All study methods received approval from the ethics committee of IRCCS San Raffaele Scientific Institute (study registration number CE 158/INT/2022) and adhered to relevant guidelines.

The proposed approach extracts topology-preserving centerlines from coronary lumen meshes extracted from DL-derived lumen segmentations (Fig. 1).

Fig 1: Overview of the proposed approach. After cardiac structures extraction using an open-source tool, CCTA images are preprocessed and then sliced along the three anatomical axes. Coronary lumen masks are obtained using a custom deep learning model, followed by mesh extraction and centerline computation.

AI-enabled Coronary Lumen Segmentation

A 2D encoder-decoder architecture was employed to segment the coronary lumen from individual CCTA slices (Fig. 2). The network utilized a pre-trained EfficientNet as the encoder’s backbone for hierarchical feature extraction, with skip connections to preserve spatial information during decoding.

Fig 2: Overview of the deep learning architecture. The proposed U-shaped encoder-decoder network incorporates skip connections to preserve fine spatial details during decoding. The encoder utilizes a pre-trained EfficientNet backbone for hierarchical feature extraction, complemented by mixing blocks that progressively refine these features.

The pre-processing pipeline included cardiac structure delineation using the open-source TotalSegmentator [1], followed by resizing to 256x256 along the z-axis, Z-score normalization, intensity clipping, and contrast-limited adaptive histogram equalization to enhance vascular structures. The same preprocessing was applied at inference stage. The model was implemented in PyTorch and optimized by minimizing the binary cross-entropy loss using AdamW optimizer with an initial learning rate of 0.01 and weight decay of 0.0001. Training was conducted with a batch size of 64 on an 80/20 train/validation split, while a Cosine Annealing scheduler was employed to dynamically adjust the learning rate. To reduce the risk of overfitting, early stopping was implemented with a patience of 3 epochs throughout the 100-epoch training process.

During inference, the model generated lumen probability maps for slices along each anatomical axis (axial, sagittal, and coronal). The final segmentation was obtained through majority voting, where voxels were retained if selected by at least two of the three orthogonal predictions.

Data

The training dataset consisted of 64 CCTA scans with coronary lumen annotations from two different cohorts. The first cohort included 40 cases from the public Automatic Segmentation of Normal and Diseased Coronary Arteries (ASOCA) dataset [2], acquired using retrospective ECG-gated protocol on a GE Lightspeed 64 slice scanner with end-diastolic reconstruction. The second cohort consisted of 24 cases from our institution (HSR) acquired on a Siemens Healthcare SOMATOM Definition Flash using both prospective and retrospective ECG-gated acquisitions. For HSR cases, the reconstruction with minimal motion artifacts was selected from multiple cardiac phases (end-diastolic, end-systolic, and 10% intervals throughout the cardiac cycle).

For validation, we used an independent test set of 20 unseen CCTA scans with varying CAD severity from our institution. Ground-truth annotations were obtained through different approaches for each cohort: majority voting from three experienced annotators for ASOCA cases, and semi-automatic segmentation using MEDIS QAngioCT suite with expert supervision for HSR cases. All images were acquired with scanners from multiple vendors using different protocols and reconstruction algorithms, with in-plane resolution ranging from 0.3 to 0.4 mm and slice thickness from 0.3 to 0.62 mm.

Mesh extraction

Surface meshes were extracted from binary lumen segmentations using the discrete flying edges algorithm, which generates triangulated surfaces at the boundaries between binary regions while preserving the topology of the segmented structures [3, 4]. The extracted surfaces underwent Taubin smoothing with 500 iterations and a passband of 0.1 to reduce noise while preserving geometric features. These closed meshes were opened at the coronary ostia regions through intersection with the aorta. Local bounding boxes at these interfaces guided the mesh clipping, creating the inlets required for centerline extraction.

Centerline computation

Centerlines were extracted using the Vascular Modeling ToolKit (VMTK), which computes the embedded Voronoi diagram from the vascular meshes [5]. The centerline paths were then obtained by tracing the steepest descent of the eikonal solution computed on the Voronoi diagram, starting from each mesh opening. This approach generated a distinct centerline for each branch in the DL-derived coronary tree. Centerlines were uniformly resampled at 0.1 mm intervals to ensure consistent point spacing along the vessel paths.

Performance evaluation

Segmentation performance was assessed against reference segmentations using Intersection over Union (IoU) and Dice similarity coefficient (DSC). Centerline quality was assessed against reference centerlines using percentage overlap (OV) and average inside distance (AD), which measures the mean distance between reference and extracted centerline for predicted centerline points that are within the radius of the reference centerline.