Beyond Segmentation: Enriching ImageCAS with Coronary Ostia Landmarks, Rooted Centerline Graphs, and Open-Weight Localization Models
Authors: Matteo Leccardi, Andrea Moglia, Marco Morandini, Michele Brembilla, Alessandro Nicosia, Andrea Penserini, Christian Brandini, Ender Konukoglu, Marco Marcon, Luca Mainardi, Pietro Cerveri
Organizations: Department of Electronics, Information and Bioengineering, Politecnico di Milano, Via Ponzio 34/5, 20133 Milano, Italy · Computer Vision Laboratory, Department of Information Technology and Electrical Engineering, ETH Zurich, Sternwartstrasse 7, 8092 Zurich, Switzerland · Department of Industrial and Information Engineering, University of Pavia, Via A. Ferrata 5, 27100 Pavia, Italy
Coronary ostia and centerlines are key reference points for automated analysis of coronary computed tomography angiography (CCTA), from surgical planning to intraoperative vessel tracking. However, the only public ostium annotations cover 77 non-gated aortic scans, not CCTA, while CAT08 and ASOCA (32 and 40 CCTA scans) mark the ostia only implicitly, as centerline start points. This work enriches ImageCAS, the largest public dataset for coronary artery segmentation, with ostium annotations for its 1000 patients and centerline graphs rooted at these new ostia, making it, to the authors' knowledge, the largest public dataset for ostium localization and the only one with centerlines following the official ImageCAS masks and starting at annotated ostia. Three annotators marked each ostium in the first 500 volumes; in the next 500, the candidates of a model trained on the first batch were corrected (human-in-the-loop). Six 3D localization models, trained on ImageCAS, CAT08 and ASOCA, were benchmarked on 58 held-out volumes (50 ImageCAS, 4 CAT08, 4 ASOCA). SwinUNETR-V2 and nnResUNet localized all ostia without failure (>5 mm error), with median errors of 0.70-0.87 mm. Trained on CAT08 and ASOCA alone, SwinUNETR-V2 fails on 34% of right and 53% of left ostia, whereas adding 250 ImageCAS volumes removes all failures. The centerlines were extracted by a deterministic, quality-controlled minimal-path procedure; on ASOCA, they cover a median of 99.5% of the reference length, with a median per-patient mean distance of 0.22 mm between matched reference points and centerlines. Annotations, centerlines, code, and weights are available at https://hal9000-lab.github.io/.
Figures & tables
Figure 1: Overview of this work: the public CAT08, ASOCA, and ImageCAS datasets and their reconciled coronary ostium definitions; the two-stage annotation protocol used to extend ImageCAS with left/right ostium landmarks and centerline graphs; the shared benchmark of localization models trained on the combined dataset; and the resulting public release.
Figure 2: Centerline extraction pipeline. (1) Inputs: the coronary segmentation mask and the annotated left and right ostia (orange dots), resampled to a 0.5 mm isotropic grid and cropped to the arteries. (2) Centered cost: the cost of Eq. ( 1 ) decreases steeply with the distance DT from the vessel wall, so that minimal paths are both short and central; each ostium is bridged to the mask. (3) Ostium fronts: a minimal-cost front propagated from each ostium gives the geodesic distance G and the minimal path (predecessor map) of every voxel, and splits the mask between the left (L) and right (R) trees according to the ostium that reaches each voxel; mask gaps of up to 10 mm are bridged. (4) Tree growth: the uncovered voxel farthest from the ostium (star) is traced back to the tree along its minimal path, which is kept as a new branch only if it adds at least 5 mm outside the region already covered (cross: rejected path), so that the result is a tree by construction. The growth is then repeated in a second pass with the wall distance capped at 1.2 times the local vessel radius (bottom arrow), which keeps the paths on the vessel axis at bulky bifurcations. (5) Centerline graph: the tips are trimmed by 2 mm and the trees smoothed branch by branch and resampled every 0.5 mm, yielding one tree per side, rooted at its ostium (square), with endpoints (circles), bifurcations, and per-node position, lumen radius, topological type and side, stored as one graph per patient. Full details are given in Appendix A.3 .
