Patient-level CAD-RADS scoring from coronary radiomic features

Synthesizing coronary radiomic data to obtain a single patient-wise Coronary Artery Disease-Reporting and Data System (CAD-RADS) score remains challenging. This work proposes four strategies for summarizing radiomic features extracted from 2779 multiplanar reconstruction images derived from coronary computed tomography angiography of 238 patients. A cascade pipeline was developed to train gradient boosting classifiers for CAD-RADS scoring through consecutive tasks, considering 80%-20% training/test split with five-fold cross-validation on the training set. Two statistical-based and two majority voting approaches were implemented to obtain patient-level classification. The former consisted in computing features average, minimum, maximum and standard deviation, across the coronary images, leading to intermediate coronary classification, followed by patient classification according to the worst coronary class. The latter consisted in single image predictions and the application of majority voting either to all the images, to obtain patient classification (MV_P), or to the images of single coronary arteries, followed by patient classification according to the worst coronary class (MV_C). Majority-voting approaches outperformed statistical-based ones, with MV_P achieving an AUC of CAD-RADS_0 = 0.94, CAD-RADS_1 = 0.92, CAD-RADS_2 = 0.97, CAD-RADS_3 = 0.77, CAD-RADS_4 = 0.88, CAD-RADS_5 = 0.85, and MV_C of CAD-RADS_0 = 0.82, CAD-RADS_1 = 0.78, CAD-RADS_2 = 0.84, CAD-RADS_3 = 0.96, CAD-RADS_4 = 0.98 and CAD-RADS_5 = 0.85. This study represents a significant advancement toward robust and reproducible coronary radiomics tools for automated CAD-RADS scoring.