List of Papers Browse by Subject Areas Author List
Abstract
Accurate correspondence matching across multiple angiographic views is the prerequisite for 3D coronary reconstruction and interventional guidance. However, the development of robust deep learning models for this task has been stifled by a fundamental data bottleneck. Obtaining ground truth for matching tasks in angiography pairs is prohibitively expensive and hard to scale. To overcome this barrier, we introduce a physically-grounded data generation framework that synthesizes high-fidelity Digital Reconstructed Radiographs (DRRs) from 3D Coronary CT Angiography (CCTA) volumes. Our framework generates dense, highly accurate 3D-to-2D correspondence labels by simulating realistic C-arm acquisition geometry on patient anatomy at zero human cost. Leveraging this dense supervision, we propose a Geometry-Informed Matching Module (GIMM) that integrates global feature and anatomical structure into correspondence learning. Unlike real angiography where assessment relies on subjective human annotation, our dataset provides 2D correspondence labels with paired images, allowing human-free evaluation. We comprehensively evaluate our method on the proposed CT-derived DRR dataset and demonstrate improvements over other matching baseline models. As the first public benchmark dataset for coronary correspondence matching, we will release our dataset and code in http://github.com/medipixel/GIMM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6167_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/medipixel/GIMM
Link to the Dataset(s)
https://drive.google.com/drive/folders/1JXRTGlTbWIbb-hvGqmrZnPXJZs86otZr?usp=sharing
BibTex
@InProceedings{LeeIn_AnatomyGrounded_MICCAI2026,
author = { Lee, In Kyu AND Seo, Sumin AND Min, Jaesik},
title = { { Anatomy-Grounded Synthetic Coronary Angiography for Geometry-Informed Multi-view Matching } },
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 a core limitation in learning-based multi-view correspondence modeling for coronary angiography—the lack of large-scale ground truth labels—by introducing a physically grounded synthetic DRR dataset generated from patient CCTA scans and a Geometry-Informed Matching Module (GIMM) that integrates anatomy-specific priors and projection geometry. The dataset contribution is data-centric yet high-impact, offering dense pixel-level supervision at zero manual cost and serving as a valuable benchmark for downstream 3D reconstruction and registration tasks. Methodologically, GIMM incorporates C-arm view conditioning (via FiLM) and epipolar geometric gating—concepts that, although individually established, are effectively adapted to the medical imaging context. By linking physics-based data generation with geometry-aware learning, the work makes a meaningful step forward in a domain long constrained by data scarcity, though the algorithmic novelty itself remains incremental.
- 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.The CT‑derived DRR dataset provides a realistic and scalable foundation for supervised learning of angiographic correspondences. Ground‑truth 3D–2D labels derived from known projection geometry eliminate the need for costly human labeling. 2.The epipolar gating and C‑arm aware conditioning explicitly encode the underlying imaging geometry, moving beyond purely appearance‑driven matchers such as LoFTR and SuperGlue. 3.Ablation tables are clear, and both 2D and 3D quantitative metrics are defined with anatomical consistency. 4.The method’s mathematical formulation (weighted 8‑point estimation, FiLM modulation) is rigorous. Implementation details are sufficient for replication. 5.Public release of a coronary DRR dataset fills a critical gap—prior NeRF or angiography datasets lacked dense correspondence 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.
1.The entire training and quantitative evaluation are on synthetic DRRs. Although qualitative visualizations on real angiograms are shown, how robust is GIMM to domain appearance shift between simulated DRRs (clean, noise‑free) and clinical X‑ray images (noisy, cluttered, device artifacts)? Did the authors consider applying domain adaptation (e. g. , style transfer, adversarial adaptation, self‑supervised finetuning) when testing on real clinical data? Can the proposed dataset generation pipeline simulate varying contrast levels, patient motion, or fluoroscopic noise models for realism? 2.Since the synthetic dataset is generated from CCTA volumes using known geometric parameters, it may perfectly align with the model’s built‑in geometric assumptions. How does the model perform when C‑arm angle metadata is perturbed or slightly inaccurate (to simulate real calibration errors)? Have the authors tested the model when epipolar geometry is estimated with mild noise, as occurs in clinical practice where detector parameters are not perfectly known? 3.In the class‑conditioning branch (FiLM): how sensitive is the model to the number or grouping of “view classes”? Could a continuous embedding of projection geometry outperform discrete class embeddings? The epipolar gating employs a hard threshold δ = 2 pixels—was a sensitivity analysis performed? How are outliers handled when the 8‑point solution fails due to degenerate correspondences? The top‑K filtering strategy makes comparison easier, but it may hide behaviour on dense correspondences. Could authors report average precision across all matches, not only top‑20? 4.The reported 3D error (≈1.17 mm) is encouraging, but how does this magnitude translate clinically (e. g. , vessel diameter, bifurcation localization)? Qualitative projections on real angiograms are visually similar, but without quantitative validation or expert scoring, the real‑world effectiveness remains uncertain. Could authors describe potential integration with downstream 3D reconstruction pipelines? Does improved matching accuracy directly translate into better 3D angiographic reconstructions? 5.The 26k pair dataset is derived from ImageCAS CCTA. How are patient demographics and anatomical variations represented? Is there a risk of bias (dominant RCA vs. LCA, lesion severity, etc. )? Are projection parameters publicly released, and can new users reproduce exactly the same projection geometry to generate matching 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 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
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 tackles the problem that the lack of dense ground‑truth supervision for multi‑view coronary angiography matching, and presents a well‑executed solution that combines physics‑based synthetic data generation with geometry‑aware learning. The proposed CT‑derived DRR dataset is both scalable and anatomically consistent, enabling supervised learning without manual annotation and offering potential community impact as a benchmark resource. The Geometry‑Informed Matching Module (GIMM), while conceptually incremental, is methodologically sound and achieves measurable performance gains over strong baselines through the integration of C‑arm view conditioning and epipolar constraints. The work is clearly written, experimentally thorough on synthetic data, and scientifically grounded.