Anatomical
Interoperator
Failure
Right ostium
1.08
1.85
5.0
Left ostium
1.45
2.59
5.0
Table 1: Distance thresholds (mm) delimiting the accuracy buckets.
Model
Developer
Architecture
Parameters
SwinUNETR-V2
NVIDIA
Swin ViT + Conv.
18.35M
ResNet50
Microsoft
Conv.
49.70M
ResNet101
Microsoft
Conv.
92.28M
nnResUNet
ACVL
Conv.
32.69M
MedNeXt-S
MIC-DKFZ
ConvNeXt
5.55M
MedNeXt-B
MIC-DKFZ
ConvNeXt
10.53M
Table 2: Specifications of the models evaluated on the test dataset.
Annotator
Left ostium
Right ostium
Combined (Left+Right)
Mean ± SD
Median [IQR]
Mean ± SD
Median [IQR]
Mean ± SD
Median [IQR]
Annotator 1
2.89 ± 0.95
2.73 [2.35–3.40]
1.98 ± 0.78
1.90 [1.47–2.40]
2.44 ± 0.98
2.37 [1.79–2.92]
Annotator 2
2.06 ± 0.95
1.99 [1.39–2.48]
1.78 ± 0.95
1.63 [1.19–2.22]
1.92 ± 0.96
1.75 [1.25–2.37]
Annotator 3
1.73 ± 0.88
1.54 [1.10–2.22]
1.29 ± 0.71
1.19 [0.78–1.61]
1.51 ± 0.83
1.34 [0.94–1.95]
Annotator 4
1.39 ± 0.90
1.19 [0.75–1.80]
1.13 ± 0.61
1.04 [0.71–1.42]
1.26 ± 0.78
1.09 [0.72–1.57]
Annotator 5
1.44 ± 0.83
1.29 [0.81–1.94]
1.19 ± 0.67
1.10 [0.69–1.55]
1.31 ± 0.76
1.20 [0.76–1.69]
Table 3: Per-annotator localization error (Euclidean distance, in mm, between each annotator’s manually placed point and the mean of the points placed on the same image by the other two annotators, i.e. leave-one-out) for the left and right coronary ostia, individually and combined, together with the pooled statistics across all six annotators. For each group we report the mean ± SD and the median with interquartile range (25th–75th percentile). The statistics are computed over 3000 placements by the six annotators (1500 landmarks per side, 250 per annotator per ostium, three annotators per volume).
Metric
Value
Images requiring manual refinement a
88 / 500 (17.6%)
Images with an undetected ostium (outlier)
10 / 500 (2.0%)
Right ostium
8
Left ostium
2
Both ostia
0
Annotation time, first batch b
≈ 24 h (144 person-h)
Table 4: Semi-automated correction statistics for the second batch of 500 ImageCAS images (patients 501–1000), annotated using SwinUNETR-V2 predictions followed by single-annotator visual review.
Right ostium
Left ostium
Both
Model
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
SwinUNETR-V2
0.53
0.74
0.83
1.02
2.07
0.61
0.85
0.98
1.32
2.51
0.68
0.86
0.91
1.10
1.94
55.6
(O: 77.59 | G: 20.69 | A: 1.72 | F: 0.00)
(O: 84.48 | G: 15.52 | A: 0.00 | F: 0.00)
(O: 68.97 | G: 29.31 | A: 1.72 | F: 0.00)
ASOCA
0.53
0.95
0.95
1.37
1.45
0.54
0.60
0.58
0.64
0.71
0.58
0.81
0.77
0.99
1.02
(O: 50.00 | G: 50.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
(O: 50.00 | G: 50.00 | A: 0.00 | F: 0.00)
CAT08
0.60
1.02
1.04
1.46
1.75
0.59
1.04
1.20
1.64
2.25
0.98
1.05
1.12
1.20
1.46
Table 5: Euclidean distance dc (mm) of the benchmark models trained on the complete ImageCAS dataset (the 950 CCTAs not in the test set) and tested on 58 CCTAs, for the right ostium, the left ostium, and both (per-patient average of the right and left distances). The line below each row gives the percentage of cases in each accuracy bucket (for both, at patient level): O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure 3: Comparison of inter-annotator variability (distance of each annotator’s marker from the mean of the other two annotators of the same image, as in Table 3 , on the 50 ImageCAS test volumes) against the benchmark models’ localization error on the test set (Table 5 ), both ostia combined.
