acceptodds
Under review as a conference paper at ICLR 2027

Projection-Refined Parametric 3D Coronary Artery Reconstruction from Sparse X-ray Angiographic Views

Abstract

Sparse-view 3D coronary reconstruction is commonly formulated as cross-view correspondence and triangulation, which is vulnerable to vessel overlap and foreshortening, or on volumetric prediction followed by vascular-graph extraction, which does not natively provide centrelines and radii. We address both limitations with a coarse-to-fine framework that avoids explicit point matching and triangulation while directly predicting a branch-structured centreline-and-radius representation instead of an intermediate volume. From a variable number of segmented views, a coarse predictor combines frozen VGGT features with learned branch queries to estimate active branches, their B-spline centreline trajectories, and approximate dense radius profiles. Projection-guided geometry and radius refiners sample input view evidence and predict residual corrections learned with 3D supervision. We evaluate representation fidelity and sparse-view 3D reconstruction quantitatively and qualitatively. On simulated angiographic masks generated from CT-derived coronary anatomy, our method achieves the best structural and centreline metrics among the evaluated methods while maintaining competitive volumetric overlap. Coarse-to-fine inference takes ms, enabling real-time reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D coronary artery reconstruction, Sparse-view reconstruction, Learning-based medical image reconstruction, Structured anatomical representation, X-ray coronary angiography, Vascular modelling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iYtXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a coarse-to-fine method for 3D coronary reconstruction from segmented X-ray views with known acquisition geometry. It uses frozen VGGT features to predict branch existence, B-spline control points, and radii, followed by projection-guided geometry and radius refinement. Experiments use simulated projections, with additional qualitative examples on real angiography.

Critical strengths

The method provides centrelines and radii through a compact representation that keeps each branch connected. It also shows promising results compared with the tested baselines.

Critical weaknesses

The contribution and cost of the proposed design need stronger evidence. Runtime comparisons do not include other VGGT-based methods, and Table 6 does not report runtime for the backbone variants. The tables also show that most gains come from geometry refinement, but the paper does not test whether a refiner trained from a simple initial shape can achieve similar accuracy without the coarse predictor. Such a comparison, including runtime, is needed to establish whether the coarse predictor is necessary.

The method and results are not described in enough detail. Section 3.3 does not specify which VGGT layer provides the features, Section 3.5 omits loss-weight values, and the differentiable rendering implementation remains unclear. Table 5 also does not make clear whether both variants include radius refinement. These gaps make the method harder to reproduce and the results harder to verify.

Questions for authors

What is the accuracy and runtime trade-off between the VGGT and ResNet-101 variants? Does the reported 121 ms include VGGT feature extraction, and how much time does that step take?

Are there other VGGT-based methods that could be compared under the same input and output settings? If so, why were they not included in the accuracy and runtime comparisons?

Can a refiner trained from a simple initial shape achieve similar accuracy without the coarse predictor, using the same image features? Please compare accuracy and runtime.

Which VGGT module and layers provide the features? Please also give the loss-weight values and explain how differentiable rendering is implemented.

Do both variants in Table 5 include radius refinement? What are the radius errors before and after refinement?

1 comment

We thank the reviewer. We have added backbone timing, a template-initialisation ablation, implementation details, and radius-error evaluation. The original Tables 5 and 6 are now Tables 6 and 7. CD denotes branch-level Chamfer distance; CC count is the number of connected components.

Weakness 1 and Question 1: Backbone accuracy and runtime

Table 7 reports accuracy after geometry and radius refinement. The timings below include online feature extraction and coarse prediction, using two masks on one RTX 4090 (Table 10; lines 826–833).

BackboneCoarse-stage time ↓RCA Dice (%) ↑RCA CD (mm) ↓LCA Dice (%) ↑LCA CD (mm) ↓
ResNet-1014.9 ms60.625.2645.199.21
VGGT60 ms66.924.7749.458.72

VGGT improves refined Dice by 6.30 for RCA and 4.26 for LCA, at approximately 55 ms additional coarse-stage cost.

Yes, the reported 121 ms includes online VGGT extraction: 58.1 ms for VGGT, 1.9 ms for coarse prediction, 14 ms for geometry refinement, and 47 ms for radius refinement (lines 1387–1392).

