acceptodds
Under review as a conference paper at ICLR 2027

CoroLift: Fast Single-View 3D Coronary Reconstruction with Anatomical Priors

Abstract

3D coronary reconstruction from 2D observations is important for clinical hemodynamic assessment, yet is inherently ill-posed, particularly in the single-view setting without multi-view geometric cues. Existing methods therefore typically rely on multiple views, leaving single-view reconstruction largely unexplored. Our key idea is to leverage anatomical priors to constrain the missing 3D geometry from a single view. We propose **CoroLift**, a coarse-to-fine framework for single-view 3D coronary reconstruction that integrates anatomical priors on topology, position, and morphology into initialization and refinement. To construct these priors, we decompose coronary trees into slots between bifurcations and endpoints, with shared anatomical identities across patients. We derive a topology prior from population-level connectivity statistics, a position prior from spatial distributions, and a morphology prior that combines family-specific statistical shape models with tree-structured inter-slot dependencies. CoroLift first formulates reconstruction as depth estimation along imaging rays. Topology and position priors guide residual-based 3D endpoint localization and continuous branch-depth estimation, yielding an initial 3D coronary tree. The morphology prior then regularizes the joint vessel geometry while maintaining consistency with the 2D observations. Experiments show that CoroLift reduces geometric reconstruction error by **26.6%** compared with existing baselines, while substantially improving morphological fidelity and reconstructing each 3D coronary tree in only **0.17 s**.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Single-view 3D reconstruction, Anatomical priors, Ill-posed inverse problems, Structured probabilistic modeling

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

This paper proposes CoroLift, a framework that leverages anatomical prior information to achieve rapid single-view 3D reconstruction of coronary arteries. The method integrates topological, positional, and morphological priors to resolve the depth ambiguity inherent in single-view reconstruction.

Critical strengths
  1. The paper addresses an interesting and clinically relevant problem. Using anatomical priors to resolve depth ambiguity in single-view coronary reconstruction is a reasonable and interpretable approach.

  2. The experimental evaluation encompasses comparisons with various baseline methods, ablation studies, and testing across different noise levels, acquisition angles, and scenarios involving incomplete observations. The reported reconstruction speed is also promising.

Critical weaknesses
  1. The evaluation relies entirely on synthetic 2D projections from CCTA data. Although the results are promising, it is difficult to assess the practical value of the reconstruction without real angiographic data or downstream hemodynamic evaluation.

  2. The final refinement improves morphological accuracy but increases CD compared with the intermediate result. It is also unclear whether the final reconstruction still satisfies the 2D reprojection constraint. In addition, the adaptation of existing baselines to the single-view setting needs further clarification.

Questions for authors
  1. Is the reported reconstruction accuracy sufficient for hemodynamic analysis? Could the authors provide vessel-radius errors or a downstream CFD evaluation?
  2. All 2D observations are synthetically generated from CCTA data. Have the authors evaluated CoroLift on real coronary angiograms?
  3. Does the final morphology refinement preserve 2D reprojection consistency? In Table 5, the intermediate result achieves a lower CD than the final reconstruction. Why is the final result preferred?
  4. How were the existing baselines adapted to the single-view setting? Were they sufficiently tuned for a fair comparison?
3 comments

Thank you for the constructive review. Where a question repeats a weakness, we answer both together.

W1 (a) / Q2. Evaluation on real coronary angiograms.

We evaluated CoroLift on the public CARDIAG dataset, which contains 645 real angiograms of 122 patients. The inputs are independent of our pipeline. The 2D traces and labels come from SYNTAX segment masks annotated by the dataset authors, and the C-arm geometry comes from the DICOM headers. No 3D information is used to prepare them. The models are trained on CCTA only and are applied without fine-tuning. Since no 3D ground truth exists, each view is reconstructed alone and scored at the branch junctions of the other views of the same patient. This gives 1527 view pairs with junction landmarks, from 109 patients. Because pairs from one patient are not independent, the intervals come from a patient-cluster bootstrap and the p-values from a Wilcoxon signed-rank test over patient and artery clusters.