Both trees
Left tree
Right tree
Recall (%)
99.5 [98.9–99.8]
99.4 [98.7–99.8]
99.7 [99.0–99.9]
Precision (%)
98.0 [96.5–99.6]
98.1 [95.7–99.8]
99.4 [97.7–100]
OV (%)
98.6 [98.2–99.0]
98.5 [97.6–99.1]
99.1 [98.4–99.5]
OT (%)
99.1 [98.8–99.9]
99.0 [98.4–100]
100 [99.0–100]
OF (%)
99.1 [96.8–99.5]
99.0 [96.3–99.6]
99.5 [98.6–99.9]
AI (mm)
0.22 [0.21–0.24]
0.22 [0.21–0.25]
0.22 [0.20–0.24]
Table 6: Agreement between the centerlines extracted from the 40 annotated ASOCA masks and the ASOCA reference centerlines, for both trees together and for each tree, as median [interquartile range] over the patients. Recall: fraction of the reference length matched (coverage); precision: fraction of the extracted length matched; OV, OT, OF: overlap, overlap of vessels at least 1.5 mm wide, and overlap until first error (first 40 mm of every path ignored); AI: average inside, the mean distance of the matched reference points from the extracted centerline; HD95: 95th-percentile symmetric Hausdorff distance; length ratio: extracted over reference length; radius error: extracted minus reference radius on matched points (signed mean) and its mean absolute value; endpoints: one-to-one pairing within 10 mm; bifurcations: fraction of reference bifurcations paired within 5 mm.
Both trees
Left tree
Right tree
ImageCAS-X trees: connectivity
Trees with detached segments (%)
15.6
28.6
2.5
Detached length, in those trees (%)
47.2 [36.1–58.5]
47.7 [38.2–59.2]
24.9 [21.1–36.9]
Detached length, all trees (%)
9.7
14.6
1.0
Detached length bridged by the mask (%)
97.7
97.6
100
ImageCAS-X start points against our ostia
Table 7: Our centerlines compared with those released with ImageCAS-X, on the 800 scans the two datasets share. Median with its interquartile range below, or percentage of the trees or of the pooled length. Detached length : part of a tree that cannot be reached from the start point when segments closer than 1 mm are joined; bridged by the mask : share of it that the ImageCAS-X mask connects (a gap of one voxel allowed). Start point : ImageCAS-X start point against our annotated ostium (interoperator threshold: Table 1 ). Inside the other masks : share of the centerline length lying within one voxel of the other set’s masks. Tree size : our extraction on the official and on the ImageCAS-X masks of the same patients (all differences significant, paired Wilcoxon signed-rank test, p<0.01 ).
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Spacing (mm)
Size (with labels)
Resolution
N° of slices
ImageCAS [ 21 ]
0.25 - 0.45
1000
512 × 512
206 - 275
CAT08 [ 28 ]
0.262 - 0.408
32
512 × 512
229 - 351
ASOCA [ 18 ]
0.625
40
512 × 512
200
Appendix
Table A.1: Comparison of the raw image characteristics of the CAT08, ASOCA and ImageCAS datasets. Spacing: in-plane pixel spacing for ImageCAS and CAT08; for ASOCA, the slice spacing, which is its largest spacing (in-plane 0.3–0.4 mm).
Figure A.1: Distance from the vessel wall, D , on which the cost of the minimal paths is built, for one ImageCAS patient: maximum projections along the superior–inferior (top) and anterior–posterior (bottom) axes, in millimeters from the corner of the volume cropped around the arteries. D peaks along the axis of every vessel, which the minimal paths therefore follow; it is largest in the proximal vessels and at their junctions, which motivates the cap applied in the second growth pass.
