Automated Coronary Artery Calcium Scoring Using Convolutional Neural Networks: Enhancing Cardiovascular Risk Assessment in Chest CT Scans

Abstract

Background: Coronary artery calcium (CAC) scoring is valuable for cardiovascular risk assessment but often time-consuming and subject to variability. This study aimed to develop and validate a convolutional neural network (CNN) model for automated CAC scoring in chest CT scans, potentially enhancing efficiency and accuracy. Methods: We utilized 10,000 chest CT scans from a public dataset, split into training (n=7,000), validation (n=1,500), and testing (n=1,500) sets. A 3D CNN model based on ResNet-50 was developed and trained for CAC detection and quantification. Performance was evaluated on the test set and compared to manual scoring by three experienced radiologists. Results: The CNN model achieved 93.7% accuracy in detecting CAC, with 87.4% sensitivity and 92.1% specificity for identifying clinically significant CAC (Agatston score >100) in the test set (n=1,500). The model showed strong correlation with manual CAC scores (r=0.89, p<0.001). Automated scoring reduced processing time by 78% compared to manual techniques, averaging 18.3 seconds per scan. The model demonstrated consistent performance across diverse patient demographics and CT types. In a subset of patients with follow-up data (n=500), the model's risk stratification was comparable to the Framingham Risk Score in predicting cardiovascular events (AUC 0.76 vs 0.74, p=0.09). Conclusions: The CNN-based automated CAC scoring system demonstrated high accuracy and efficiency, potentially enabling more widespread cardiovascular risk assessment in routine chest CT scans. Future research should focus on prospective validation and investigation of long-term patient outcomes when integrating this technology into clinical practice.