Single-view methodLandmark error (mm) [95% CI]CoroLift winspLCARCA60–90° apart
Depth plane8.11 [7.07, 9.17]80%< 10⁻¹⁷8.427.0112.32
Population prior5.24 [4.78, 5.75]58%0.0045.604.537.38
Endpoint-linear depth5.11 [4.61, 5.67]59%< 10⁻³5.294.716.62
CoroLift, Stage I4.78 [4.40, 5.25]52%0.025.004.426.17
CoroLift4.70 [4.37, 5.21]––4.974.346.16
Multi-view oracle3.48 [3.02, 4.07]––3.822.994.11

CoroLift has the lowest landmark error overall, on both arteries, and at view separations of 60–90°. It closes 74% of the gap between a depth plane (8.11 mm) and the multi-view oracle (3.48 mm). Its depth correlates with the triangulated depth at r = 0.79, compared with 0.05 for the depth plane, and it reprojects onto its input view with a median error of 0.25 px.

W1 (b) / Q1. Hemodynamic evaluation: vessel-radius errors and CFD.

We now report the radius error, a reduced-order FFR proxy at 877 outlets, and a 3D CFD check.

Radius (GT median 0.93 mm)MAE (mm)Bias (mm)Rel. MAEr
CoroLift (width read on the vessel axis)0.113+0.01611.9%0.91
Population profile0.1510.00015.5%0.84
FFR proxy vs. CCTA tree (877 outlets, axis-read width)MAEBiasr
CoroLift centerline + CoroLift radius0.056−0.0240.92
CoroLift centerline + GT radius0.010+0.0041.00
GT centerline + CoroLift radius0.055−0.0280.93
CoroLift centerline + population radius0.090+0.0450.72
3D CFD (17 trees, 92 outlets)MAEMedianr
CoroLift vs. CCTA, 3D CFD0.0200.0040.89
CoroLift vs. CCTA, 1D proxy0.0230.0040.89
1D proxy vs. 3D CFD, CCTA0.030–0.97
1D proxy vs. 3D CFD, CoroLift0.024–0.99

Replacing our radius with the ground-truth radius lowers the FFR-proxy error from 0.056 to 0.010, while replacing our centerline with the ground-truth centerline leaves it at 0.055. Under this model the remaining error is therefore dominated by the radius rather than by the reconstructed centerline. The 3D CFD solution agrees, with a difference of 0.020 and a median of 0.004.

W2 (a, b) / Q3. Reprojection consistency of the final result, and why it is preferred over the intermediate result.

The first table reports the reprojection error at each stage on the 82 test patients at 2 px noise.

StageMedian (px)P95 (px)Max (px)> 2 px
x0x^00.0000.0000.00010.00%
PC(x0)P_C(x^0)0.0000.0000.00010.00%
RM(x0)R_M(x^0)0.3582.46419.3787.89%
PC(RM(x0))P_C(R_M(x^0))0.2841.5101.9990.00%
x∗x^* (final)0.3632.39217.4027.74%

The final x∗x^* has a median reprojection error of 0.36 px, which is 0.16 mm, and 92.3% of its samples are within 2 px. The intermediate PC(RM(x0))P_C(R_M(x^0)) keeps all samples within 2 px. We have corrected the wording in the paper accordingly. On real angiograms the median is 0.25 px.

The second table reports accuracy and morphology at each stage.

StageCD ↓HD95 ↓OV ↑Tort. ↓Curv. ↓Bif. angle ↓
x0x^03.8328.8340.1270.2430.59533.32°
PC(x0)P_C(x^0)3.8328.8340.1270.2430.59533.32°
RM(x0)R_M(x^0)3.9739.3380.1050.0410.05524.33°
PC(RM(x0))P_C(R_M(x^0))3.6968.7510.1420.1080.29030.50°
x∗x^* (final)3.8549.1900.1100.0420.05424.19°

Compared with the intermediate result, the final step changes CD by 0.16 mm and reduces the curvature error by 5.4 times, the tortuosity error by 2.6 times, and the bifurcation-angle error by 6.3°. We use x∗x^* by default because hemodynamic analysis needs smooth vessels. PC(RM(x0))P_C(R_M(x^0)) remains available when strict 2 px consistency is required.