Weakness 1 and Question 2: Other VGGT-based methods

Related approaches include VGGT-Occ (arXiv:2605.16911) and VGOcc (arXiv:2607.18078). Like ours, they use frozen VGGT features with calibrated multi-view images, but target occupancy prediction in RGB driving scenes rather than coronary reconstruction. A meaningful comparison would require adapting and retraining these methods for projected artery masks, which we have not completed within the rebuttal period. As volumetric predictors, they would also need the same centreline extraction as AutoCAR and DeepCA, which we already compare against.

Our controlled backbone comparison isolates VGGT’s contribution within our framework. Even with ResNet-101, the refined model has lower branch-level CD and CC counts than the evaluated baselines for both arteries (Tables 2, 3 and 7). These advantages therefore do not depend on VGGT alone.

Weakness 1 and Question 3: Template versus learned coarse initialisation

We replace predicted geometry and radii with a per-branch mean template from the training set and retrain both refiners using the same procedure and image features. Coarse-predicted branch existence is retained (Tables 16–17; lines 1400–1453).

InitialisationRCA Dice (%) ↑RCA CD (mm) ↓LCA Dice (%) ↑LCA CD (mm) ↓
Template5.0720.563.7424.01
Coarse19.4710.9814.9214.40
Template + refiners57.906.2349.1410.34
Coarse + refiners66.924.7749.458.72

The refiners recover substantial accuracy from the template. Learned coarse initialisation nevertheless improves RCA Dice by 9.02 and reduces CD by 1.46 mm for RCA and 1.62 mm for LCA, although LCA Dice is similar. This supports a better geometric starting point, not an absolute requirement for refinement.

In an additional timing run of the template pipeline, VGGT takes 58.34 ms, the coarse predictor 1.18 ms, and both refiners, including preparation, 75.80 ms: 135.31 ms overall. Subtracting the coarse predictor gives an estimated 134.14 ms, not measured as coarse-free runtime: branch existence and branch tokens are still required, and VGGT features are reused by the geometry refiner. Thus, eliminating that computation would save only 1.18 ms in this run. These timings are separate from the original 121 ms benchmark.

Weakness 2 and Question 4: Implementation details

  • Features: We use the frozen VGGT-1B aggregator without its task-specific prediction heads. The final (24th) alternating-attention stage concatenates 1024-dimensional frame-attention and global-attention outputs into 2048-dimensional tokens (lines 791–797).
  • Loss weights: Table 8 lists all weights, including intermediate and final refinement-stage supervision.
  • Differentiable rendering: We construct perpendicular rings of 24 surface points at the predicted radii, with additional rings between centreline samples. These points are cone-beam projected and Gaussian-splatted onto the detector (σ\sigma = 1.25 pixels), producing a soft mask differentiable with respect to radius (lines 865–884).

Weakness 2 and Question 5: Radius refinement and accuracy

Yes, both dense-centreline and B-spline variants include geometry and radius refinement (Table 6 caption). Tables 2–3 report radius mean absolute error (RadMAE) and relative RadMAE before and after refinement:

ArteryCoarse RadMAE, mm (%)Refined RadMAE, mm (%)
RCA0.219 (12.48%)0.149 (8.65%)
LCA0.311 (15.55%)0.260 (13.05%)

Refinement reduces radius MAE by 32% for RCA and 16% for LCA, directly demonstrating radius improvement beyond changes in Dice.

Reviewer_VzKAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reconstructs 3D coronary arteries from a few calibrated vessel masks. A frozen VGGT encoder supports initial predictions of branches, spline centrelines, and radii. Two modules then use projection evidence to refine the geometry and radii. Quantitative tests use simulated views from ImageCAS. ASOCA and real angiograms provide qualitative examples.

Critical strengths
  1. Learned refinement gives a large improvement from the same coarse prediction. RCA Dice reaches 65.89%, versus 41.02% for direct projection fitting. This supports the value of learned corrections beyond the representation itself.
Critical weaknesses
  1. Better connectivity does not necessarily mean more accurate anatomy. With two views, Dice of the proposed mtheod is 49.45%, versus 56.56% for DeepCA in Table 9. clDice is also lower, despite better Chamfer distance. A connected tree can still have missing branches or incorrect junctions. Without branch and junction accuracy, I cannot tell whether better connectivity offsets the lower overlap.