Figure A.2: Flattening of the coronary trees of one ImageCAS patient (343), seen from superior (anterior up, patient’s right on the right). (a) Centerline graph projected on the axial plane: vessels running toward the apex are foreshortened and branches overlap. (b) Epicardial unfolding: the trees are mapped from a spheroid fitted around the heart onto a plane centered on the aortic root; vessels no longer overlap, but distances across the meridians are stretched toward the apex (bottom of the right tree). (c) Length-preserving flattening: each tree is rebuilt from its ostium with the directions of (b) and the true length of every segment. Black lines: centerline graph; green dots: ostia (right with a red, left with a blue outline); red dots: endpoints. The vessels (gray) are drawn as tubes of the lumen radius stored in the graph, with a slightly irregular outline added for display, not from the segmentation masks. Scale bar: 20 mm. Flattened views of this kind were also used in the visual review of the released centerlines, to assess the quality of the extraction and the defects of the segmentation masks. See Appendix A.5 .
Annotator
Left ostium
Right ostium
Combined (Left+Right)
Mean ± SD
Median [IQR]
Mean ± SD
Median [IQR]
Mean ± SD
Median [IQR]
Annotator 1
1.93 ± 0.64
1.82 [1.56–2.28]
1.32 ± 0.52
1.27 [0.98–1.60]
1.63 ± 0.66
1.58 [1.19–1.95]
Annotator 2
1.37 ± 0.64
1.33 [0.93–1.65]
1.19 ± 0.63
1.09 [0.79–1.49]
1.28 ± 0.64
1.16 [0.83–1.58]
Annotator 3
1.15 ± 0.59
1.03 [0.73–1.47]
0.86 ± 0.47
0.79 [0.52–1.08]
1.00 ± 0.55
0.90 [0.63–1.30]
Annotator 4
0.93 ± 0.60
0.79 [0.50–1.20]
0.75 ± 0.41
0.70 [0.48–0.95]
0.84 ± 0.52
0.73 [0.48–1.05]
Annotator 5
0.96 ± 0.56
0.86 [0.54–1.29]
0.79 ± 0.45
0.74 [0.46–1.04]
0.88 ± 0.51
0.80 [0.50–1.13]
Appendix
Table A.2: Per-annotator distance (Euclidean distance, in mm) between each annotator’s manually placed point and the released consensus of that image, i.e. the mean of the points of all its annotators, the annotator’s own point included, for the left and right coronary ostia, individually and combined, together with the pooled statistics across all six annotators, computed over the same 3000 placements (1500 landmarks per side, 250 per annotator per ostium, three annotators per volume, two ostia per volume - left and right). Same layout and statistics as Table 3 , which reports the leave-one-out distances; on an image annotated by three annotators each distance here is exactly two thirds of the corresponding leave-one-out distance.
Right ostium
Left ostium
Model
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
SwinUNETR-V2
0.63
0.82
10.3
1.21
150.0
0.65
0.94
5.17
1.31
131.6
50.7
(O: 65.52 | G: 24.14 | A: 1.72 | F: 8.62)
(O: 77.59 | G: 18.97 | A: 0.00 | F: 3.45)
ASOCA
0.91
1.05
1.18
1.32
1.88
1.29
1.72
29.2
29.7
112.8
(O: 50.00 | G: 25.00 | A: 25.00 | F: 0.00)
(O: 25.00 | G: 50.00 | A: 0.00 | F: 25.00)
CAT08
98.3
118.6
121.2
141.5
150.0
0.82
1.50
33.8
34.5
131.6
Appendix
Table B.1: Euclidean distance dc (mm) of SwinUNETR-V2 trained with the first 500 annotated ImageCAS images. Means and percentiles include the failed localizations. The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 .
