Coronary segmentation performance

The method showed high accuracy in segmentation with IoU of 76%±3 and DSC of 0.87±0.02 (Fig. 3).

Fig 3: Example of coronary lumen segmentations obtained from the three anatomical axes (axial, sagittal, and coronal) and the final binary mask resulting from majority voting fusion.

Centerline extraction performance

Quantitative evaluation against reference centerlines yielded an OV of 92%±0.09 and AD of 0.48±0.06 mm, indicating high extraction accuracy. Representative segmentation and centerline results are shown in Fig. 4, with straightened MPR views comparing automated and reference centerlines along major coronary vessels shown in Fig. 5.

Fig 4: Comparison of deep learning lumen segmentations and centerlines across two cases (A, B). For each case, axial cross-sections (left, middle) and 3D reconstruction (right) demonstrate the algorithm's performance in different vessel orientations and anatomical contexts.

Fig 5: Straightened multiplanar reconstructions (MPR) comparing inferred (violet and cyan) and reference (yellow) centerlines along the right coronary artery (RCA), left main-left anterior descending (LM-LAD), and left circumflex (LCX) arteries.

Computational efficiency

The pipeline extracts coronary tree centerlines automatically in under a minute. Training required approximately 60 hours on a NVIDIA RTX A6000 48GB.