W2 (c) / Q4. Adaptation and tuning of the baselines.

The baselines of Table 1 receive the same noisy 2D traces, labels, vessel widths and camera as CoroLift. They are trained on the same 671 trees and scored by the same pipeline. NeCA and 3DGR-CAR take a rendered vessel image, following their official protocol.

BaselineSingle-view adaptationTuning
NN Retrievalretrieve training tree with best-matching projectionvalidation split
Shape ModelPCA model fitted by reprojection errorvalidation split
Implicit Fieldimplicit field conditioned on labeled trace + test-time reprojection refinementvalidation split
Soft Templateour topology + position priors, no morphologyvalidation split
Tree Regressiontree-recurrent 3D regression along the topologyvalidation split
NeCA, 3DGR-CARofficial code, one viewthreshold best on test set

The results on the 82 test patients at 2 px noise are listed below. CD, HD95 and Gap are in mm, and T, P and M denote the topology, position and morphology priors.

MethodPriorsCD ↓HD95 ↓OV ↑Tort. ↓Curv. ↓Bif. angle ↓Gap ↓Time
NeCA (2024)–11.1852.040.051––––6.5 min
3DGR-CAR (2024)–12.6634.690.052––––18 min
NN RetrievalT, M7.88417.5560.0170.0370.03626.71°16.621.27 s
Shape ModelT, M6.42714.6340.0270.0550.05326.68°22.575.43 s
Implicit FieldP6.18713.7830.0330.0770.13035.19°12.845.72 s
Soft TemplateT, P6.14515.3180.0740.2530.61134.65°24.330.11 s
Tree RegressionT, P5.25111.1720.0350.1570.26148.64°6.130.10 s
CoroLiftT, P, M3.8549.1900.1100.0420.05424.19°0.000.17 s

Where a choice existed, it favors the baselines. NeCA and 3DGR-CAR use the threshold that is best on the test set, and a Shape Model that is given our slot correspondence improves from 6.43 to 5.21 mm. Among the methods of Table 1, CoroLift is still the only one that returns a connected tree.

Thank you for the additional experiments and clarifications, particularly the real-angiogram evaluation and hemodynamic analysis, which address most of my concerns. The 3D CFD validation remains limited in scope, and the hemodynamic benefit of the final refinement over the intermediate reconstruction has not yet been demonstrated. I will maintain my original rating.

Reviewer_EzJdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CoroLift reconstructs a 3D coronary tree from a single calibrated view containing anatomically labeled 2D centerline traces. It builds population priors for vessel topology, position, and morphology from CCTA-derived trees. An endpoint-localization model and continuous depth predictor provide a ray-consistent initialization; morphology regularization and a projection-consistency operator then refine the reconstruction. On synthetic projections of 82 held-out ImageCAS patients, CoroLift reports a median Chamfer distance of 3.854 mm at 2-pixel trace noise, compared with 5.251 mm for the strongest listed baseline, with approximately 0.17-second inference time. The paper also reports component ablations and tests across noise levels, acquisition views, vessel families, and incomplete observations.

