基于深度学习的肺结节筛检和定性诊断分析

Pulmonary Nodule Screening and Qualitative Diagnosis Based on Deep Learning

目的探讨基于深度学习的人工智能(AI)在肺结节检测和定性诊断中的临床价值。方法收集行胸部CT平扫的250例患者。分为住院医(A组)、AI(B组)和住院医结合AI(C组)三组,比较三组对肺结节检出的误诊率、漏诊率、敏感度、阳性预测值和平均诊断时间。同时分别比较实性结节和磨玻璃结节(GGN)良恶性的AI量化参数,对有统计学差异的参数行ROC曲线分析。结果以两名高年资主任医师共同阅片结果为参照标准,确认有2230个结节。B组的误诊率明显高于A、C两组,阳性预测值明显小于A、C两组(P<0.05)。A组的漏诊率明显高于B、C两组,敏感度明显低于B、C两组(P<0.05)。B组平均诊断时间明显少于A、C两组(P<0.05)。实性良、恶性结节的长径、最大面积、体积、最小CT值和恶性概率差异均有统计学意义(P<0.05),ROC曲线下面积(AUC)大于0.7的参数为:长径、最大面积、体积、恶性概率。GGN良、恶性结节的长径、最大面积、体积、平均CT值、最大CT值和恶性概率差异均有统计学意义(P<0.05),对各参数行ROC曲线分析,AUC均大于0.7。结论AI协助阅片可明显提高工作效率和肺结节检出敏感度,并减少误诊率和漏诊率,同时AI对肺结节良恶性的预判具有一定参考价值。

Objective To explore the clinical application value of deep learning-based artificial intelligence(AI)in the detection and related quantitative measurement of pulmonary nodules.Methods We collected 250 cases of chest CT scan and compared the misdiagnosis rate,missed diagnosis rate,sensitivity,positive predictive value and average diagnosis time of pulmonary nodules among group A(hospitalized),group B(AI)and group C(hospitalized+AI).Meanwhile,AI quantization parameters of solid nodules and ground glass nodules(GGN)were compared,and ROC curve analysis was performed for the parameters with statistical difference.Results A total of 2230 nodules were identified.The misdiagnosis rate of group B was significantly higher than those of group A and C,and the positive predictive value was significantly lower than those of group A and C(P<0.05).The rate of missed diagnosis in group A was significantly higher than those in group B and C,and the sensitivity was significantly lower than those in group B and C(P<0.05).There were statistically significant differences in the long diameter,maximum area,volume,minimum CT value and malignant probability between solid benign and malignant nodules(P<0.05).The indexes of area under the ROC curve(AUC)greater than 0.7 were:long diameter,maximum area,volume and malignant probability.There were statistically significant differences in the length,maximum area,volume,average CT value,maximum CT value and malignant probability between GGN benign and malignant nodules(P<0.05).ROC curve analysis of all parameters showed that AUC was greater than 0.7.Conclusion AI-assisted film reading could significantly improve work efficiency and sensitivity of pulmonary nodules detection and reduce the rates of misdiagnosis and missed diagnosis.Meanwhile,it has certain reference value for the prediction of benign and malignant pulmonary nodules.