Radiomics versus Visual and Histogram-based Assessment to Identify Atheromatous Lesions at Coronary CT Angiography: An ex Vivo Study.

Background Visual and histogram-based assessments of coronary CT angiography have limited accuracy in the identification of advanced lesions. Radiomics-based machine learning (ML) could provide a more accurate tool. Purpose To compare the diagnostic performance of radiomics-based ML with that of visual and histogram-based assessment of ex vivo coronary CT angiography cross sections to identify advanced atherosclerotic lesions defined with histologic examination. Materials and Methods In this prospective study, 21 coronary arteries from seven hearts obtained from male donors (mean age, 52.3 years ± 5.3) were imaged ex vivo with coronary CT angiography between February 23, 2009, and July 31, 2010. From 95 coronary plaques, 611 histologic cross sections were coregistered with coronary CT cross sections. Lesions were considered advanced if early fibroatheroma, late fibroatheroma, or thin-cap atheroma was present. CT cross sections were classified as showing homogeneous, heterogeneous, or napkin-ring sign plaques on the basis of visual assessment. The area of low attenuation (<30 HU) and the average Hounsfield unit were quantified. Radiomic parameters were extracted and used as inputs to ML algorithms. Eight radiomics-based ML models were trained on randomly selected cross sections (training set, 75% of the cross sections) to identify advanced lesions. Visual assessment, histogram-based assessment, and the best ML model were compared on the remaining 25% of the data (validation set) by using the area under the receiver operating characteristic curve (AUC) to identify advanced lesions. Results After excluding sections with no visible plaque ( = 134) and with heavy calcium ( = 32), 445 cross sections were analyzed. Of those 445 cross sections, 134 (30.1%) were advanced lesions. Visual assessment of the 445 cross sections indicated that 207 (46.5%) were homogeneous, 200 (44.9%) were heterogeneous, and 38 (8.5%) demonstrated the napkin-ring sign. A radiomics-based ML model incorporating 13 parameters outperformed visual assessment (AUC = 0.73 with 95% confidence interval [CI] of 0.63, 0.84 vs 0.65 with 95% CI of 0.56, 0.73, respectively; = .04), area of low attenuation (AUC = 0.55 with 95% CI of 0.42, 0.68; = .01), and average Hounsfield unit (AUC = 0.53 with 95% CI of 0.42, 0.65; = .004) in the identification of advanced atheromatous lesions. Conclusion Radiomics-based machine learning analysis improves the discriminatory power of coronary CT angiography in the identification of advanced atherosclerotic lesions. Published under a CC BY 4.0 license.