Critical strengths
  1. The method addresses the substantial depth ambiguity of single-view reconstruction with an explicit, structured anatomical prior. Table 1 shows a meaningful geometric improvement over the evaluated single-view baselines, while the patient-level split keeps prior estimation and model training separate from the 82 test patients.
  2. The ablations make the roles and trade-offs of the components relatively clear. Endpoint localization and continuous depth prediction jointly improve initialization (Table 3); morphology regularization greatly reduces tortuosity and curvature errors, while projection correction recovers geometric accuracy at an intermediate stage (Tables 2 and 5). These results are more informative than a single final-score comparison.
Critical weaknesses
  1. Clinical generalization remains uncertain. The evaluation projects CCTA-derived 3D trees into synthetic angiographic observations and supplies correctly labeled 2D slots, shared bifurcation topology, vessel widths, and calibrated camera geometry. The simulated trace noise preserves that topology, and the incomplete-observation experiment removes complete distal branches. Real angiograms can additionally contain ambiguous or incorrect branch labels, missed junctions, occlusions, overlapping vessels, and calibration error. The reported results therefore establish performance under a controlled reconstruction protocol, but do not yet establish performance from routinely obtained clinical observations. The absolute median Rotterdam overlap of 0.110 at 2-pixel noise also warrants interpretation for the proposed hemodynamic use.
  2. The claimed consistency of the final reconstruction with the observed 2D trace is not demonstrated. The output is x∗=RM(PC(RM(x0)))x^\ast=R_M(P_C(R_M(x^0))): a morphology projection follows the observation-consistency step and may move points away from their observed image locations. Table 14 reports reprojection errors only through the intermediate PC(RM(x0))P_C(R_M(x^0)), omitting x∗x^\ast. Table 5 also shows that this intermediate has better Chamfer distance and overlap than the final output, although the final output improves morphology. Separately, the apparent 4-pixel-noise results differ between Table 1 (CD 3.847 mm, OV 0.089) and Table 12(a) (CD 3.999 mm, OV 0.076). These points need clarification before the joint geometry-and-observation claim can be fully assessed.
Questions for authors
  1. Why do Tables 1 and 12(a) report different results for the same nominal final reconstruction at 4-pixel noise?
  2. Is there evaluation on real angiograms with independently obtained 2D traces and anatomical labels?
  3. Please report the median, 95th percentile, maximum, and fraction exceeding the 2-pixel reprojection tolerance for the final x∗x^\ast, as in Table 14.
3 comments

Thank you for the careful review. Where a question repeats a weakness, we answer both together.

W1 (a) / Q2. Clinical generalization, and evaluation on real angiograms with independently obtained 2D traces and labels.

We evaluated CoroLift on the public CARDIAG dataset, which contains 645 real angiograms of 122 patients. The inputs are independent of our pipeline. The 2D traces and labels come from SYNTAX segment masks annotated by the dataset authors, and the C-arm geometry comes from the DICOM headers. No 3D information is used to prepare them. The models are trained on CCTA only and are applied without fine-tuning. Since no 3D ground truth exists, each view is reconstructed alone and scored at the branch junctions of the other views of the same patient. This gives 1527 view pairs with junction landmarks, from 109 patients. Because pairs from one patient are not independent, the intervals come from a patient-cluster bootstrap and the p-values from a Wilcoxon signed-rank test over patient and artery clusters.

Single-view methodLandmark error (mm) [95% CI]CoroLift winspLCARCA60–90° apart
Depth plane8.11 [7.07, 9.17]80%< 10⁻¹⁷8.427.0112.32
Population prior5.24 [4.78, 5.75]58%0.0045.604.537.38
Endpoint-linear depth5.11 [4.61, 5.67]59%< 10⁻³5.294.716.62
CoroLift, Stage I4.78 [4.40, 5.25]52%0.025.004.426.17
CoroLift4.70 [4.37, 5.21]––4.974.346.16
Multi-view oracle3.48 [3.02, 4.07]––3.822.994.11

CoroLift has the lowest landmark error overall, on both arteries, and at view separations of 60–90°. It closes 74% of the gap between a depth plane (8.11 mm) and the multi-view oracle (3.48 mm). Its depth correlates with the triangulated depth at r = 0.79, compared with 0.05 for the depth plane, and it reprojects onto its input view with a median error of 0.25 px.

W1 (b). Wrong labels, missed junctions, overlaps and calibration error.