Figure C.1: Right and left coronary ostium positions of all three datasets, after removing each patient’s own absolute position: every ostium is expressed relative to that patient’s own right-left midpoint, then rotated into a common frame whose x axis is the population-average left-to-right direction and whose y / z axes are the anatomical anterior/superior directions made orthogonal to it. Top: coronal-like projection ( x - y ). Bottom: sagittal-like projection ( x - z ). ASOCA (triangles) and CAT08 (squares) are shown as raw points; ImageCAS (circles, semi-transparent) additionally shows, for each side and projection, a 2 σ confidence ellipse computed from the ImageCAS points only. Its axes are the principal axes (principal component analysis) of the covariance of the projected points, and its semi-axes are twice the standard deviation along them, so it is an ellipse, not a circle, because the spread of the points differs between directions; it encloses about 86% of a bivariate normal distribution. Right ostium in firebrick, left ostium in navy.
Right ostium
Left ostium
ImageCAS images
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
0
1.18
1.93
34.53
53.89
170.58
2.08
36.89
50.94
99.01
139.27
48.7
(O: 24.14 | G: 22.41 | A: 18.97 | F: 34.48)
(O: 18.97 | G: 13.79 | A: 13.79 | F: 53.45)
ASOCA
0.64
1.17
15.79
16.31
60.34
2.94
3.25
21.07
21.39
75.31
(O: 50.00 | G: 25.00 | A: 0.00 | F: 25.00)
(O: 0.00 | G: 25.00 | A: 50.00 | F: 25.00)
CAT08
1.16
16.31
43.08
58.24
138.70
0.94
2.04
32.58
33.68
125.74
Appendix
Table D.1: Euclidean distance dc (mm) of SwinUNETR-V2 trained with an increasing number of ImageCAS images (training and validation sets combined; 1000 denotes the complete dataset, i.e., the 950 images not in the test set), always together with the whole of CAT08 and ASOCA (excluding the held-out cases of the test set). All the models are evaluated on the same test set. The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.1: Euclidean distance dc (mm) of SwinUNETR-V2 vs. the number of ImageCAS images used for training (equidistant axis; 1000 denotes the complete dataset, i.e., the 950 images not in the test set; CAT08 and ASOCA always included). Top: medians (lines) and p25–p75 (bands) of the right and left ostium over the whole test set, on a logarithmic axis. Bottom: percentage of failed localizations (error above 5 mm). See Section 3.8 .
Right ostium
Left ostium
Spacing (mm)
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
0.5
0.53
0.74
0.83
1.02
2.07
0.61
0.85
0.98
1.32
2.51
55.6
(O: 77.59 | G: 20.69 | A: 1.72 | F: 0.00)
(O: 84.48 | G: 15.52 | A: 0.00 | F: 0.00)
ASOCA
0.53
0.95
0.95
1.37
1.45
0.54
0.60
0.58
0.64
0.71
(O: 50.00 | G: 50.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
CAT08
0.60
1.02
1.04
1.46
1.75
0.59
1.04
1.20
1.64
2.25
Appendix
Table D.2: Euclidean distance dc (mm) of SwinUNETR-V2 trained and tested at the isotropic spacing (mm) reported in the first column. The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.2: Euclidean distance dc (mm) of SwinUNETR-V2 trained and tested at the same isotropic spacing (equidistant axis). Top: medians (lines) and p25–p75 (bands) of the right and left ostium over the whole test set. Bottom: processing time per patient. See Section 3.8 .
Right ostium
Left ostium
Spacing (mm)
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
0.5
0.52
0.70
0.83
1.13
2.03
0.68
0.87
0.96
1.21
2.32
41.1
(O: 72.41 | G: 24.14 | A: 3.45 | F: 0.00)
(O: 89.66 | G: 10.34 | A: 0.00 | F: 0.00)
ASOCA
0.61
0.67
0.77
0.83
1.20
0.58
0.74
0.67
0.83
1.06
(O: 75.00 | G: 25.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
CAT08
0.68
1.22
1.17
1.71
1.74
1.03
1.17
1.05
1.20
1.23
Appendix
Table D.3: Euclidean distance dc (mm) of nnResUNet trained and tested at the isotropic spacing (mm) reported in the first column. The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.3: Euclidean distance dc (mm) of nnResUNet trained and tested at the same isotropic spacing (equidistant axis). Top: medians (lines) and p25–p75 (bands) of the right and left ostium over the whole test set. Bottom: processing time per patient. See Section 3.8 .
