Automatic coronary calcium scoring in chest CT using a deep neural network in direct comparison with non-contrast cardiac CT: A validation study

Abstract

To evaluate deep-learning based calcium quantification on Chest CT scans compared with manual evaluation, and to enable interpretation in terms of the traditional Agatston score on dedicated Cardiac CT. Methods: Automated calcium quantification was performed using a combination of deep-learning convolution neural networks with a ResNet-architecture for image features and a fully connected neural network for spatial coordinate features. Calcifications were identified automatically, after which the algorithm automatically excluded all non-coronary calcifications using coronary probability maps and aortic segmentation. The algorithm was first trained on cardiac-CTs and refined on non-triggered chest-CTs. This study used on 95 patients (cohort 1), who underwent both dedicated calcium scoring and chest-CT acquisitions using the Agatston score as reference standard and 168 patients (cohort 2) who underwent chest-CT only using qualitative expert assessment for external validation. Results from the deep-learning model were compared to Agatston-scores(cardiac-CTs) and manually determined calcium volumes(chest-CTs) and risk classifications. Results: In cohort 1, the Agatston score and AI determined calcium volume shows high correlation with a correlation coefficient of 0.921(p < 0.001) and R 2 of 0.91. According to the Agatston categories, a total of 67(70 %) were correctly classified with a sensitivity of 91 % and specificity of 92 % in detecting presence of coronary calcifications. Manual determined calcium volume on chest-CT showed excellent correlation with the AI volumes with a correlation coefficient of 0.923(p < 0.001) and R 2 of 0.96, no significant difference was found (p = 0.247). According to qualitative risk classifications in cohort 2, 138(82 %) cases were correctly classified with a k-coefficient of 0.74, representing good agreement. All wrongly classified scans (30(18 %)) were attributed to an adjacent category. Conclusion: Artificial intelligence based calcium quantification on chest-CTs shows good correlation compared to reference standards. Fully automating this process may reduce evaluation time and potentially optimize clinical calcium scoring without additional acquisitions.

Automatic coronary artery calcium scoring on routine chest computed tomography (CT): comparison of a deep learning algorithm and a dedicated calcium scoring CT

“…As for cardiac risk classification, our deep learning algorithm accurately classified 81% of internal and 79% of external cases, and the Kappa values were 0.82 and 0.80, respectively, which were higher than those of Cano-Espinosa et al (11). The miscategorization rate was also in line with that of van Assen et al (16) and Wolterink et al (5). Most misclassified cases were within one category, with only 1% of cases showing more than one category in internal scans.…”

“…The sensitivity of our deep learning algorithm was higher than in previous studies, reducing the rate of missed diagnosis. The diagnostic accuracy was slightly higher than that of van Assen et al (16), who reported an accuracy of 90%. The quantification of the CAC score on 3-mm-slice thickness CT was almost perfect in internal scans (ICC =0.94) and good in external validation (ICC =0.83).…”

“…Results showed that both 1 and 3 mm scans achieved good agreement with the gold standard, and chest CT with a slice thickness of 3 mm was slightly more accurate in CAC detection, with a sensitivity of 99% and a diagnostic accuracy of 94%. Previous studies have reported a sensitivity of 82% to 94% in detecting CAC (9)(10)(11)16,17). The sensitivity of our deep learning algorithm was higher than in previous studies, reducing the rate of missed diagnosis.…”

“…From the perspective of automatic scoring using deep learning-based algorithms, few studies have adopted dedicated calcium scoring CT as a reference. van Assen et al (16) examined 95 paired scans within 1.5 years and Eng et al (17) examined 447 paired scans, with a maximum time interval of 2 years. CAC may show progression, and the accuracy of the deep learning algorithm may be affected.…”

“…Additionally, the reference standard for most of the aforementioned studies was a manual result on chest CT, which differs from that measured on a dedicated calcium scoring CT of the same patient, as motion artifacts are more likely to occur on chest CT images due to the lack of ECG control (15). Recently, several studies investigated automatic deep learning algorithms using a dedicated calcium scoring CT as a reference, but the sample sizes were relatively small (16), the interval between the paired scans was long (17), and the influence of slice thickness on the algorithm was not compared.…”