Patient-level CAD-RADS scoring from coronary radiomic features

Area of Science:

• Cardiovascular Imaging
• Radiomics
• Machine Learning

Background:

• Synthesizing patient-wise Coronary Artery Disease-Reporting and Data System (CAD-RADS) scores from coronary radiomic data is challenging.
• Coronary computed tomography angiography (CCTA) generates extensive multiplanar reconstruction (MPR) images crucial for cardiac assessment.

Purpose of the Study:

• To develop and evaluate four distinct strategies for summarizing coronary radiomic features to derive a single patient-level CAD-RADS score.
• To compare the performance of statistical-based versus majority voting approaches for automated CAD-RADS classification.

Main Methods:

• Extracted radiomic features from 2779 MPR images of 238 patients undergoing CCTA.
• Developed a cascade pipeline with gradient boosting classifiers for CAD-RADS scoring.
• Implemented two statistical (average, min, max, std dev) and two majority voting (MV_P, MV_C) methods for patient-level scoring.
• Utilized an 80%-20% training/test split with five-fold cross-validation.

Main Results:

• Majority voting approaches outperformed statistical-based methods in CAD-RADS scoring.
• The MV_P approach achieved high AUCs across various CAD-RADS categories (e.g., 0.94 for CAD-RADS 0, 0.97 for CAD-RADS 2).
• The MV_C approach also demonstrated strong performance, particularly for higher CAD-RADS grades (e.g., 0.96 for CAD-RADS 3, 0.98 for CAD-RADS 4).

Conclusions:

• Majority voting strategies, particularly MV_P, offer a robust and reproducible method for automated patient-wise CAD-RADS scoring.
• This work represents a significant step towards reliable coronary radiomics tools for clinical application.
• The proposed methods enhance the potential for automated CAD-RADS assessment in clinical practice.