CABGenie: Deep Learning-Based Coronary Artery Segmentation with Optuna-Driven Hyperparameter Tuning for CABG Planning

Abstract

Coronary artery disease (CAD) remains a leading cause of mortality worldwide. In severe cases, coronary artery bypass grafting (CABG) as a critical surgical intervention. Accurate preoperative planning relies heavily on the precise delineation of coronary arteries from computed tomography angiography (CTA) images, a time-consuming and variable manual task. This study presents CABGenie, a deep learning-based system that integrates five state-of-the-art semantic segmentation models (UNet, VNet, DynUNet, SegResNet, and UNETR). These models are optimized via an automated hyperparameter tuning using Optuna to perform coronary artery segmentation from 3D CCTA images in the publicly available ImageCAS dataset. The system includes a web interface that supports image upload, segmentation inference, and interactive 3D visualization using ITK/VTK viewers. Results demonstrate improved model performance after tuning, with DynUNet achieving the highest Dice Similarity Coefficient of 0.7657. The performance could still be improved and there is a need for further model refinement to enhance segmentation accuracy.