Automatic coronary artery calcium scoring from unenhanced-ECG-gated CT using deep learning

Area of Science:

• Cardiology
• Radiology
• Artificial Intelligence

Background:

• Coronary artery calcium (CAC) scoring is crucial for cardiovascular risk assessment.
• Manual CAC scoring from CT scans is time-consuming and prone to inter-observer variability.
• Automated CAC quantification could improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for automated CAC scoring.
• To utilize convolutional neural networks (CNNs) for estimating CAC from unenhanced ECG-gated cardiac CT.
• To compare the algorithm's performance against manual radiologist estimations.

Main Methods:

• A set of five 3D U-Net CNNs was trained on 783 CT examinations.
• The algorithm detects and segments coronary artery calcifications in 3D volumes.
• The Agatston score was computed from segmentation masks and compared to ground truth.

Main Results:

• The developed algorithm achieved a high concordance index (C-index) of 0.951 on an independent testing set.
• The model demonstrated robust performance in CAC quantification.
• Minor errors were noted for small, low-density calcifications or those near the mitral valve/ring.

Conclusions:

• The deep learning-based method provides fast and robust automated CAC scoring from cardiac CT.
• The accuracy is comparable to existing AI methods, enhancing workflow efficiency.
• Automated Agatston score calculation eliminates manual segmentation time, optimizing clinical practice.