Straightening the path to clarity: A subpixel-level vessel segmentation framework in X-ray coronary angiography

Abstract

Coronary artery disease (CAD) is the leading cause of death globally. X-ray coronary angiography (XCA) is the standard method for routine evaluation of coronary artery disease and its precise quantitative analysis relies heavily on contour segmentation. However, existing direct contour segmentation algorithms for XCA vessels can only achieve pixel-level accuracy, compromising the reliability of quantitative measurements. To address this, we propose the first deep learning-based framework for subpixel-level XCA vessel segmentation, achieving an average error of less than one pixel. The framework includes an automated vessel landmarks localization to identify main vessels and stenotic lesions, followed by a planar coordinate transformation to convert vessels into a straightened view. Subsequently, we designed an efficient SuPP-Net for subpixel contour prediction on the straightened view, which was ultimately transformed back to the original image coordinates. Our method achieved state-of-the-art performance on clinical data, with a contour MSE of 0.53 ± 0.30 pixel and an average diameter stenosis error of 3.23 ± 2.51%. Beyond achieving subpixel-level accuracy, our framework specifically addresses diverse stenotic lesion types, optimizes labeling techniques, and enables a fully automated workflow. Moreover, the method demonstrates robust generalization across different image quality, vessel perturbation levels, and external noise levels. This subpixel analysis of XCA vessels meets the precision demands of coronary anatomical and physiological assessments, thereby may enhance CAD diagnosis and treatment strategies.