Generalizable Pipeline for Automated Coronary Calcium Scoring Across CT Protocols

• Applicant: Bernhard Föllmer, Ph.D.
• Subject Area: Radiology ; Methods in Artificial Intelligence and Machine Learning
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 568331058

Cardiovascular disease (CVD) remains the leading cause of mortality worldwide. Coronary artery calcium (CAC) is a key indicator of coronary atherosclerosis and a strong predictor of CVD risk and patient outcomes. Traditionally, CAC scoring is performed using electrocardiographic-gated (ECG-gated), non-contrast CT scans. However, low-dose chest CT scans, widely used for lung cancer screening, offer an alternative for CAC assessment, particularly as their prevalence is expected to increase in the coming years. Additionally, recent studies have demonstrated the feasibility of CAC scoring in coronary CT angiography (CTA), presenting an opportunity to reduce radiation exposure by eliminating the need for separate calcium scoring scans. However, generalizing these models towards a variety of scan protocols remains challenging due to variations in acquisition and reconstruction parameters. A generalizable pipeline for automated CAC scoring across CT protocols could address these limitations by adapting deep learning models to varying data domains and acquisition settings. Despite recent advances, existing methods often fail to integrate physicians into the development and evaluation process by user-friendly interfaces and explainable methods (XAI) to enhance physician trust, improve transparency, and facilitate patient acceptance. Our goal is to develop and implement a generalizable deep learning pipeline for CAC assessment across CT protocols. This involves adapting existing CAC scoring models from ECG-gated, non-contrast CT to low-dose chest CT, CTA, and other coronary CT protocols with varying acquisition and reconstruction parameters. To achieve this, we will employ active domain adaptation (ADA) to enhance model robustness across different scan types and imaging domains. To support model development and validation, we will establish the CAC-Scoring Development Hub, a platform for testing and refining advanced calcium scoring methodologies. Additionally, we will design a Human-AI interface for image quality assessment, performance evaluation, and explainable AI (XAI), ensuring meaningful physician engagement in the decision-making process. The CAC scoring pipeline will be trained and internally validated on a dataset comprising more than 3,500 ECG-gated, non-contrast CTs (DISCHARGE, CAD-Man (Extend), SCOT-HEART, Charité routine data) and over 22,500 low-dose chest CTs (Charité routine data, NLST), alongside various other scan protocols. External validation will be conducted using more than 5,000 CT scans from the Scottish Medical Imaging dataset and over 28,000 CTs from the SCAPIS trial to ensure robustness and clinical applicability. We anticipate that the findings from our generalizable CAC scoring pipeline will facilitate its integration into clinical practice, enhancing automated CAC assessment, and increasing physician confidence in AI-driven decision-making.

• DFG Programme: Research Grants
• International Connection: Sweden, United Kingdom
• Co-Investigator: Professor Dr. Marc Dewey
• Cooperation Partners: Professor Dr. Göran Bergström; Professorin Dr. Michelle Williams