Identification of noncalcified coronary plaque characteristics using machine learning radiomic analysis of non-contrast high-resolution computed tomography.

Abstract

Novel imaging and analysis techniques may offer the ability to detect noncalcified or high-risk coronary plaques on a non-contrast computer tomography (CT) scan, advancing cardiovascular diagnostics. We aimed to explore whether machine learning (ML) radiomic analysis of low-dose high-resolution non-contrast electrocardiographically (ECG) gated cardiac CT scan allows for the identification of noncalcified coronary plaque characteristics. We prospectively enrolled 125 patients with noncalcified plaques and adverse plaque characteristics (APC) and 25 controls without visible atherosclerosis on coronary CT angiography (CCTA). All patients underwent non-contrast CT exam before CCTA. Four hundred and nineteen radiomic features were calculated to identify the presence of any coronary artery disease (CAD), obstructive CAD (stenosis >50%), plaque with ≥2 APC, degree of calcification, and specific APCs. ML models were trained on a training set (917 segmentations) and tested (validation) on a separate set (292 segmentations). Among the radiomic features, 88.3% were associated with a plaque, 0.9% with obstructive CAD, and 76.4% with the presence of at least two APCs. Overall, 80.2%, 88.5%, and 36.5%, of features were associated with calcified, partially calcified, and noncalcified plaques, respectively. Regarding APCs, 61.1%, 61.8%, 84.2%, and 61.3% of features were associated with low attenuation (LAP), napkin-ring sign (NRS), spotty calcification (SC), and positive remodeling (PR), respectively. ML models outperformed conventional methods for the presence of plaque obstructive stenosis, and the presence of 2 APCs, as well as for noncalcified plaques and partially calcified plaques, but not for calcified plaques. ML models also significantly outperformed identification of LAP and PR, but neither NRS nor SC. Radiomic analysis of non-contrast cardiac CT exams may allow for the identification of specific noncalcified coronary plaque characteristics displaying the potential for future clinical applications.