Coronary computed tomographic angiography (CCTA) is a non-invasive imaging technique widely used for diagnosing coronary artery disease (CAD), one of the leading causes of mortality in developed countries. Accurate and automatic segmentation of coronary arteries from CCTA is essential for extracting both anatomical and pathological information. Existing deep learning methods suffer from noise artifacts and vessel discontinuities, while classical image processing methods including fixed Hounsfield unit (HU) threshold are highly dependent on scanner characteristics. In this study, we proposed a novel hybrid method that integrated deep learning with our unique mathematical integration of image processing filters, featuring a contour detection algorithm that exploited intensity gradients. The performance of our method was quantitatively evaluated using CCTA scans from 84 patients (internal validation set) and 40 patients from a public dataset (external validation set), with segmentation results compared against manually annotated reference data. We also evaluated existing deep learning-only and classical fixed HU threshold methods against the same reference data for comparison. Our hybrid method demonstrated superior performance with a Dice score of 0.92 (95% confidence interval CI: 0.91–0.93), significantly outperforming deep learning-only (0.68, 95% CI: 0.66–0.69, p < 0.001) and fixed HU threshold methods (0.55, 95% CI: 0.53–0.56, p < 0.001). External validation on public datasets confirmed significantly better performance with a Dice score of 0.82 (95% CI: 0.81–0.82) compared to deep learning-only (0.76, 95% CI: 0.74–0.77, p < 0.001) and fixed HU threshold methods (0.76, 95% CI: 0.75–0.77, p < 0.001). These results indicate that our hybrid method enables robust and consistent automatic coronary artery segmentation from CCTA, demonstrating potential to aid CAD assessment in clinical practice.
Park et al. (Thu,) studied this question.