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

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Coronary Artery Disease (CAD) is a major cause of death globally.
• Accurate stenosis grading is vital for treatment but manual assessment is time-consuming and variable.
• Existing automated methods lack integration of radiomic and deep learning features.

Purpose of the Study:

• To develop and evaluate a machine learning approach combining radiomic and autoencoder (AE)-based features for stenosis grade evaluation.
• To improve the accuracy and reduce variability in CAD stenosis assessment using cardiac computed tomography angiography (CCTA) images.
• To compare the performance of combined features against individual radiomic and AE-based features.

Main Methods:

• Utilized 2,548 CCTA-derived multiplanar reconstructed (MPR) images from 220 patients.
• Extracted 64 AE-based and 465 2D radiomic features, processed separately and combined.
• Employed a random forest classifier with a cascade pipeline for a three-class stratification (no CAD, non-obstructive CAD, obstructive CAD).

Main Results:

• The combined model achieved a balanced accuracy of 0.91, sensitivity of 0.91, and specificity of 0.94 on the test set.
• The combined model significantly outperformed models using only AE-based (0.68 balanced accuracy) or radiomic features (0.82 balanced accuracy).
• Feature selection identified relevant AE-based, radiomic, and combined features for accurate classification.

Conclusions:

• Integrating radiomics and deep learning (AE-based features) offers a promising approach for automated stenosis assessment in CAD.
• This hybrid method enhances diagnostic accuracy and efficiency in evaluating coronary artery stenosis.
• The findings suggest a potential for improved clinical decision-making in CAD management.