Deep learning-based detection of coronary artery calcification in non-contrast and contrast-enhanced CT scans

Coronary artery calcification (CAC) assessed using computed tomography (CT) scans is a clinically-validated biomarker that is highly prognostic for coronary heart disease (CHD) and adverse cardiac events. Clinical assessment of CAC relies on a dedicated coronary electrocardiogram (ECG)-synchronised non-contrast CT scan. However, millions of CT scans are acquired every year for various indications that include the heart in the field-of-view yet visible CAC is often not reported in these scans. This is a significant missed opportunity for incidental detection of a powerful cardiac risk factor. Our study was conducted on a set of 295 unselected, consecutive CT scans from the National Health Service (NHS) Golden Jubilee Hospital. These were annotated for CAC and used for model training and testing. We developed and validated a deep learning model to accurately quantify CAC on any CT scan including the heart, regardless of the presence or phase of contrast agent, reason for the scan, or use of ECG-synchronisation. The model achieved substantial agreement with the manual human assessment (Cohen’s Kappa=0.61, Bland-Altman mean difference=-40.8mm ³ ). Additionally, we found no correlation between arterial brightness (a surrogate metric for the level of contrast agent present) and agreement between manual and automated measurements (Spearman correlation R=-0.005). Early intervention is vital to improve patient outcomes. The automated CAC scoring method demonstrated here could be applied to all chest CT scans that include the heart, greatly expanding the opportunities for early detection of subclinical cardiovascular disease when preventative interventions have more impact. The promising accuracy achieved here by our deep learning model on a set of unselected sequential CT scans shows the potential for large-scale implementation to reduce the burden of coronary heart disease through systematic, opportunistic CAC screening.