  2. Radius recovery is central to the method, but its accuracy is not measured directly. Radius refinement adds about one Dice point in Table 10. The qualitative example does not establish accurate recovery of local narrowing. This leaves an important part of the method untested, even though radius labels are available.

Questions for authors
  1. What are the radius MAE and relative errors before and after refinement? Please report RCA and LCA separately.

  2. Does the method recover more correct LCA branches and junctions than DeepCA? This would help interpret the mixed results in Table 9.

  3. Does the reported 121 ms include online VGGT feature extraction? Does it also include the preprocessing required by the refiners?

1 comment

We thank the reviewer for these constructive comments. We have added branch-level geometry and radius metrics, per-branch and vessel-size comparisons with DeepCA, and a detailed runtime breakdown.

Metrics. Tables 2–3 now report branch-level Chamfer distance (CD, mm), ground-truth side-branch recall (BrRec, %), side-branch attachment-position error (AttErr, mm), and absolute and relative radius errors (RadMAE, mm; RadMAE_rel, %). Geometry and radius errors are computed on matched predicted and ground-truth branches; BrRec accounts for missing side branches. CC is the connected-component count. Definitions appear in Section 3.6, with branch matching detailed in Appendix A.3.2 (lines 1083–1159).

Weakness 1 and Question 2: Connectivity, LCA branches and junctions

We agree that connectivity alone does not establish anatomical accuracy. We therefore added branch-level evaluation in Tables 2–3 and a per-branch comparison with DeepCA in Tables 12–13.

LCADice (%) ↑CC ↓CD (mm) ↓BrRec (%) ↑Main branches matched (each out of 75) ↑AttErr (mm) ↓
DeepCA56.569.9719.9655.2567 / 5310.80
Ours49.451.0538.7288.7275 / 7512.89

Our model recovers more LCA branches and achieves lower branch-level CD, but DeepCA has lower LCA attachment error. For every branch entry evaluated by both methods, our mean CD is lower; main-branch CD is three to five times lower (Tables 12–13). Attachment errors are computed over each method’s matched branches, so the evaluated subsets differ.

DeepCA retains higher LCA Dice and clDice. Our advantage is more complete branch recovery and lower branch-level geometry and radius errors under the evaluated graph-extraction and branch-matching protocol. These provide evidence of anatomical recovery beyond connectivity alone, without implying superiority on every metric.

Weakness 2 and Question 1: Radius accuracy and narrowing

Tables 2–3 now report radius errors before and after refinement, separately for RCA and LCA. Entries below show RadMAE in mm, with relative RadMAE in parentheses.

ArteryCoarseRefinedBest comparison method
RCA0.219 mm (12.48%)0.149 mm (8.65%)AutoCAR: 0.406 mm (20.77%)
LCA0.311 mm (15.55%)0.260 mm (13.05%)DeepCA: 0.946 mm (40.04%)

Refinement reduces radius MAE by 32% for RCA and 16% for LCA. This directly demonstrates improved radius accuracy, which the small change in Dice alone does not capture.

We additionally evaluate radius error by ground-truth vessel size (Table 14; category definitions at lines 1293–1295). Each entry reports RadMAE in mm (relative RadMAE in %).

ArteryGT radius (mm)DeepCAOurs
RCA<1.50.292 (21.32%)0.157 (11.71%)
RCA1.5–2.50.656 (32.19%)0.131 (7.04%)
RCA>2.51.084 (39.12%)0.251 (9.13%)
LCA<1.50.348 (24.04%)0.260 (18.18%)
LCA1.5–2.50.686 (33.89%)0.225 (11.90%)
LCA>2.51.657 (57.56%)0.358 (12.49%)

Our model achieves lower absolute and relative radius errors in every category for both arteries. These errors concern matched branches; Table 14 also reports the number of evaluable cases.

