Noting the limitations with traditional coronary artery calcium (CAD) scoring and the time-consuming challenges with manual CAC scoring of coronary artery segments, the authors of a new study suggested a deep learning model may automate CAC scoring on non-contrast computed tomography (CT) with improved localization and quantification of calcifications.
For the multicenter study, recently published in Insights into Imaging, researchers reviewed data from 1,514 patients (mean age of 60) with stable chest pain. After developing a neural network that segments calcifications at the coronary artery segment level as well as the regional level in a training/validation set of 1,059 patients, the study authors tested the deep learning model in 455 patients (1,797 calcifications), according to the study.
The researchers found that the deep learning model had a 73.2 percent accuracy rate for assigning calcifications to coronary artery segments, had a micro-average specificity of 97.8 percent and achieved an 80.8 percent segment-level agreement in comparison to 80.9 percent between the two reviewing radiologists.
“Coronary calcium scoring on the segment level might improve the predictive value of calcium scoring since information on calcium distribution across the segments of the coronary artery tree is not lost, and the presence of bifocal/trifocal calcified lesions can be considered in risk prediction. The proposed model could also assist in CT angiography planning and serve as a gatekeeper to avoid unnecessary subsequent angiography in the presence of a high calcium score in high-risk segments,” wrote lead study author Bernard Follmer, M.D., who is affiliated with the Department of Radiology at Charite Berlin University Hospital in Berlin, and colleagues.
For multiclass segmentation of calcified lesion volume, the study authors noted an average micro-F1 score of 73.2 percent and over a 20 percent higher average micro-F1 score of 94.2 percent at the vessel level. When they examined the use of a binary module for classifying CAC, the researchers found a 90.7 percent F1 score for binary segmentation.
The researchers conceded that the deep learning model had low sensitivity for the side right coronary artery (RCA) (0 percent), the distal left anterior descending (LAD) artery (50 percent) and the mid-LAD artery (53 percent) at the segment level.
Three Key Takeaways
1. Deep learning for automated CAC scoring. The study proposes a deep learning model to automate coronary artery calcium (CAC) scoring on non-contrast CT, providing improved localization and quantification of calcifications at the segment and regional levels, which may enhance predictive value and clinical utility.
2. Accuracy and specificity. The deep learning model showed a 73.2 percent accuracy in assigning calcifications to coronary artery segments, high micro-average specificity (97.8 percent) and an 80.8 percent agreement at the segment level, similar to that of radiologists.
3. Limitations and utility. Although the model had lower sensitivity in certain coronary artery segments (e.g., RCA and LAD), it showed high sensitivity in proximal segments like the proximal RCA (94 percent) and proximal LCX (92 percent), making it potentially useful for identifying proximal CAC, a marker of major adverse cardiovascular events.
However, the study authors also pointed out high sensitivity rates for the model with respect to the proximal left circumflex (LCX), the distal RCA (92 percent) and the proximal RCA (94 percent).
“Both sensitivity and F1 scores were higher for proximal segments compared with side branches, indicating that the model may be particularly useful for automatically identifying proximal CAC, which is an independent marker of major adverse cardiovascular events,” noted Follmer and colleagues.
(Editor’s note: For related content, see “FDA Clears Software for Enhancing CCTA Assessment of Atherosclerosis,” “Can AI Enhance CT Detection of Incidental Extrapulmonary Abnormalities and Prediction of Mortality?” and “FDA Clears Updated AI Software for CT-Based Coronary Artery Calcium Assessment.”)
In regard to study limitations, the authors acknowledged a high interobserver variability with CAC segment scoring, which was particularly evident for side branches of the coronary artery tree. They also conceded that the reference standard for training and validation of the deep learning model relied on annotations from one radiologist with three years of experience.
