| Objective:Investigate the effectiveness of junior and senior physicians with the help of artificial intelligence(AI)software for the detection of pulmonary nodule,and evaluate the clinical value of AI software based on deep learning(3D convolutional neural network model)in the diagnosis of pulmonary nodules.Methods:206 patients with pulmonary nodules were randomly selected from the radiology department of Zhongshan Hospital Affiliated to Dalian University From January2019 to October 2019.The total number of pulmonary nodules was 1690,including 84males(40.8%),122 females(59.2%),and the average age is(51.5±14.8)years old.Two experienced radiologists(engaged in chest report review for more than 15 years)took the consistency of lung nodules detected by chest CT of 206 patients as the"gold standard",then two junior doctors and two senior residents tested the lung nodules on chest CT of 206patients independently,and recorded the maximum diameter,location,density and detection time of each cases of CT.After 14-day washout period,206 cases of pulmonary nodules on chest CT were detected again by the same low and high-aged doctors(two doctors each)assisted by AI software,and the relevant information of the target pulmonary nodules was marked and recorded(ibid.).All pulmonary nodules were further classified according to the diameter((27)3.0mm??3.0mm),location(subpleural nodule,paravascular nodule,other location nodule),density(solid,hyposolid,calcification).Compared the detection results of junior and senior doctors to the"gold standard"results,compared and analyzed the lung nodule detection efficiency of jonior and senior doctors by calculating the detection rate,false positive rate,and average detection time of each group.Results:1.General clinical data:A total of 1690“gold standard”lung nodules of 206 patients were obtained in this study.The average number of nodules in each CT was(8.2±5.0),and the diameter(d)was 2.0mm~12.4mm(3.6mm±1.2mm);980(58.0%)with d?3.0mm,710(42.0%)with d(27)3.0mm;879(52.0%)were solid nodules,594(35.2%)were subsolid nodules,217(12.8%)were calcified nodules;141(8.3%)were subpleural lung nodules,129 were paravascular nodules(7.6%),and 1420(84.1%)were the others.2.Pulmonary nodule detection results:(1)879 and 952 true positive pulmonary nodules were detected by junior and senior doctors independently,with a total detection rate of52.0%and 56.3%(?~2(28)6.351,(49)(27)0.05),the average detection time was 10.28±2.11min,6.42±2.26min((49)(27)0.01);the consistency of lung nodule detection by two radiologists in the group was good(Kappa=0.632,0.747);The detection rate of d(27)3.0mm,solid nodules and subpleural nodules were statistically significant((49)(27)0.05),but there was no statistical significance for other classifications((49)(29)0.05).(2)The total detection rate of junior physicians assisted by AI software was 81.7%,29.7%,and the average detection time was shortened from 10.28±2.11min to 5.43±2.79min((49)(27)0.01).Among them,the detection rate of d(27)3.0mm and subpleural nodules increased most significantly.(3)The total detection rate of senior physicians assisted by AI software was 82.3%,26.0%,and the average reading time was shortened from 6.42±2.26min to 2.83±0.97min((49)(27)0.01),Among them,the detection rate of d(27)3.0mm,solid nodules and paravascular pulmonary nodules were the most significant.(4)There was no significant difference in the overall detection rate of junior-and senior-age physicians after using AI((49)(29)0.05),The overall independent detection rate of senior doctors was significantly lower than that of junior doctors combined with AI(?~2(28)253.346,(49)(27)0.01).(5)The order of false-positive rate among the four groups was as follows:junior physicians independently(29)junior physicians combined with AI(29)senior physicians independently(29)senior physicians combined with AI,The false-positive rates were 17.6%,13.1%,7.8%,3.4%.Conclusion:1.AI software can significantly improve the overall detection efficiency of lung nodules of junior-and senior physicians,especially the junior physicians,making their pulmonary nodule detection ability exceed the ability of senior physicians independently.2.AI software can significantly improve the detection of subpleural and paravascular pulmonary nodules,as well as sub solid and solid nodules,especially the solid nodules.3.There is a certain false-positive rate in AI software.The screening and judgment of AI software is related to the working experience of the junior and senior doctors,The technology and algorithm of AI software also need to be improved constantly. |
PDF Full Text Request