The test set lacks sufficient annotated narrowing cases for systematic stenosis-specific evaluation. We cannot perform this evaluation with the annotations at hand and plan further evaluation once clinical data with stenosis annotations become available (lines 535–539).

Question 3: Runtime

Yes. For two input masks on one NVIDIA RTX 4090, the 121 ms includes online VGGT feature extraction (58.1 ms), coarse prediction (1.9 ms), geometry refinement (14 ms), and radius refinement (47 ms). Refiner preprocessing takes approximately 2 ms and is included in these stage timings, not added separately. Table 10 reports the total runtime, and lines 1385–1392 provide the stage breakdown.

Reviewer_SbkHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a coarse-to-fine framework for reconstructing 3D coronary arteries from sparse X-ray angiographic views. The method directly predicts a branch-structured representation consisting of B-spline centerlines and vessel radii, using a frozen VGGT backbone and learned branch queries for coarse prediction, followed by projection-guided geometric and radius refinement. Experiments on simulated angiographic projections demonstrate improved centerline accuracy and structural connectivity compared with existing methods. The main contribution is an efficient reconstruction framework that produces explicit vascular geometry without requiring cross-view point correspondence or volumetric post-processing.

Critical strengths
  1. Explicit and anatomically structured vessel representation.

Directly predicting B-spline centerlines and dense radius profiles is a meaningful design choice. Unlike voxel-based methods that require additional centerline extraction, the proposed representation naturally preserves within-branch connectivity and provides geometric information useful for downstream applications. The experimental results demonstrate substantial improvements in structural connectivity. It addresses a practical limitation of existing reconstruction methods and reduces the need for potentially error-prone post-processing.

  1. Efficient reconstruction with effective projection-guided refinement.

The coarse-to-fine framework achieves competitive reconstruction accuracy while requiring only 121 ms per case. The ablation studies show that the B-spline representation and projection-guided refinement substantially improve reconstruction quality. It also demonstrates that the proposed approach can achieve a reasonable balance between accuracy and computational efficiency, making it potentially suitable for time-sensitive applications.

Critical weaknesses
  1. Limited validation on realistic angiographic data.

The quantitative evaluation relies entirely on synthetic projections generated from CT-derived coronary anatomy. These projections do not fully capture the challenges of clinical angiography, particularly cardiac motion between acquisitions, imperfect segmentation, and calibration errors. Although qualitative results on real angiograms are included, they are insufficient to demonstrate reliable generalization. This is a major concern because the method's practical value depends on its robustness to real-world imaging conditions, and the current evidence does not establish whether the reported improvements can be maintained in clinical settings.

  1. Reconstruction accuracy and clinical relevance remain insufficiently demonstrated.

Despite strong connectivity results, the method does not consistently outperform existing approaches in reconstruction accuracy. For example, in the two-view LCA experiments, DeepCA achieves higher Dice and clDice scores (56.56% and 71.20%) than the proposed method (49.45% and 61.30%). Furthermore, although direct radius prediction is a key contribution, the paper does not provide a quantitative evaluation of radius reconstruction accuracy or stenosis-related geometric measurements. This weakness is consequential because improved connectivity alone does not necessarily indicate more accurate vascular reconstruction, and the lack of quantitative radius evaluation makes it difficult to assess the method's advantages for downstream applications such as quantitative coronary analysis.

Questions for authors
  1. Can the authors provide a quantitative evaluation on real angiographic data, or at least investigate robustness to realistic inter-view cardiac motion, segmentation errors, and projection calibration errors? Such experiments would help establish whether the improvements observed on synthetic data transfer to clinical conditions.

  2. Can the authors evaluate centerline and radius reconstruction accuracy separately, particularly for small side branches and regions with vessel narrowing? Since explicit geometric reconstruction is a primary motivation of the proposed method, these measurements would help clarify its advantages over existing volumetric approaches.

  3. The proposed method achieves substantially better connectivity but lower Dice and clDice than DeepCA on two-view LCA reconstruction. Can the authors explain this trade-off and clarify whether the connectivity advantage comes primarily from the imposed parametric representation rather than more accurate recovery of the underlying 3D anatomy?