Right ostium
Left ostium
Inference spacing (mm)
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
0.5
0.53
0.74
0.83
1.02
2.07
0.61
0.85
0.98
1.32
2.51
55.6
(O: 77.59 | G: 20.69 | A: 1.72 | F: 0.00)
(O: 84.48 | G: 15.52 | A: 0.00 | F: 0.00)
ASOCA
0.53
0.95
0.95
1.37
1.45
0.54
0.60
0.58
0.64
0.71
(O: 50.00 | G: 50.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
CAT08
0.60
1.02
1.04
1.46
1.75
0.59
1.04
1.20
1.64
2.25
Appendix
Table D.4: Euclidean distance dc (mm) of SwinUNETR-V2 trained at 0.5 mm and tested at the isotropic spacing (mm) reported in the first column. The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.4: Euclidean distance dc (mm) of SwinUNETR-V2 trained at 0.5 mm vs. inference spacing (equidistant axis, logarithmic distance). Lines: medians; bands: p25–p75 over the whole test set. Numbers at the bottom: p75, median (bold) and p25 of the right (red), left (blue) and both (gray) ostia. See Section 3.8 .
Right ostium
Left ostium
Window (vox)
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
5
0.59
0.77
0.84
1.00
2.04
0.63
0.86
0.98
1.33
2.54
55.8
(O: 77.59 | G: 20.69 | A: 1.72 | F: 0.00)
(O: 84.48 | G: 15.52 | A: 0.00 | F: 0.00)
ASOCA
0.56
0.94
0.95
1.33
1.43
0.53
0.60
0.58
0.65
0.70
(O: 50.00 | G: 50.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
CAT08
0.70
1.08
1.07
1.45
1.74
0.59
1.00
1.19
1.60
2.29
Appendix
Table D.5: Euclidean distance dc (mm) of SwinUNETR-V2 for different sizes of the centroid window (voxels; default 11). The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.5: Euclidean distance dc (mm) of SwinUNETR-V2 vs. centroid window size (equidistant axis; default 11). Lines: medians; bands: p25–p75 over the whole test set. Numbers at the bottom: p75, median (bold) and p25 of the right (red), left (blue) and both (gray) ostia. See Section 3.8 .
Right ostium
Left ostium
Overlap (vox)
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Time (s)
SwinUNETR-V2
0
0.52
0.77
0.86
1.09
2.09
0.67
0.97
1.12
1.46
3.02
30.1
(O: 72.41 | G: 20.69 | A: 6.90 | F: 0.00)
(O: 74.14 | G: 22.41 | A: 3.45 | F: 0.00)
ASOCA
0.66
0.85
0.81
1.00
1.10
0.57
0.64
0.61
0.68
0.75
(O: 75.00 | G: 25.00 | A: 0.00 | F: 0.00)
(O: 100.00 | G: 0.00 | A: 0.00 | F: 0.00)
Appendix
Table D.6: Euclidean distance dc (mm) of SwinUNETR-V2 and nnResUNet for different overlaps between adjacent sliding-window patches at inference (voxels, with 1283 patches; default 64). The line below each row gives the percentage of cases in each accuracy bucket for the right and left ostium: O = Optimal, G = Good, A = Acceptable, F = Failed. See Section 3.8 . In each distance column, the best value is in bold and the second best is underlined, comparing the rows of the same kind (overall and, separately, each dataset).
Figure D.6: Euclidean distance dc (mm) of SwinUNETR-V2 and nnResUNet vs. sliding-window overlap at inference (equidistant axis; default 64 voxels, 1283 patches). Lines: medians; bands: p25–p75 over the whole test set. Numbers at the bottom: p75, median (bold) and p25 of the right (red), left (blue) and both (gray) ostia. See Section 3.8 .
Right ostium
Left ostium
p25
Med.
Mean
p75
Max
p25
Med.