We injected each type of error into the observations of all 82 test patients, while the solver kept the nominal camera and labels.

Perturbation (82 patients)CD (mm)OVΔCD, median [IQR]
None3.8540.110–
Label swap, all cases3.9380.1160.00 [0.00, 0.00]
Label swap, 26 affected cases4.0250.116+0.15 [−0.10, +0.46]
Missed junction, 10%3.8810.107+0.00 [−0.02, +0.02]
Missed junction, 20%3.8840.107+0.00 [−0.02, +0.02]
Overlap bridged, 10%3.8670.109+0.01 [−0.00, +0.02]
Overlap bridged, 20%3.8940.101+0.02 [+0.00, +0.06]
Pixel spacing +2%3.8490.101+0.04 [−0.04, +0.09]
C-arm angles, σ = 2°3.9850.094+0.12 [+0.04, +0.25]

For annotation errors and mild calibration errors, the median paired change in CD is at most 0.15 mm. For a larger angle error of 5°, the CD is 4.45 mm. Every perturbed run remains better than the best baseline of the original submission (5.25 mm).

W1 (c). OV = 0.110 and the hemodynamic use.

OV counts a point as correct only if it lies within one vessel radius of the ground truth, and the median radius is 0.93 mm. We therefore tested the hemodynamic effect directly.

FFR proxy vs. CCTA tree (877 outlets, axis-read width)MAEBiasr
CoroLift centerline + CoroLift radius0.056−0.0240.92
CoroLift centerline + GT radius0.010+0.0041.00
GT centerline + CoroLift radius0.055−0.0280.93
CoroLift centerline + population radius0.090+0.0450.72
3D CFD (17 trees, 92 outlets)MAEMedianr
CoroLift vs. CCTA, 3D CFD0.0200.0040.89
CoroLift vs. CCTA, 1D proxy0.0230.0040.89
1D proxy vs. 3D CFD, CCTA0.030–0.97
1D proxy vs. 3D CFD, CoroLift0.024–0.99

Replacing our radius with the ground-truth radius lowers the FFR-proxy error from 0.056 to 0.010, while replacing our centerline with the ground-truth centerline leaves it at 0.055. Under this model the remaining error is therefore dominated by the radius rather than by the reconstructed centerline. The 3D CFD solution agrees, with a difference of 0.020 and a median of 0.004.

W2 (a) / Q3. Reprojection statistics of the final x∗x^*.

The table reports all stages on the 82 test patients at 2 px noise, pooled over all samples.

StageMedian (px)P95 (px)Max (px)> 2 px
x0x^00.0000.0000.00010.00%
PC(x0)P_C(x^0)0.0000.0000.00010.00%
RM(x0)R_M(x^0)0.3582.46419.3787.89%
PC(RM(x0))P_C(R_M(x^0))0.2841.5101.9990.00%
x∗x^* (final)0.3632.39217.4027.74%

The final x∗x^* has a median reprojection error of 0.36 px, which is 0.16 mm, and 92.3% of its samples are within 2 px. The intermediate PC(RM(x0))P_C(R_M(x^0)) keeps all samples within 2 px. We have corrected the wording in the paper accordingly. On the 645 real angiograms the median is 0.25 px.

W2 (b). The intermediate PC(RM(x0))P_C(R_M(x^0)) has better CD and OV than the final output.

StageCD ↓HD95 ↓OV ↑Tort. ↓Curv. ↓Bif. angle ↓
x0x^03.8328.8340.1270.2430.59533.32°
PC(x0)P_C(x^0)3.8328.8340.1270.2430.59533.32°
RM(x0)R_M(x^0)3.9739.3380.1050.0410.05524.33°
PC(RM(x0))P_C(R_M(x^0))3.6968.7510.1420.1080.29030.50°
x∗x^* (final)3.8549.1900.1100.0420.05424.19°

Compared with the intermediate result, the final step changes CD by 0.16 mm and reduces the curvature error by 5.4 times, the tortuosity error by 2.6 times, and the bifurcation-angle error by 6.3°. We use x∗x^* by default because hemodynamic analysis needs smooth vessels. PC(RM(x0))P_C(R_M(x^0)) remains available when strict 2 px consistency is required.

