Automated CAD-RADS scoring from multiplanar CCTA images using radiomics-driven machine learning

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Radiology
• Cardiovascular Disease Assessment

Background:

• Coronary Artery Disease-Reporting and Data System (CAD-RADS) scoring from CCTA is manual, time-consuming, and variable.
• Deep learning automation exists, but radiomics-based approaches with better interpretability are needed.
• This study addresses the need for explainable AI in CAD-RADS assessment.

Purpose of the Study:

• To develop and validate a novel radiomics-based machine learning model for automated CAD-RADS scoring.
• To compare the performance of radiomic, clinical, and combined models for CAD-RADS classification.
• To evaluate the utility of radiomics in therapy-oriented classification of coronary artery stenosis.

Main Methods:

• Retrospective study of 251 patients undergoing CCTA.
• Automated image segmentation, radiomic feature extraction, and data preprocessing.
• Development of a cascade pipeline for 6-class CAD-RADS and 4-class therapy-oriented classification using clinical, radiomic, and combined models with 5-fold cross-validation.

Main Results:

• Radiomic and combined models significantly outperformed the clinical model for CAD-RADS scoring (AUC 0.88 and 0.90 vs. 0.66).
• Radiomic and combined models showed superior performance in therapy-oriented classification (AUC 0.93 and 0.97 vs. 0.79).
• The models demonstrated significant improvements in classifying specific stenosis severity levels.

Conclusions:

• This study presents the first radiomics-based model for CAD-RADS classification.
• The developed model offers enhanced explainability and accuracy in coronary artery stenosis assessment.
• Radiomics provides a promising, interpretable AI tool to support radiologists in CCTA analysis.