Mean
p75
Max
Trained on ImageCAS only, all external volumes
SwinUNETR-V2
ASOCA (40)
0.90
1.18
1.27
1.55
2.65
0.67
1.03
4.05
1.35
118.1
(O: 40.00 | G: 47.50 | A: 12.50 | F: 0.00)
(O: 80.00 | G: 12.50 | A: 5.00 | F: 2.50)
CAT08 (32)
0.61
0.77
0.96
1.20
2.65
0.74
0.89
1.18
1.41
3.49
Appendix
Table D.7: External validation: Euclidean distance dc (mm) of SwinUNETR-V2 and nnResUNet trained on the ImageCAS images only (training and validation sets), with no volume of ASOCA or CAT08 used for training, validation or model selection. Upper part: all the volumes of ASOCA (40) and CAT08 (32), and the 50 ImageCAS test volumes as the in-domain reference. Lower part: the 8 volumes of ASOCA (4) and CAT08 (4) that belong to the test set of the benchmark, compared with the models of the benchmark trained on the three datasets (Table 5 ); these 8 volumes were not used for training by either kind of model. The line below each row gives the share of cases in each accuracy bucket (O = Optimal, G = Good, A = Acceptable, F = Failed).
Right ostium
Left ostium
Both
Method
Mean
Med.
Mean
Med.
Mean
Med.
Zheng et al. [42] Private (1360)
1.73
1.31
1.66
1.45
1.70
-
Jang and Jeon [23] CAT08
1.94
1.80
2.22
1.70
2.08
1.59
Noothout et al. [22] Private (672)
-
1.45
-
1.55
-
-
Gajowczyk et al. [20] Private (201)
1.29
0.98
1.18
1.14
1.24
-
Leccardi et al. [44] ASOCA (40), CAT08 (32)
5.28
4.16
10.57
7.79
7.93
5.24
Appendix
Table E.1: Euclidean distance dc (mm), as mean and median (Med.), reported in the manuscripts of previously published methods for coronary ostium localization. The data used by each work, which differ from ours and are mostly private, are indicated in gray below the reference (number of volumes in brackets). The methods were not reimplemented and the values are copied from the respective works; a dash indicates a statistic that the work does not report. Both: combined right and left values. The last row is our earlier published model, evaluated on its own test split.
Both trees
Left tree
Right tree
Recall (%)
100 [100–100]
100 [100–100]
100 [100–100]
Precision (%)
99.9 [99.8–100]
99.9 [99.7–100]
100 [100–100]
OV (%)
99.9 [99.9–100]
99.9 [99.8–100]
100 [100–100]
OF (%)
100 [100–100]
100 [100–100]
100 [100–100]
AI (mm)
0.23 [0.22–0.24]
0.23 [0.22–0.25]
0.22 [0.21–0.24]
HD95 (mm)
0.50 [0.47–0.53]
0.50 [0.46–0.55]
0.48 [0.44–0.52]
Appendix
Table F.1: Agreement between the centerlines extracted from the 800 ImageCAS-X masks and the ImageCAS-X centerlines, for both trees together and for each tree, as median [interquartile range] over the patients. The metrics are those of Table 6 ; the overlap until first error and the bifurcation score are computed only over the patients in which they are defined (all of them for both trees together). The overlap of wide vessels and the radius errors are not reported, since the ImageCAS-X centerlines carry no radius.
Joining tolerance (mm)
Scans
Left trees
Right trees
1
30.4 (243)
28.6 (229)
2.5 (20)
2
18.1 (145)
17.3 (138)
1.3 (10)
3
5.4 (43)
5.1 (41)
0.4 (3)
5
0.6 (5)
0.6 (5)
0 (0)
10
0.1 (1)
0.1 (1)
0 (0)
Appendix
Table F.2: ImageCAS-X trees that remain split into pieces when the ends of their segments closer than the joining tolerance are joined, as a percentage (number) of the 800 scans and of the 800 left and 800 right trees.
