Threshold quantification of coronary artery blockage using a cascaded ensemble U-Net-based pipeline

Abstract

Coronary artery disease remains a primary driver of morbidity and mortality world- wide, with precise quantification of arterial stenosis essential for risk stratification

and treatment planning. We propose a fully automated, cascaded–ensemble deep

learning framework for 3D CTA analysis that integrates coarse whole‐volume seg- mentation with multi‐model patch‐level refinement and skeleton‐based stenosis as- sessment. First, a 3D U‐Net coarse network localizes the coronary tree across the

entire volume. Regions of interest are then cropped via dilated centerline extrac- tion and passed through an ensemble of four patch‐level architectures—3D U‐Net,

3D ResU‐Net, Attention U‐Net, and U‐Net++—to capture fine vessel details. Pre- dicted centerlines are skeletonized and compared against ground truth to compute

percentage stenosis, leveraging distance‐transform radii measurements. We evaluate performance over 40 training epochs, reporting AUC, Dice, precision, recall, F1 for training, validation, and test cohorts, alongside bootstrapped 95% CIs, Wilcoxon p‐values, and correlation matrices. Visual comparisons of all model outputs on the same Frangi‐enhanced slices demonstrate complementary strengths, while ensemble voting yields more robust boundary delineation. On a 1000‐case CTA dataset, this research’s approach achieves a mean Dice >0.80 and significantly less stenosis error, illustrating its potential and reliability than other prior researches and 2D Unet Models.