Automated segment-level coronary artery calcium scoring on non-contrast CT: a multi-task deep-learning approach

Area of Science:

• Cardiology
• Radiology
• Artificial Intelligence

Background:

• Coronary artery calcium (CAC) scoring on non-contrast computed tomography (CT) is crucial for cardiovascular risk assessment.
• Manual CAC scoring is time-consuming and prone to interobserver variability.
• Automating CAC scoring can improve efficiency and accuracy.

Purpose of the Study:

• To develop and evaluate a multi-task deep-learning (DL) model for automated segment-level CAC scoring.
• To precisely localize and quantify calcifications within the coronary artery tree.
• To assess the model's agreement with human observers.

Main Methods:

• A multi-task neural network was developed using data from 1514 patients in the DISCHARGE trial.
• The model performed segment-level calcification segmentation and auxiliary segmentation of coronary artery segments.
• Performance was evaluated using sensitivity, specificity, F1-score, and Cohen's κ, with interobserver variability analysis.

Main Results:

• The model correctly assigned 73.2% of calcifications to the appropriate coronary artery segment in the test set.
• Achieved a micro-average sensitivity of 0.732, specificity of 0.978, and F1-score of 0.717.
• Demonstrated good segment-level agreement (Cohen's κ = 0.808), comparable to interobserver agreement (0.809).

Conclusions:

• Automated segment-level CAC scoring using a multi-task DL model shows significant potential.
• The DL approach achieves good agreement with human observers, reducing subjectivity.
• This technology can contribute to more efficient and accurate coronary artery calcification classification.