TY  - EJOU
AU  - Gladston, Angelin 
AU  - Velraj, Swedha 
AU  - Nadarajan, Harini 

TI  - Hybrid Multi-Scale 3D Segmentation Framework for Automated Stenosis Detection
T2  - Structural and Congenital Heart Disease

PY  - 2025
VL  - 20
IS  - 6
SN  - 3071-1738

AB  - <b>Background:</b> Coronary artery disease (CAD) is a major global health concern requiring efficient and accurate diagnostic methods. Manual interpretation of coronary computed tomography angiography (CTA) images is time-consuming and prone to interobserver variability, underscoring the need for automated segmentation and stenosis detection tools. <b>Methods:</b> This study presents a hybrid multi-scale 3D segmentation framework utilizing both 3D U-Net and Enhanced 3D U-Net architectures, designed to balance computational efficiency and anatomical precision. Processed CTA images from the ImageCAS dataset underwent data standardization, normalization, and augmentation. The framework applies ensemble learning to merge coarse and fine segmentation masks, followed by advanced post-processing techniques, including connected component analysis and centerline extraction, to refine vessel delineation. Stenosis regions are detected using the Enhanced 3D U-Net and morphological operations for accurate localization. <b>Results:</b> The proposed pipeline achieved near-perfect segmentation accuracy (0.9993) and a Dice similarity coefficient of 0.8539 for coronary artery delineation. Precision, recall, and F1 scores for stenosis detection were 0.8418, 0.8289, and 0.8397, respectively. The dual-model approach demonstrated robust performance across varied anatomical structures and effectively localized stenotic regions, indicating clear superiority over conventional models. <b>Conclusion:</b> This hybrid framework enables highly reliable and automated coronary artery segmentation and stenosis detection from 3D CTA images. By reducing reliance on manual interpretation and enhancing diagnostic consistency, the proposed method holds strong potential to improve clinical workflows for CAD diagnosis and management.
KW  - Coronary artery disease; computed tomography angiography; three-dimensional U-Net; stenosis detection; deep learning; medical image segmentation; ensemble learning

DO  - 10.32604/chd.2025.068879