However, the reliance on synthetic data limits quantitative validation on real angiography, and the algorithmic novelty—though well contextualized—remains modest. Stronger evidence of real‑world generalization or additional analysis of model robustness would significantly strengthen the submission. Overall, the paper is above the acceptance threshold due to its practical significance, technical soundness, and potential dataset contribution, but a borderline case dependent on how well the authors address generalization concerns in rebuttal.
- 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 #2
- Please describe the contribution of the paper
This paper proposes a new matching method for coronary angiography that incorporates the class-conditioning module and the epipolar geometry module.
- 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.
- Given the difficulty of obtaining annotations for real angiography data, the authors trained and evaluated the proposed network using a synthetic CT-derived digital reconstructed radiograph (DRR) dataset.
- The comparative evaluation and ablation study demonstrate the effectiveness of the proposed method on the synthetic DRR dataset.
- 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.There are already a number of existing methods for DRR generation, including the following work: “Fast Auto-differentiable Digitally Reconstructed Radiographs for Solving Inverse Problems in Intraoperative Imaging” However, Section 2.1 does not provide sufficient citation or discussion of related DRR generation methods. Since the dataset is presented as one of the contributions of this work, the authors should clarify how their DRR generation approach is similar to, or different from, prior methods.
2.Although the proposed method achieves competitive results on the synthetic DRR dataset, the paper does not provide quantitative evidence on real data. Only a limited number of qualitative examples are shown in Figure 4.I understand that annotation of real data is costly and challenging; however, even sparse manual annotations, such as correspondences between bifurcation points, could provide meaningful quantitative validation. In its current form, the paper provides very limited detail regarding the real-data experiments, and the lack of evidential results on real data makes the study feel incomplete.
3.I found that several parts of the paper require clearer explanation.
- Figure 2 is difficult to interpret due to insufficient explanation. In particular, the meanings of M_c, M_f, and E are not clearly defined, even in the main text.
- The implementation details section provides only part of the necessary experimental details.
- The definition of the distance metrics is also a little unclear. In particular, it is difficult to understand what exactly is meant by the “centerline.” It is not clear whether this refers to a centerline segment or to individual pixels/points on the centerline.
- Please rate the clarity and organization of this paper
Poor
- 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?
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?
While the proposed method appears promising on the synthetic DRR dataset, the paper has notable limitations in terms of related-work positioning, experimental validation on real data, and clarity of presentation. In particular, the absence of quantitative evidence on real angiography data significantly weakens the overall contribution.
- 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 paper has significant limitations regarding related work, experimental validation on real data, and clarity.
Review #3
- Please describe the contribution of the paper
This paper addresses the problem of multi-view correspondence matching in coronary angiography. It proposes a Geometry-Informed Matching Module (GIMM) integrated into a coarse-to-fine matching framework.
- 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.To alleviate the lack of ground-truth correspondence in real angiography, the authors construct a CT-derived Digital Radiographic Reconstruction (DRR) dataset based on CCTA data and provide dense 3D–2D projection labels. Experiments on this dataset show that the proposed method achieves improvements over several baseline methods. 2.The paper addresses a clinically relevant problem in coronary angiography with clear practical motivation. 3.The idea of generating dense 3D–2D correspondence labels from CCTA via DRR synthesis is interesting and provides a controlled experimental setting for this task. 4.Incorporating view information and geometric constraints into the matching process is intuitively reasonable.