W2 (c) / Q1. "Why do Tables 1 and 12(a) report different results for the same nominal final reconstruction at 4-pixel noise?"

The two tables used different random draws of the 4 px trace noise. For consistency, both now use the same draw, which gives CD 3.999, HD95 9.253 and OV 0.076 at 4 px. The 2 px row is unchanged, and the ranking of the methods at 4 px is not affected, since the next best CD is 5.198 mm. These table numbers refer to the original submission. In the revised PDF the appendix tables shift by three, so Table 12(a) becomes Table 15(a). The values at all noise levels are as follows.

Noise (px)CD ↓HD95 ↓OV ↑Tort. ↓Curv. ↓Bif. angle ↓
03.8529.4730.1360.0390.05015.91°
13.8759.2600.1340.0370.05220.35°
23.8549.1900.1100.0420.05424.19°
33.9169.2200.0860.0460.05827.99°
43.9999.2530.0760.0480.06130.88°

From 0 to 4 px of trace noise, CD increases by 0.15 mm and HD95 stays between 9.19 and 9.47 mm.

Thank you for the additional experiments and clarifications. The CARDIAG results provide useful evidence of cross-view consistency at branch junctions, while the 3D accuracy of the full coronary tree on real angiograms remains uncertain. The final-stage reprojection results also clarify the trade-off between vessel smoothness and adherence to the observed trace. Given these remaining concerns, I am keeping my original rating.

Reviewer_p9vMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper starts with an idea that leverages anatomical priors to constrain the missing 3D geometry from a single view. In this regard, the authors propose a coarse-to-fine framework for single-view 3D coronary reconstruction that integrates anatomical priors on topology, position, and morphology into initialization and refinement. The experiments show the advantage of the proposed method.

Critical strengths
  1. Clear description of the proposed method with enough detail to follow.
  2. The idea of leveraging anatomical priors to constrain the missing 3D geometry from a single view is reasonable.
Critical weaknesses
  1. The key idea is not new. How to leverage prior knowledge to reconstruct the missing 3D geometry from a single view can be found in existing work. The key difference from existing works is not significant.

  2. The way to construct the prior is not new. Also, the reason for using only these three priors is missing.

  3. Lack of comparison with recent methods.

Questions for authors
  1. The key idea is not new. How to leverage prior knowledge to reconstruct the missing 3D geometry from a single view can be found in existing work. The key difference from existing works is not significant.

  2. The way to construct the prior is not new. Also, the reason for using only these three priors is missing.

  3. Lack of comparison with recent methods.

3 comments

Thank you for the review. Q1–Q3 repeat W1–W3, so we answer each of them once.

W1 / Q1. "The key idea is not new. How to leverage prior knowledge to reconstruct the missing 3D geometry from a single view can be found in existing work. The key difference from existing works is not significant."

Our contribution is the way the single-view problem is parameterized and the way the priors enter a connected tree.

Existing methods regress free 3D coordinates or fit a population template. CoroLift instead uses the calibrated projection. Each observed 2D point defines a 3D ray, so only one unknown remains for that point, which is its depth. Branches that meet at a junction share one anatomical node, so the tree is connected by construction. The initialization reproduces the observed projection exactly, and the later morphology refinement gives up a small, measured part of it (median 0.36 px) for a plausible vessel shape.

3D coordinate regression (Tree Regression)CoroLift Stage I (ray depth)
Unknowns per point31
Observed 2D positionpredicted againkept exactly at initialization (0.000 px)
Branches at a junctionpredicted separatelyone shared node
Connected treenot guaranteed (gap 6.13 mm)by construction (gap 0 mm)
Topology and position priors act onall 3 coordinatesthe unobserved depth

The following comparison uses the same topology and position priors, the same training data and the same evaluation. The morphology prior is not used.

