AI对非门控胸部LDCT平扫冠状动脉钙化积分危险分层的预测价值

目的 探讨人工智能（AI）在非门控胸部低剂量CT（LDCT）平扫冠状动脉钙化积分（CACS）危险分层中的预测价值。 方法 回顾性收集接受冠状动脉CT血管成像（CCTA）检查及非门控LDCT平扫的病人152例（训练集与测试集比例为2∶1）,训练集为上海长征医院收集的102例病人;测试集为武汉同济医院收集的50例病人。分别由AI模型和2名中年资影像医师在所有病人影像上勾画钙化,获得CACS后进行CACS危险分层（低危、中危和高危）,使用手动标注非门控LDCT的训练集数据,构建非门控LDCT的CACS及其危险度分层的AI模型。将测试集数据导入AI模型进行验证,与心电门控CT平扫获得的标准CACS及其危险分层进行对比分析,分别记录放射科医师手动评估及AI模型自动评估测试集CACS所需时间。采用分类准确度、组内相关系数（ICC）、Kappa检验和Bland-Altman一致性分析评估AI模型的性能,并采用Wilcoxon符号秩检验比较AI模型与标准CACS危险分层间的差异。采用配对t 检验比较AI、影像医师评估CACS危险分层所需时间。 结果 在训练集和测试集中,标准CACS的中位数分别为165.89（36.04,425.76）、96.50（25.75,346.75）,AI模型测得CACS的中位数分别为167.07（43.17,449.11）、75.51（24.30,250.74）,两者一致性均较好[ICC分别为0.977（0.965, 0.984）、0.989（0.980, 0.994）]。在测试集中进行Bland-Altman一致性分析,结果显示AI模型评估的CACS与标准CACS差值在95%一致性界限内的病例有48例,界限外的只有2例。在训练集和测试集中,AI模型预测的CACS危险度分层与标准CACS危险度分层的一致性均较好（κ值分别为0.895、0.899,均P<0.001）。AI模型预测训练集CACS危险分层的分类准确度为97.1%,其中对高危、中危、低危的分类准确度分别为96.9%、95.1%、100%。AI模型预测测试集CACS危险分层的分类准确度为94.0%,其中对高危、中危、低危的分类准确度分别为100.00%、82.40%、100.00%。AI模型预测测试集CACS危险分层与标准CACS危险分层的差异无统计学意义（Z=2.00,P=0.564）。采用AI模型评估不同CACS危险分层所需时间均较放射科医师少（P<0.001）。 结论 AI模型能够较为准确地评估LDCT平扫的CACS及其危险分层,明显提高工作效率,具有一定的临床应用价值。

Prediction of coronary calcification score risk stratification on non-gated chest low-dose CT with artificial intelligence

Objective To predict the risk stratification of coronary calcification score(CACS) risk stratification on non-gated low-dose computed tomography (LDCT) scan with artificial intelligence. Method All of 152 patients who had undergone coronary CT angiography (CCTA) examination and non-gated LDCT scans were retrospectively collected (ratio of training set to test set was 2∶1), the training set including 102 patients was collected from Shanghai Changzheng Hospital. The test set including 50 patients was collected from Wuhan Tongji Hospital. AI model and 2 well-experienced radiologists outlined calcifications on all patients’ images. After obtaining CACS, the CACS risk stratification (low, intermediate, and high risk) was performed, and the training set data of manually labeled non-gated LDCT was used to construct non-gated LDCT CACS and AI model of risk stratification, then applied the AI model to the test set data for verification.We compared the results with the standard CACS and risk stratification obtained by ECG-gated CT plain scan, and recorded the times required for the radiologist’s manual evaluation and the automatical AI model in evaluating the CACS of the test set. The classification accuracy, intra-group correlation coefficient (ICC), Kappa test and Bland-Altman consistency analysis were used to evaluate the performance of the AI model, and the Wilcoxon signed-rank test was used to compare the difference between the AI model and the standard CACS risk stratification. The paired t-test was used to compare the time required for AI and radiologists to assess CACS risk stratification. Results In the training set and test set, the median standard CACS were 165.89 (interquartile range 36.04-425.76) and 96.50 (interquartile range 25.75-346.75), and the CACS measured by the AI model were 167.07 (interquartile range 43.17-449.11) and 75.51 (interquartile range 24.30-250.74), respectively. The consistencies were good [ICC was 0.977 (95%CI: 0.965-0.984), 0.989 (95%CI: 0.980-0.994), respectively]. Bland-Altman consistency analysis showed that the difference between the CACS evaluated by the AI model and the standard CACS was within the 95% consistency limit of 48 cases, and only 2 cases were outside the limitin the test set. In the training set and test set, the CACS risk stratifications predicted by the AI model were in good agreement with the standard CACS risk stratification (κ values were 0.895 and 0.899, respectively, and both P<0.001). The AI model achieved 97.1% classification accuracy of CACS risk stratification in the training set,among which the classification accuracy of high-risk, intermediate-risk, and low-risk were 96.9%, 95.1%, and 100%, respectively. The AI model achieved 94.0% classification accuracy of CACS risk stratification in the test set, among which the classification accuracy of high-risk, intermediate-risk, and low-risk were 100.00%, 82.40%, and 100.00%, respectively. There was no statistical difference in CACS risk stratification between the AI model and the standard methodin the test set (Z=2.00, P=0.564). The AI model used less time to assess CACS risk stratification than radiologists(P<0.001). Conclusions The AI model could accurately evaluate the CACS of non-gated LDCT scans and its risk stratification, significantly improve work efficiency, and has certain clinical application value.