The Application Value And Influencing Factors Of Coronary Artery Calcification Score Algorithm Based On Artificial Intelligence In Non-gated Chest CT

Objective : To explore the application value of artificial intelligence software in the measurement of coronary artery calcification score(CACS)of non-gated chest CT and its influencing factors.Methods:A total of 237 patients who underwent both coronary computed tomography angiograph(CCTA)and chest CT within one month were included in this study retrospectively.ECG-gated cardiac CT calcification was delineated by radiologists with different experiences(Low/High,and the two groups were labeled as gated-L and gated-H),Then the ECG-gated Agatson scores were obtained by CACS software(Heartbeat CS,v3.5.0.2254,Philips Healthcare,Netherland),take the gated-H CACS as the gold standard.The artificial intelligence(AI)software(CACScore Doc,v170.170.051acuda110,Shu Kun Technology,Beijing)automatically obtain the non-gated Agatson scores,and the results were checked and reedited by radiologist with low experience(the two groups were labeled as AI and AI+L).these Agatson scores and the time taken for each case were recorded.The correlation and bias of gated-L,AI,AI+L and gated-H(gold standard)CACS were compared by Spearman correlation analysis and Bland-Altman consistency analysis.Kappa test was used to evaluate the consistency of CACS risk categorization among AI,AI+L,gated-H and gated-L.The chi-square test was used to evaluate the accuracy of risk categorization among gated-L,AI,AI+L and gated-H.The time required for CACS measurement in AI,AI+L,gated-H and gated-L was compared by paired sample wilcoxon test.The Mann-Whitney U test was used to compare the heart rate difference between the cases classified correctly and the cases reclassified.The independent sample t test was used to compare the BMI difference between the cases classified correctly and the cases reclassified.And the chi-square test was used to evaluate the accuracy of CACS risk categorization among patients with different heart rates and different BMI in the AI.a P<0.05 was considered statistically significant.Results : The post-processing and calculating time on each case of AI was 42.0s(37.0～46.0s),reduced by 39%,68% compared with gated-H,gated-L(all P<0.01),respectively.The checking and reediting times on each case of AI+L was 60.0s(51.5～68.0s),reduced by 15%,55% compared with gated-H,gated-L(all P<0.01),respectively.The correlation of the gated-L,AI,AI+L-CACS with the gated-H CACS was r = 0.99,0.65,0.98(all P<0.01).The Bland-Altman plot(gated-H CACS minus gated-L,AI,AI+L-CACS)showed a mean difference of 29.5,299.8,102.1,and 95% limits of agreement of-310.7～369.6,-807.4 to 1405.0,-351.3 to 555.6.Kappa analysis showed that the CACS risk categorization consistency in gated-L and AI+L was better than that in AI,and the Kappa coefficients(κ)were 0.73,0.89 and 0.36,respectively,P<0.01.The accuracy of CACS risk categorization in AI+L was 92.0%,which was significantly better than that in gated-L and AI(80.2% and 52.3%,respectively,P<0.01).Heart rate has a significant effect on the accuracy of CACS risk categorization in AI software.patients with heart rate ≥ 80 bpm are twice as likely to be misclassified as patients with heart rate< 60 beats / min(66.7% vs 34.3%).Body mass index(BMI)has no significant effect.Conclusion : AI non-gated chest CT CACS software greatly shortens the time for radiologists to measure CACS,and it is more likely to underestimate CACS by using AI merely.While the AI+doctor reading model can significantly improve the evaluating performance and the accuracy of risk categorization,which is more suitable for clinical risk grade screening of patients with coronary heart disease.In addition,high heart rate will have a negative impact on the accuracy of CACS risk categorization by AI software,which provides a basis for software optimization in the future..