Same priors (topology + position)FormulationCD ↓HD95 ↓OV ↑Gap ↓
Soft Templatepopulation template6.14515.3180.07424.33
Tree Regression3D coordinate regression5.25111.1720.0356.13
CoroLift, Stage Idepth along observed rays at shared nodes3.8328.8340.1270.00

With the same priors, CD drops from 6.15 and 5.25 mm to 3.83 mm, and the junction gap drops from 24.3 and 6.1 mm to 0. This gain is obtained with matched priors, by using the projection geometry and shared nodes.

The table below lists each component, the conventional choice it replaces, and its measured effect.

StepComponentRoleConventional counterpartEffect
143 anatomical slotscorrespondence across patientsnearest-neighbor correspondenceCD 3.959 → 3.854
2Shared nodes at slot endsone node per junctionindependent endpointsGap 8.9 → 0.0 mm
3Ray depth between nodesonly the unobserved depth is estimatedpopulation initializationCD 8.687 → 3.832
4Residual shape anchored at the nodesshape recovered with nodes fixedcoordinate PCACD 4.792 → 3.931
5Gaussian–Markov tree on the slot graphdependency between connected branchesindependent GaussiansCurv. 0.102 → 0.055

The resulting performance is shown below. NeCA and 3DGR-CAR are designed for two or more views, and we ran their official code with one and with two views.

NeCA, 1 view3DGR-CAR, 1 viewNeCA, 2 views3DGR-CAR, 2 viewsTree Regression, 1 viewCoroLift, 1 view
CD (mm) ↓11.1812.663.905.775.253.85
HD95 (mm) ↓52.0434.6915.3921.1011.179.19
Gap (mm) ↓no treeno treeno treeno tree6.130.00
Time ↓6.5 min18 min6.5 min18 min0.10 s0.17 s

With one view, the CD of CoroLift is about one third of that of NeCA and 3DGR-CAR, and CoroLift is more than 2000 times faster. It also reaches the CD of NeCA with two views (3.85 vs. 3.90 mm) at a 40% lower HD95. Compared with Tree Regression, the best single-view baseline, CD is 26.6% lower and HD95 is 17.7% lower, and CoroLift is the only one of the tree-reconstruction methods in Table 1 that returns a connected tree.

W2 / Q2. "The way to construct the prior is not new. Also, the reason for using only these three priors is missing."

In our centerline-tree representation, three objects specify the reconstruction completely. These are the graph p(s)p(s), which says which slots connect, the node positions YjY_j, and the edge shapes rsr_s. Every slot is recovered from them as Xs=ℓs+rsX_s = \ell_s + r_s, where ℓs\ell_s is the straight chord between the two nodes of the slot and rsr_s is its deviation from that chord. Nothing else is needed to rebuild the tree.

This split is established. Statistical tree-shape analysis separates topology from branch geometry (Feragen et al., TPAMI 2013; Duncan et al., AoAS 2018), shape theory separates position from shape (Kendall, 1984; Dryden & Mardia, 1998), and the named-segment graph is the clinical standard for coronary anatomy (Austen et al., Circulation 1975).

A single view leaves one part of each object unknown, and each prior supplies that part:

Tree factorObjectOne view givesOne view lacksPrior
Graphparent map p(s)p(s)2D tracesbranch identity, 3D connectivityTopology
Nodes3D keypoints YjY_jray (2 of 3 coordinates)depth along the rayPosition
Edgesresidual curves rsr_s2D projectionout-of-plane shapeMorphology

In this representation the three factors determine the centerline completely, so each prior has its own role, and the ablation below shows that each one contributes. Radius needs no prior, because it follows from the 2D width and the reconstructed depth. With the width read on the vessel axis, the error is 0.113 mm.

