人工智能对基层影像医生筛查肺结节的辅助价值

目的探讨基层影像医生结合人工智能辅助诊断系统对厚层CT中4 mm以上结节检出效能的差异。方法前瞻性收集2019年1月1日至1月31日在我院接受常规胸部CT检查的118例患者并进行层厚5 mm骨算法重建,由两位十年以上诊断经验的主治医师和一位十五年以上诊断经验的副主任医师借助人工智能软件(InferRead CT_Lung 6.0,Infervision,Beijing,推想科技)确定肺结节金标准。由基层医院两位主治医师对CT图像进行独立阅片,记录结节的数量、位置、长径和标记时间,两周后再次借助AI对同一批图像进行阅片,计算医生独立阅片(A组)及借助AI(B组)两种情况下的结节检测敏感度、假阳性率,同时比较两种情况标记时间。结果A组和B组对于4 mm以上肺结节检出总数分别为172和293个,其中真阳性结节数分别为112和171个,假阳性结节数分别为60和122个。B组肺结节的检测敏感度显著高于A组(P0.01)。A、B两组医生肺结节平均检测时间比较,差异有统计学意义(P0.05)。结论 AI对于肺结节检出具有较好的辅助能力,借助AI辅助诊断系统的基层影像医生对厚层图像4 mm以上肺结节的标记时间更短,结节诊断敏感度更高。

Research on the auxiliary value of artificial intelligence in screening pulmonary nodules by grassroots imaging physicians

Objective To investigate the difference of detection efficiency of grassroots imaging physicians combined with artificial intelligence(AI) aided diagnosis system for nodules above 4 mm in thick slice CT.Methods A total of 118 patients admitted to our hospital and treated with conventional chest CT examination from January 1 to January 31,2019 were prospectively collected and reconstructed with 5 mm thick bone algorithm.The pulmonary nodules gold standard was determined with the help of AI software(InferRead CT_Lung 6.0,Infervision,Beijing,Infervision Technology) by two attending physicians with more than 10 years of diagnosis experience and one deputy chief physician with more than 15 years of diagnosis experience.CT images were read independently by two attending physicians of grassroots hospitals,and the number,location,long diameter and marking time of nodules were recorded.Two weeks later,the same batch of images were read again with the aid of AI,and the sensitivity and false positive rate of nodule detection in two cases of independent reading by physicians(group A) and AI(group B) were calculated.The paired sample of t-test was used for comparison,and the time consumptions of the two cases were compared at the same time.Results In group A and group B,the total numbers of pulmonary nodules over 4 mm were 172 and 293 respectively,among which 112 and 171 were correct nodules and 60 and 122 were false positive nodules,respectively.The detection sensitivity of pulmonary nodules in group B was higher than that in group A(P<0.01).There was significant difference of the average time spent of imaging physicians between the group A and the group B(P<0.05).Conclusion AI has a good auxiliary ability to detect pulmonary nodules.The grassroots imaging physicians who use AI-aided diagnosis system can mark pulmonary nodules larger than 4 mm in thick layer images in a shorter time and have higher sensitivity in nodule diagnosis.