Beyond plaque segmentation: a combined radiomics-deep learning approach for automated CAD-RADS classification

Coronary Artery Disease (CAD) is a leading cause of global mortality, accurate stenosis grading is crucial for treatment planning, it currently requires time-consuming manual assessment and suffers from interobserver variability. Few deep learning methods have been proposed for automated scoring, but none have explored combining radiomic and autoencoder (AE)-based features. This study develops a machine learning approach combining radiomic and AE-based features for stenosis grade evaluation from multiplanar reconstructed images (MPR) cardiac computed tomography (CCTA) images.

The dataset comprised 2,548 CCTA-derived MPR images from 220 patients, classified as no-CAD, non-obstructive CAD or obstructive CAD. Sixty-four AE-based and 465 2D radiomic features, were processed separately or combined. The dataset was split into training (85%) and test (15%) sets. Relevant features were selected and input to a random forest classifier. A cascade pipeline stratified the three classes via two sub-tasks: (a) no CAD vs. CAD, and (b) nonobstructive vs. obstructive CAD.

The AE-based model identified 17 and 6 features as relevant for the sub-task (a) and (b), respectively, while 44 and 30 features were selected in the radiomic model. The two models reached an overall balanced accuracy of 0.68 and 0.82 on the test set, respectively. Fifteen and 35 features were indeed selected in the combined model which outperformed the single ones achieving on the test set an overall balanced accuracy, sensitivity and specificity of 0.91, 0.91, and 0.94, respectively.

Integration of radiomics and deep learning shows promising results for stenosis assessment in CAD patients.