- 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.A key contribution of this work is the construction of a CT-derived DRR dataset with dense 3D–2D correspondence labels. However, the dataset is entirely synthetic, and the paper does not provide any validation of its realism or its consistency with real coronary angiography data. In addition, there is no cross-domain evaluation (e.g., synthetic to real) to demonstrate that models trained on this dataset generalize to real clinical data. Therefore, it remains unclear whether the dataset can serve as a valid proxy for the real-world problem. 2.The proposed GIMM is built upon existing matching frameworks, and its core components largely consist of combining and extending existing ideas. In its current form, the method appears to be an incremental improvement under strong supervision rather than introducing a fundamentally new modeling approach. 3.All quantitative experiments are conducted on the synthetic DRR dataset constructed by the authors, with no quantitative evaluation on real coronary angiography data. Meanwhile, the paper discusses the clinical significance of the results. However, such conclusions are drawn from experiments on idealized synthetic data, which do not account for noise, motion, and imaging artifacts present in real angiography. Therefore, the current experimental setup does not sufficiently support the claimed clinical applicability. 4.The ablation study is conducted by progressively adding components on top of QuadTree. However, the rationale for selecting QuadTree as the base for ablation is not clearly explained. Additional justification would be helpful. 5.In Table 2, “Epipolar Gating (Soft)”, “Epipolar Gating (Logit)”, and “Ours” are all marked with the same module configuration (two checkmarks), but the distinctions between these configurations are not explicitly clarified. This makes the comparison difficult to interpret and should be clarified. 6.The discussion of limitations of existing methods lacks direct experimental or literature support. In addition, attributing the difficulty of the task mainly to the lack of ground truth appears somewhat simplified. Furthermore, in the results section, the error is compared with coronary artery diameter to argue clinical significance; however, this conclusion is based on synthetic data and should be stated more cautiously.
- 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 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?
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.
(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 problem addressed (multi-view correspondence matching in coronary angiography) is highly relevant and challenging in the clinical domain. The integration of geometry (view information, geometric constraints) into the matching process is a reasonable and potentially effective approach for improving correspondence accuracy.
- Reviewer confidence
Very confident (4)
- [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 have provided a constructive response to the reviewers’ concerns, and their future revisions will sufficiently address the key issues raised during the review process. In particular, the discussion on real-data validation is thoughtful and realistic. While dense pixel-level ground truth is indeed infeasible in angiography, the authors justify their alternative evaluation strategies well, including sparse landmark validation and consistent trends between real and synthetic data. This demonstrates that the proposed method generalizes beyond purely synthetic settings.
On the evaluation side, the authors provide a reasonable justification for using top-K metrics instead of average precision, given the extreme density of potential correspondences in their formulation.
Finally, the commitment to improve clarity, provide implementation details, and release code and dataset generation procedures will significantly enhance reproducibility.
Overall, the paper presents a novel and well-motivated GIMM for 3D coronary artery reconstruction with solid technical contributions.
Author Feedback
We thank the reviewers for their valuable feedback and recognizing the contribution of our study including anatomically grounded CT DRR for correspondence matching and geometry-aware learning conditioned on clinical prior. We acknowledge the concerns regarding real-data validation, clinical claims, DRR-related prior work, and reproducibility. We will address these points in the revision as follows. For R1/2/3, we agree that clinical claims should be cautious and require quantitative validation on real angiography. However, dense pixel-level GT would require annotating correspondences for nearly every pixel, which is infeasible in real angiography. A more practical alternative, used in prior work, is to manually judge whether each predicted match is correct. Yet this still requires model-specific annotation because each method predicts different source points. Thus, comparing N methods with K predictions per image requires up to N×K manual labels, making real-data validation costly and difficult to scale. As shown in Fig 4, we had performed a sparse real data analysis on 150 real image pairs with manually annotated bifurcations. We omitted it because sparse landmarks are limited for assessing dense correspondence quality. Nevertheless, the result supports the same trend observed in synthetic DRR. GIMM trained on our synthetic dataset achieved the best F1@10px, 0.332, compared with LoFTR 0.300 and QuadTree 0.308 trained on the same data. For R2/3, we acknowledge missing citations for prior DRR methods. We will add relevant references and clarify that our contribution is not a new DRR renderer, but a CT DRR framework to generate angiography-like image pairs with labels for matching. We will also note in the Discussion that our framework can incorporate modern fast or differentiable DRR renderers to improve realism, efficiency, and future domain adaptation to real angiography. For R1, we agree that evaluating only top-20 candidates may not fully reflect dense correspondence behavior. However, computing AP requires a finite GT correspondence set, but in our setting GT matches can be sampled from any centerline point, yielding extremely dense labels and poorly interpretable recall. Instead of AP, we report top-K performance across varying K values (Fig. 3), which provides an intuitive comparison of performance at the same number of predicted matches across models. R1 raised the concern regarding robustness to imperfect C-arm geometry and calibration errors in clinical settings. Our DRR generation already includes random perturbations in primary and secondary angles, exposing the model to intra-class geometric variability. In addition, epipolar gating estimates geometry from tentative matches rather than preset projection angles. Thus, small C-arm angle errors are unlikely to directly degrade matching. We will clarify this and consider systematic calibration robustness analysis in future work. R1/3 raised the questions regarding threshold selection and the distinction between different gating formulations. The 2-pixel threshold was chosen to enforce strict geometric consistency by filtering inconsistent correspondences while preserving valid local matches. If fundamental matrix estimation fails, the batch is marked invalid and epipolar gating is disabled, so matching relies on feature similarity. Table 2 variants differ only in gating function: soft gating uses exp(−d/τ^2), logit gating uses a sigmoid, and our model uses hard thresholding. A broader threshold sensitivity analysis remains future work. For R1/2, we agree that implementation and notation details should be clearer. We will clarify that Mc and Mf denote coarse and fine matches, E denotes expectation following LoFTR, and “centerline” refers to individual centerline points. We will provide dataset composition, projection parameters, and full DRR generation/training code as stated in the submission for reproducibility.
