Coronary artery calcium score quantification using a deep-learning algorithm

AIM: To investigate the impact of a deep-learning algorithm on the quantification of coronary artery calcium score (CACS) and the stratification of cardiac risk. MATERIALS AND METHODS: Computed tomography data of 530 patients who underwent CACS scan were included retrospectively. The scoring (including Agatston, mass, and volume scores) was done manually. The deep-learning method was trained using data from 300 patients to calculate CACS based on the manual calculation. The automated method was validated on a set of data from 90 patients and subsequently tested on a new set of data from 140 patients against manual CACS. For the data from 140 patients that were used to analyse the accuracy of deep-learning algorithm, the total CACS obtained manually and by using the deep-learning algorithm was recorded. Agatston score categories and cardiac risk categorisation of the two methods were compared. RESULTS: No significant differences were found between the manually derived and deep-learning Agatston, mass, and volume scores. The Agatston score categories and cardiac risk stratification displayed excellent agreement between the two methods, with kappa = 0.77 (95% confidence interval [CI]=0.73–0.81); however, a 13% reclassification rate was observed. CONCLUSION: Deep-learning algorithm can provide reliable Agatston, mass, and volume scores and enables cardiac risk stratification.