(P191) DEEP LEARNING ANALYSIS OF AORTIC VALVE AND CORONARY ARTERY CALCIUM SCORES FROM NON-GATED COMPUTED TOMOGRAPHY SCANS: COMPARISON WITH THE ECG-GATED REFERENCE STANDARD

Background: Coronary artery calcium (CAC) and aortic valve calcium (AVC) scores are clinically important measures of cardiovascular risk and aortic stenosis severity, respectively. Non-gated chest computed tomography (CT) scans are frequently performed and can be opportunistically used to assess CAC and AVC. This study aimed to measure CAC and AVC scores from non-gated scans using an automated deep learning (DL) pipeline, and validate these measurements against the reference standard from ECG-gated scans.

METHODS AND RESULTS: A retrospective cohort was assembled at the Royal Victoria Hospital and Jewish General Hospital between 2019-2023. Patients who underwent at least one ECG-gated cardiac CT scan with reported CAC and/or AVC scores, along with a non-gated chest CT scan (contrast-enhanced "C+" or non-contrast-enhanced "C-'') within the one year of each other, were included. Non-gated scans were downloaded and analyzed using a 3-dimensional deep learning convolutional neural network previously trained and validated in our laboratory to segment CAC and AVC.

Among 164 patients with ECG-gated CAC scores, 97 had non-gated C+ scans (mostly CT pulmonary embolism) and 79 had C- scans during the one year timeframe. Spearman's correlation analysis showed very good agreement between reference CAC scores and DL-CAC scores from non-gated C+ scans (R 0.566, p< 0.001) and C- scans (R 0.652, p< 0.001). ROC analysis showed a high discrimination to identify CAC scores ≥100 by DL from non-gated C+ scans (AUC 0.864, 95% CI 0.789 to 0.938) and C- scans (AUC 0.888, 95% CI 0.814 to 0.961). Among 321 patients with reported AVC scores, 202 had non-gated C+ scans and 147 had C- scans. Spearman's correlation analysis revealed very good agreement between reference AVC scores and DL-AVC scores from non-gated C+ scans (R 0.603, p< 0.001) and less so C- scans (R 0.221, p< 0.001). ROC analysis showed a high discrimination to predict AVC scores ≥1200 by DL from gated C+ scans (AUC 0.830, 95% CI 0.773 to 0.886) and less so C- scans (AUC 0.649, 95% CI 0.558 to 0.740).

Conclusion: CAC and AVC scores measured from non-gated CT scans using an automated DL model are well correlated with the reference standard, and for the most part, have high discrimination to identify patients exceeding clinically actionable thresholds. Further research is underway to improve the DL model's calibration and enrich the training set with diverse scans.