基于人工智能诊断系统的胸部CT肺结节检出率及良恶性诊断研究

The detection rate and diagnosis of benign and malignant pulmonary nodules on chest CT based on artificial intelligence diagnostic system

目的探讨基于人工智能(artificial intelligence,AI)诊断系统的胸部CT肺结节检出率及良恶性诊断效能。方法收集2021年2月至2022年3月于金华市人民医院行胸部CT扫描且同时接受肺结节AI诊断系统检查的97例患者的临床资料,由2名高级职称影像诊断医生共同制定“金标准”并比较人工检出和AI诊断系统的真阳性结节检出率、假阳性结节检出率、假阴性结节漏检率、敏感度与阳性预测值。结果AI诊断系统共检出结节532个,其中实性结节302个,混合磨玻璃结节81个,含钙化结节75个,纯磨玻璃结节74个,假阳性结节79个,假阴性结节2个,敏感度99.6%,阳性预测值84.4%;人工诊断检出结节419个,假阴性结节32个,敏感度92.9%,阳性预测值100%。人工诊断的敏感度显著低于AI诊断系统(χ2=27.508,P<0.001),阳性预测值显著高于AI诊断系统(χ2=71.621,P<0.001)。AI诊断系统各结节漏诊率分别为实性结节0.41%、混合磨玻璃结节1.25%;人工各结节漏诊率分别为实性结节4.51%、混合磨玻璃结节7.50%、纯磨玻璃结节21.43%;AI诊断系统实性结节、混合磨玻璃结节、纯磨玻璃结节的漏诊率均显著低于人工诊断(P<0.05)。AI诊断系统结果与病理诊断符合率为72.28%,与病理结果一致性的Kappa值为0.430;人工诊断与病理诊断符合率为90.02%,与病理结果一致性的Kappa值为0.791;人工诊断符合率显著高于AI诊断系统(P<0.05)。结论基于AI诊断系统的胸部CT肺结节检出假阳性率较高,假阴性率低;在结节的定性诊断上与人工诊断还有一定差距,可作为人工诊断的参考。

Objective To investigated detection rate and diagnosis of benign and malignant pulmonary nodules on chest CT based on artificial intelligence(AI)diagnostic system.Methods From February 2021 to March 2022,97 patients who underwent chest CT scan in Jinhua People’s Hospital and were examined by AI diagnostic system for pulmonary nodules were collected.The“gold standard”was jointly formulated by two senior diagnostic imaging doctors,and the detection rate of true positive nodules,false positive nodules,false negative nodules missed detection rate,sensitivity and positive predictive value were compared between manual detection system and AI diagnosis system.Results The AI diagnostic system detected a total of 532 nodules,included 302 solid nodules,81 mixed ground glass nodules,75 calcified nodules,74 pure ground glass nodules,and 79 false positive nodules,2 false negative nodules,with a sensitivity of 99.6%and a positive predictive rate of 84.4%.419 nodules and 32 false negative nodules were detected by manual diagnosis,with a sensitivity of 92.9%and a positive predictive value of 100%.The sensitivity of manual diagnosis was significantly lower than that of AI diagnosis system(χ2=27.508,P<0.001),and the positive predictive value was significantly higher than that of AI diagnosis system(χ2=71.621,P<0.001).The missed diagnosis rates of nodules in AI diagnosis system were 0.41%for solid nodules and 1.25%for mixed ground glass nodules.The missed diagnosis rates of nodules in manual diagnosis were 4.51%for solid nodules,7.50%for mixed ground glass nodules,and 21.43%for pure ground glass nodules.The missed diagnosis rate of solid nodules,mixed ground glass nodules and pure ground glass nodules by AI diagnosis system was significantly lower than that by manual diagnosis(P<0.05).The coincidence rate between the results of the AI diagnostic system and the pathological diagnosis was 72.28%,and the Kappa value consistent with the pathological results was 0.430;the coincidence rate between the artificial diagnosis system and the pathological diagnosis was 90.02%,and the Kappa value consistent with the pathological results was 0.791.The coincidence rate of manual diagnosis was significantly higher than that of AI diagnosis system(P<0.05).Conclusion The false positive rate of pulmonary nodules detected by chest CT based on AI diagnosis system is high,and the false negative rate is low.There is a certain gap between the qualitative diagnosis of nodules and manual diagnosis,which can be used as a reference for manual diagnosis.