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.
The reviewers acknowledge a useful dataset contribution and a sensible geometry-informed matcher, with consistent gains over strong baselines on the synthetic benchmark. Two are weakly positive, finding the method reasonable and valuing the dataset, while one is strongly negative on grounds of insufficient context relative to state-of-the-art differentiable-DRR work, the absence of quantitative real-data evaluation, and clarity issues (undefined Fig 2 symbols, ambiguous centerline metric, thin implementation details). Notably, one of the positive reviewers also shares the synthetic-only concern, flagging that the clinical-significance claim is drawn from idealized data and should be stated more cautiously. These concerns warrant a rebuttal phase.
For the rebuttal, the most pressing items are the absence of quantitative real-data validation and an explicit differentiation of the DRR generation pipeline relative to prior differentiable-DRR work. Further items raised by the reviewers — clarifying Fig 2 notation and the centerline-metric definition, robustness analysis under perturbed C-arm angles, and reporting AP/precision-recall curves beyond top-20 only — should also be addressed, alongside a clarification of the clinical-translation claim.
- 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 presents a CT-derived DRR dataset for multi-view coronary angiography matching with dense 3D–2D supervision, alongside a Geometry-Informed Matching Module (GIMM) integrating C-arm view conditioning and epipolar constraints, with consistent gains over strong baselines on the synthetic benchmark. After rebuttal, R1 remains weakly positive, R3 maintains accept, and R2 maintains reject on the grounds of insufficient differentiation from prior differentiable-DRR work, absence of quantitative real-data evaluation, and presentation clarity. I share several of R2’s concerns — particularly the limited specification of methodological details, including basic network structures — though I find the synthetic-data reliance less pressing in light of the authors’ justification for sparse landmark-based validation. R2’s post-rebuttal justification, however, is relatively brief and does not fully develop these points against the more detailed positive endorsements. I recommend acceptance, with the expectation that the camera-ready version substantially expand the methodological description and incorporate quantitative sparse-landmark validation on real angiography.
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.
While the paper received diverse opinions from the reviewers, I do agree with the major weakness identified by the reviewers that the method failed to justify the performance quantitively on real data. The authors argue in the rebuttal that justification in real data is very challenging, but without this, it is difficult to convince others. The authors may want to think about justification in sparsely annotated data quantitively. Another major weakness is that the authors did not compare with other generation methods while the claimed contribution is the data synthesis. Therefore, I don’t think the method is properly justified.
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.
This is a borderline paper with mixed reviews. Reviewers agreed that the paper addresses a clinically relevant and technically challenging problem of establishing dense 3D–2D correspondences for coronary angiography, and that the proposed geometric integration of epipolar constraints and C-arm awareness yields clear improvements over appearance-based baselines. However, they also raised serious concerns about the paper’s significance and soundness: (1)the entire evaluation relies on synthetic DRR data, with no quantitative validation on real angiograms to demonstrate robustness to clinical noise, artifacts, and calibration inaccuracies; (2)the methodological novelty is moderate, as GIMM primarily combines and extends existing components rather than introducing fundamentally new modeling principles; and (3)the realism and generalizability of the proposed dataset remain unverified, and several experimental details—such as ablation baselines, metric definitions, and sensitivity analyses—lack clarity or completeness. The rebuttal offered thoughtful justifications and committed to improving clarity and releasing code and dataset procedures, but did not fully resolve concerns regarding real-data validation and novelty. After careful deliberation, I recommend accepting this manuscript, as the core contribution—explicitly embedding imaging geometry into dense correspondence learning—represents a principled and practically valuable advance for 3D coronary reconstruction, with strong potential to generalize beyond the synthetic setting.