1 comment

We thank the reviewer for these constructive comments. We have revised the manuscript to address these concerns by explicitly stating the input assumptions, adding robustness experiments and evaluating branch-level geometry and radius accuracy beyond connectivity alone.

Weakness 1 and Question 1: Real data and robustness

Our method was developed for phase-matched, segmented views with known acquisition geometry, assumptions now stated explicitly in Section 3.1, lines 143–147. We agree that clinical angiography can depart from these conditions through inter-view motion, segmentation errors and calibration inaccuracies. We therefore further investigate robustness to direction and segmentation errors (Tables 18–19).

For direction errors, we perturb the supplied direction of one view without changing either mask. Below, CD is branch-level Chamfer distance and RadMAE is radius mean absolute error, both computed on matched branches. Entries show DeepCA / our refined model.

ArteryOffsetDice (%) ↑CD (mm) ↓RadMAE (mm) ↓
RCA0°63.49 / 66.9215.14 / 4.770.742 / 0.149
RCA4°53.93 / 59.1518.83 / 5.290.857 / 0.152
LCA0°56.56 / 49.4519.96 / 8.720.946 / 0.260
LCA4°38.66 / 44.2925.80 / 8.871.089 / 0.261

At 4°, our model degrades less than DeepCA on all three metrics for both arteries. Direction errors mainly reduce our Dice; CD changes modestly and radius errors remain nearly unchanged.

Segmentation errors reduce accuracy for all methods. RCA Dice drops are similar, while ours has the smallest LCA drop (Table 19). Our method nevertheless remains sensitive to segmentation errors.

We have not evaluated cardiac motion: ImageCAS lacks phase-resolved anatomy, so simulation would require additional deformation assumptions. Our evaluation is limited to phase-matched inputs.

We lack real angiography with paired 3D reference anatomy for quantitative evaluation and do not claim established clinical generalisation. We are in the process of obtaining paired data through a collaborating hospital as our next step towards coronary artery reconstruction from real X-ray angiography data.

Weakness 2 and Question 2: Centreline and radius accuracy, side branches and narrowing

Tables 2–3 report CD and RadMAE alongside side-branch recall (BrRec), attachment-position error (AttErr) and relative radius MAE (RadMAE_rel). Definitions and branch matching details appear in Section 3.6 and Appendix A.3.2 (lines 1083–1159). Geometry and radius errors are computed on matched predicted and ground-truth branches; BrRec measures the fraction of ground-truth side branches matched to a prediction branch.

ArteryMethodCD (mm) ↓BrRec (%) ↑RadMAE (mm) ↓RadMAE_rel (%) ↓
RCADeepCA15.1475.000.74233.87
RCAOurs4.7792.050.1498.65
LCADeepCA19.9655.250.94640.04
LCAOurs8.7288.720.26013.05

Tables 12–13 evaluate individual main and side branches, including less frequent entries. We match 73% of RCA side branch 4 versus 33% for DeepCA, and 71% of LCA side branch 5 versus 35%. This addresses side-branch recovery. We additionally stratify radius errors by ground-truth radius (<1.5, 1.5–2.5, and >2.5 mm; Appendix A.3.5, Table 14). Our model has lower absolute and relative radius errors than DeepCA in every category for both arteries. For radii <1.5 mm, our RadMAE is 0.157 versus 0.292 mm for RCA and 0.260 versus 0.348 mm for LCA.

The test set lacks sufficient annotated narrowing cases for systematic stenosis-specific evaluation. We cannot perform this evaluation with the currently available annotations and plan to undertake it once clinical data with stenosis annotations become available.

Weakness 2 and Question 3: Connectivity versus Dice

We agree that our representation promotes connectivity, but this does not rule out better anatomical recovery. Higher branch recall and lower CD and radius errors provide separate evidence of improved branch recovery and geometric accuracy, beyond connectivity alone.

DeepCA has higher two-view LCA Dice and clDice and lower AttErr. Its volumetric representation directly targets vascular occupancy, whereas ours directly models branch geometry and radius profiles. These results highlight different strengths. Our radius accuracy may benefit downstream QCA and potentially geometry-based FFR estimation, although these applications remain to be validated.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.