Accurate segmentation of the coronary vessel lumen is a prerequisite for quantitative assessment of atherosclerotic plaque and perivascular adipose tissue in coronary computed tomography angiography (CCTA). Cardiologists rely on semi-automated methods for this task because manual vessel tracing and segmentation are labour-intensive. Although many automated methods have been proposed, their validation remains limited by the lack of large, high-quality publicly available datasets. We provide a new dataset of voxel-wise annotations of the vessel lumen and coronary segments, alongside centerlines, and mesh surfaces for 800 scans from the publicly available ImageCAS dataset. Using this dataset, we benchmark established lumen segmentation methods against inter-observer variability, stratifying performance by disease, image quality, coronary dominance, coronary segment, vessel diameter, and lumen attenuation. These labels allow segmentation accuracy to be described in anatomical and clinical context rather than reported as a single aggregate score. The dataset supports the development and validation of methods for lumen segmentation, plaque and perivascular quantification, and haemodynamic modelling.
Kit M. Bransby, Esther Øksnebjerg, Kristoffer Kjær +7
DTU Compute, Technical University of Denmark, Kongens Lyngby, Denmark · Cardiovascular Research Unit, Copenhagen University Hospital – Rigshospitalet, Copenhagen, Denmark
Coronary artery disease, a leading cause of cardiovascular mortality worldwide, can be assessed non-invasively by coronary computed tomography angiography (CCTA). Although deep learning has advanced automated CCTA analysis, clinical translation remains constrained by the scarcity of expert-annotated data and by the spatial sparsity of coronary pathology, which occupies only a small fraction of each scan. Widely used label-free pretraining strategies, such as masked image modeling and contrastive learning, optimize for global anatomical reconstruction and tend to under-represent these tiny localized pathological features. Here we present CORA, an annotation-efficient model for comprehensive coronary artery disease assessment. Rather than reconstructing background anatomy, CORA learns from volumetric CCTA through a synthesis-driven self-supervised strategy: an anatomy-guided engine inserts diverse synthetic calcified and non-calcified lesions into unlabeled scans, reframing pretraining as an abnormality-detection task that biases representation learning toward clinically relevant disease features. We pretrained CORA on 10,138 unlabeled CCTA volumes and evaluated it across datasets from nine independent hospitals. Across plaque characterization, stenosis detection, and coronary artery segmentation, CORA consistently outperformed strong self-supervised pretraining baselines, with the largest gains on external multi-center data, indicating robust generalization under distributional shift. Coupling the imaging encoder with structured clinical variables further enabled near-term major adverse cardiac event (MACE) risk stratification. Our results show that pathology-centric, synthesis-driven pretraining is an effective and scalable strategy for annotation-efficient coronary artery disease assessment from CCTA.
Jinkui Hao, Gorkem Durak, Halil Ertugrul Aktas +4
Department of Radiology, Northwestern University, Chicago, IL, USA · Department of Biomedical Engineering, Northwestern University, Evanston, IL, USA · Department of Cardiology, Northwestern University, Chicago, IL, USA +1
Accurate coronary artery segmentation on coronary computed tomography angiography (CCTA) is essential for diagnosing coronary artery disease. Deep networks are conventionally trained and evaluated with the Dice coefficient, but volume-overlap metrics are poorly suited to thin, tubular anatomy: since most voxels belong to a few thickproximal segments, a missing distal branch barely affects Dice despite severely disrupting the connectivity required for clinical use. We introduce a surface metric that matches predicted and reference surface points via bipartite assignment under a localized, vessel-radius tolerance, reporting precision, recall, and F1 with decoupled false positives (spurious branches) and false negatives (missed branches) a distinction the symmetric Dice cannot make. With it we show that a strong Dice-trained baseline omits far more vessel surface than it hallucinates, an asymmetry its high Dice hides. Building on this, we propose a differentiable surface loss that simultaneously suppresses spurious mass and recovers absent structure, validated by fine-tuning three backbones (nnU-Net, SwinUNETR, NexToU) on two public benchmarks (Image-CAS, ASOCA). Against a matched-epoch control, it significantly improves surface F1 by recovering missed distal vessels at comparable Dice. Our findings argue for measuring and optimizing the vessel surface, not the volume it overlaps. Code
Rafael Velasquez, Esther Puyol-Antón, Pablo Arbeláez
Universidad de los Andes, Bogotá, Colombia · School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK