HALE: Healthy area of lumen estimation for vessel stenosis quantification

Abstract

One of the most widely used non-invasive clinical metric for diagnosing patients with symptoms of coronary artery disease is %stenosis derived from cCTA. Estimation of %stenosis involves two steps-the measurement of local diameter and the measurement of a reference healthy diameter. The estimation of a reference healthy diameter is challenging,especially in diffuse,ostial and bifurcation lesions. We develop a machine learning algorithm using random forest regressors for the estimation of healthy diameter using downstream and upstream properties of coronary tree vasculature as features. We use a population-based estimation,in contrast to single patient estimation that is used in the majority of the literature. We demonstrate that this method is able to predict the diameter of healthy sections with a correlation coefficient of 0.95.We then estimate %stenosis based on the ratio of the local vessel diameter to the estimated healthy diameter. Compared to a reference anisotropic kernel regression method,the proposed method,HALE (Healthy Area of Lumen Estimation),has a superior area under curve (0.90 vs 0.83) and operating point sensitivity/specificity (90%/85% vs 82%/76%) for the detection of stenoses. We also demonstrate superior performance of HALE against invasive quantitative coronary angiography (QCA),compared to the reference method (mean absolute error: 14% vs 31%,p<0.001).