-background separation. The method demonstrates strong robustness across decreasing angular separations, outperforming baselines under realistic settings, while also achieving significant computational efficiency.
- 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 article does not provide a theoretical analysis of the fundamental ambiguity of reconstructing 3D structure from two narrow-angle projections, leaving identifiability unclear. All experiments rely on synthetic projections without validation on real clinical data, which limits their practical credibility. The reconstruction resolution (80x80x80) is relatively low for medical applications, and scalability to higher resolutions is not discussed. Additionally, the method performs per-case optimisation rather than learning a generalisable mapping, and the proposed regularisation terms remain heuristic without deeper theoretical justification.
- 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.
- 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?
Not applicable
- 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
None
- 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The limitations in experimental validation and the lack of deeper analysis of the underlying inverse problem reduce its overall strength.
- 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.
The theoretical weakness of the article can not be justified by the rebuttal. It’s the fundamental flaw in the article’s problem design.
Author Feedback
We wholeheartedly thank the Area Chair and the three Reviewers for their constructive suggestions. We will carefully revise the manuscript accordingly and include the code link.
Low-resolution reconstruction [AC, R1, R4]: The high Dice scores are indeed influenced by downsampling. Following Reviewer’s suggestion, we evaluated by upsampling the generated reconstructions to 256^3, achieving Dice scores of 0.71 and 0.78 on the ImageCAS and ASOCA datasets and outperforming the best competing approach by 30% and 27%, respectively. We completely agree that upsampling the reconstruction does not preserve smaller anatomical details. We will discuss this limitation and explore more efficient representations, such as point clouds, in our future work.
Parameter selection and sensitivity [AC, R2 (Q2 - Q5)]: We did sensitivity analysis on the binarization threshold, morphological dilation, loss weights, and hash levels. A threshold of 70% of the maximum grey value attained the best performance, while varying it by ±10% caused only minor changes in Dice (2.1%). Cube dilation sizes of 0, 3, and 6 voxels produced nearly identical metrics, suggesting that binarization also captures thin and distal arteries. For possible discontinuity in segmentation, we retained dilation. Varying λ_ray from 6.5 to 8.5 and λ_bin from 1 to 3 altered Dice by <2.3%. We tested 12-18 hash levels and found that increasing beyond 16 improved Dice by <0.001.These parameters were selected experimentally on validation set and applied consistently across datasets and angular settings.
Generalization to Clinical X-ray Angiography (XRA) [AC, R1, R4, R2 (Q1)]: Use of 2D projections from real CCTA data enabled quantitative evaluation of the proposed reconstruction losses and regularizations whilst isolating the impact of motion artifacts. The challenges addressed in this work arise from acquisition geometry and vascular anatomy, which also exist in clinical XRA. So, the proposed framework can be applicable to clinical XRA following a motion correction step [1].
Performance under uncalibrated XRA and other vascular regions (e.g., cerebral or peripheral) [R2 (Q8, Q9), R4]: Due to its self-supervised learning framework, NeCA++ performs patient-specific deformation modelling rather than learning a generalizable mapping, making it largely modality-insensitive. Additionally, ray regularization and multi-view consistency constraints improve robustness to calibration perturbations. The method can also be extended to other vascular regions with minimal modification.
LCx reconstruction [R1]: We acknowledge that we used both left (including LCx) and right coronary trees rather than single branches.
Lack of comparison with recent NeRF (e.g., mip-NeRF 360, KiloNeRF) [R2 (Q6)]: These methods are designed for natural scenes and rely on photometric consistency assumptions that do not hold for angiography.
Subjective assessment by clinicians [R2 (Q7)]: Clinical assessment is valuable, and we plan to conduct an inter-observer study in future work.
Theoretical ambiguity of reconstructing 3D structures from two narrow-angle projections [R4]: The theoretical ambiguity has been previously analysed [1], which is now explicitly cited and discussed in the revised manuscript.
Justification of regularization term [R4]: Section 2.3 discusses that L_ray promotes continuous vascular structures while suppressing outliers through gradient constraints, and L_bin encourages binary occupancy through a quadratic formulation.
Diagnostic value of CCTA for CAD [R1]: We agree and will revise the Introduction accordingly.
XRA calibration [R1]: We agree that XRA calibration is a necessary step due to noise in the acquisition parameters. We will revise the Introduction accordingly.
Comparison with DeepCA [23] [R1]: Our self-supervised NeCA++ requires only 2D projections for 3D reconstruction, whereas supervised methods such as DeepCA rely on paired 3D and 2D XRA images for training.
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 paper proposes a method for 3D reconstruction of the coronary artery from 2 narrow-angle x-ray projections, relying on implicit reconstruction (NeRF) techniques. The method is tested on synthetic and real data, where 2D X-Ray projections are simulated from 3D CCTA volumes.
All reviewers agree that the paper is generally well presented, well motivated, and proposes a sound solution for a relevant problem.
Reviewers mention concrete and actionable issues, where rebuttal response would influence final scores (R1,R2):
- What is the clinical usability of the proposed method when the generated reconstructions are very low-resolution (R1,R4)?
- Further details on parameter selection and how they impact result sensitivity (R2)
- Gap between real X-ray data and the 2D projections from 3D volumes utilised in the experiments. How much confidence do we have that this method can work on real X-rays? (R1,R4)
- 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.
Disagreement remains after rebuttal (2 accept, 1 reject).
I am leaning towards recommending acceptance on the basis that the limitations in dealing with low-resolution volumes have been more clearly articulated in the rebuttal, satisfying R1, R2.
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.
After rebuttal, all reviewers maintain their original scores, with two acceptance and one rejection. One reviewer suggests acceptance did not submit the final score and comments. One reviewer insists that there are fundamental flaws in the problem design, and the other review also think its usability is not that convincing.
Comprehensively consider the rebuttal and the comments raised by the reviewers, this paper cannot be accepted in its current shape.
It is suggested that the authors further consider reviewers’ suggestions and comments, revise the paper, and submit it to another conference or journal.
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 has a critical criticism in terms of theoretical justification. But reviewers agreed this is an interesting application and may have a potential impact for artery reconstruction. Considering that two experts are happy with their feedback. I am lean to recommend acceptance because of the potential application.