Tianyi Guan

Abstract

Coronary computed tomography angiography (CCTA) segmentation masks are widely used as intermediate outputs in medical image-analysis pipelines, but binary masks alone do not provide structured descriptors for downstream analysis. This work introduces CoroConcept, a three-dimensional post-segmentation image-processing pipeline that converts CCTA coronary masks and corresponding intensity data into case-level, branch-level, candidate-level, quality-control, and visual-audit descriptors. The pipeline preserves the coronary tree, assigns major branches, extracts graph-based centrelines, computes length, diameter, and tortuosity profiles, identifies peripheral high-Hounsfield-unit candidates, estimates local centreline-derived geometric descriptor changes, and applies reliability-controlled filtering. CoroConcept was evaluated on 1000 ImageCAS CCTA cases, generating 3996 branch-level outputs and 2086 broad computational candidate outputs, of which 1367 satisfied primary reliability criteria. All cases were successfully processed, all core visual outputs were available, and the median automated quality-assessment score was 100. In a random 100-case manual validation, branch-identification agreement ranged from 82% to 93%. Among correctly identified branches, the proportions of centreline-length and mean-diameter measurements within 15% of the manual estimates ranged from 83% to 97% and from 98% to 100%, respectively. A 100-case perturbation analysis showed that global vessel descriptors were relatively stable under one-voxel mask perturbations, whereas local centreline-derived candidate descriptors were more sensitive to erosion and dilation. These findings support CoroConcept as an auditable 3D image-processing pipeline for converting CCTA coronary segmentation masks into structured quantitative descriptors.

Keywords: CCTA; 3D image processing; segmentation masks; centreline extraction; descriptor extraction; robustness analysis