P6166Radiomics-based machine learning versus histogram analysis and visual assessment to identify advanced atherosclerotic lesions on coronary computed tomography angiography

Abstract

Abstract Background Currently used coronary CT angiography (CTA) plaque classification and histogram-based methods have limited accuracy to identify advanced atherosclerotic lesions. Radiomics-based machine learning (ML) could provide a more robust tool to identify high-risk plaques. Purpose Our objective was to compare the diagnostic performance of radiomics-based ML against histogram-based methods and visual assessment of ex-vivo coronary CTA cross-sections to identify advanced atherosclerotic lesions as defined by histology. Methods Overall, 21 coronaries of seven hearts were imaged ex vivo with coronary CTA. From 95 coronary plaques 611 histological cross-sections were obtained and classified based-on the modified American Heart Association scheme. Histology cross-sections were considered advanced atherosclerotic lesions if early, late fibroatheroma or thin-cap atheroma was present. Corresponding coronary CTA cross-section were co-registered and classified into homogenous, heterogeneous, napkin-ring sign plaques based on plaque attenuation pattern. Area of low attenuation (<30HU) and average CT number was quantified. In total, 1919 radiomic parameters describing the spatial complexity and heterogeneity of the lesions were calculated in each coronary CTA cross-section. Eight different radiomics-based ML models were trained on randomly selected cross-sections (training set: 75% of the cross-sections) to identify advanced atherosclerotic lesions. Plaque attenuation pattern, histogram-based methods and the best ML model were compared on the remaining 25% of the data (test-set) using area under the receiver operating characteristic curves (AUC) to identify advanced atherosclerotic lesions using histology as a reference. Results After excluding sections with heavy calcium (n=32) and no visible atherosclerotic plaque on CTA (n=134), we analyzed 445 cross-sections. Based on visual assessment, 46.5% of the cross-sections were homogeneous (207/445), 44.9% heterogeneous (200/445) and 8.6% were with napkin-ring sign (38/445). Radiomics-based ML model incorporating 13 parameters significantly outperformed visual assessment, area of low attenuation and average CT number to identify advanced lesions (AUC: 0.73 vs. 0.65 vs. 0.55 vs. 0.53; respectively; p<0.05 for all). Conclusions Radiomics-based ML analysis may be able to improve the discriminatory power of CTA to identify high-risk atherosclerotic lesions.