Each prior improves the quantity it is responsible for (82 test patients, 2 px noise):

PriorMetricWithoutWithImprovement
TopologyCD (mm) ↓3.9593.8542.7%
HD95 (mm) ↓9.3689.1901.9%
OV ↑0.1050.1104.8%
PositionCD (mm) ↓6.2713.85438.5%
HD95 (mm) ↓14.7699.19037.8%
OV ↑0.0590.11086.4%
MorphologyTort. error ↓0.2430.04282.7%
Curv. error ↓0.5950.05490.9%
Bif. angle error ↓33.33°24.19°27.4%

Every metric improves with its prior (improvement = relative gain in the direction of the arrow).

What is new in the construction is the set of 43 anatomical slots. They give point-level correspondence across 671 patients, which allows the three factors to be estimated separately and then combined under one calibrated projection.

W3 / Q3. "Lack of comparison with recent methods."

We added NeCA (2024) and 3DGR-CAR (2024), the two most recent coronary reconstruction methods with public code. Both were run with their official code on the 82 test patients, with the same view, trace noise and evaluation as CoroLift. 3DGR-CAR is initialized by filtered back-projection because the weights of its center predictor are not released. The two-view rows use standard clinical view pairs and serve as a reference, since they are not paired with the single test view.

MethodViewsThresholdCD (mm) [IQR]HD95 (mm)Time
NeCA1default35.76 [34.02, 37.88]103.096.5 min
NeCA1test-best11.18 [10.52, 11.99]52.046.5 min
3DGR-CAR1default17.06 [15.17, 20.79]51.3618 min
3DGR-CAR1test-best12.66 [11.48, 14.83]34.6918 min
NeCA2default8.51 [7.17, 9.83]45.346.5 min
NeCA2test-best3.90 [3.22, 4.44]15.396.5 min
3DGR-CAR2default8.19 [6.96, 9.95]42.4418 min
3DGR-CAR2test-best5.77 [4.55, 7.30]21.1018 min
CoroLift1none3.859.190.17 s

With one view, the CD of CoroLift is three times lower than that of both methods, even when their threshold is chosen on the test set. With one view CoroLift also reaches the CD of NeCA with two views and is 1.9 mm better than 3DGR-CAR with two views. It needs no threshold and runs in 0.17 s instead of 6.5 to 18 minutes.

We also added an evaluation on real angiograms. We evaluated CoroLift on the public CARDIAG dataset, which contains 645 real angiograms of 122 patients. The inputs are independent of our pipeline. The 2D traces and labels come from SYNTAX segment masks annotated by the dataset authors, and the C-arm geometry comes from the DICOM headers. No 3D information is used to prepare them. The models are trained on CCTA only and are applied without fine-tuning. Since no 3D ground truth exists, each view is reconstructed alone and scored at the branch junctions of the other views of the same patient. This gives 1527 view pairs with junction landmarks, from 109 patients. Because pairs from one patient are not independent, the intervals come from a patient-cluster bootstrap and the p-values from a Wilcoxon signed-rank test over patient and artery clusters.

Single-view methodLandmark error (mm) [95% CI]CoroLift winspLCARCA60–90° apart
Depth plane8.11 [7.07, 9.17]80%< 10⁻¹⁷8.427.0112.32
Population prior5.24 [4.78, 5.75]58%0.0045.604.537.38
Endpoint-linear depth5.11 [4.61, 5.67]59%< 10⁻³5.294.716.62
CoroLift, Stage I4.78 [4.40, 5.25]52%0.025.004.426.17
CoroLift4.70 [4.37, 5.21]––4.974.346.16
Multi-view oracle3.48 [3.02, 4.07]––3.822.994.11

CoroLift has the lowest landmark error overall, on both arteries, and at view separations of 60–90°. It closes 74% of the gap between a depth plane (8.11 mm) and the multi-view oracle (3.48 mm). Its depth correlates with the triangulated depth at r = 0.79, compared with 0.05 for the depth plane, and it reprojects onto its input view with a median error of 0.25 px.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.