| Objective:This study aims to explore the application value of artificial intelligence(AI,artificial intelligence)system in the screening and identification of pulmonary nodules.There are three main research objectives:1.Compare the detection efficiency of AI system and physician scan for all nodules and different types of nodules,and analyze the missed diagnosis and misjudgement of the two methods.2.Evaluate the accuracy of AI system in evaluating the malignant risk of nodules.3.Quantitatively analyze the characteristic parameters of the nodules that AI system extract,find the critical value of the characteristic parameters,and evaluate their value in the identification of benign and malignant pulmonary nodules.Methods:Pulmonary nodules were retrospectively included from the PACS system(picture archiving and communication system),and all enrolled cases were imported into the LinkDoc system for screening and the results were recorded.The results of physician dual-reading mode screening were subject to the image report which archived in the imaging diagnosis system.1.All true nodules were classified according to different sizes、types and locations,and the sensitivity,misjudgment rate,missed diagnosis rate and accuracy of the overall and three different classification criteria of nodules by AI system and physician screening were compared respectively.The reasons for missed diagnosis and misjudgement by the two methods were analyzed.2.The nodules with pathological results were classified as benign and malignant,and the malignant risk degree given by the AI system was recorded.ROC curve(Receiver Operating characteristic Curve)was used for analysis.3.The nodule characteristic parameters that extracted by AI system were recorded.The parameters were firstly analyzed by factor analysis,and then the diagnostic value of the parameters was analyzed by ROC curve.For the valuable parameters,binary Logistic regression analysis was used to find the critical value of benign and malignant.Results:A total of 174 cases were included in this study,869 true nodules were determined by gold standard,and 180 with pathological results.1.The AI system has a sensitivity of 90.2%,a missed diagnosis rate of 9.8%,a misjudgement rate of 2.8/per cases and an accuracy of 58.0%for the overall nodules,and a sensitivity of 55.7%,a missed diagnosis rate of 44.3%,a misjudgement rate of 0.11/per case and an accuracy of 54.4%for the double-read mode.For solid nodules,mGGN(mixed ground-glass nodules)and pGGN(pure ground-glass nodules),sensitivity of AI system respectively was 95.0%,96.7%and 76.4%,and sensitivity of double-read mode was 70.0%,56.7%and 30.1%.For subpleural、paravascular,、and solitary pulmonary nodules,AI system sensitivity respectively was 89.9%,87.3%,and 92.6%,misjudgement rates were 2.8/per case,3.8/per case,and 1.6/per case;doublereading mode sensitivity respectively was 64.7%,37.7%,and 61.1%,misjudgement rates were 0.17/per case,0.06/per case,and 0.10/per case.For nodules less than 5mm,5 to 8mm,and greater than 8mm,AI sensitivity was 91.1%,82.5%,and 96.7%respectively,and double-read mode sensitivity was 30.0%,70.4%,and 93.7%.All the above differences were statistically significant(P<0.05).2.The 180 nodules with pathological results were divided into benign group and malignant group.The risk score of malignancy which was assessed by AI system as the independent variable,while the actual malignancy of the nodules was used as the dependent variable,then use ROC curve,AUC=0.687,P<0.001,the maximum value of yoden index was 0.299,corresponding to malignant risk of 73.7%,sensitivity of 48.8%,specificity of 81.1%.3.Record the maximum diameter,area,volume,CT value and nodule type which extracted by AI system,conduct KMO test and Bartlett’s spherical test for the five characteristic parameters,KMO coefficient 0.525,P<0.001,so they are not suitable for factor analysis.ROC curves were performed for the five parameters,and the AUC values were maximum diameter 0.683,area 0.819,volume 0.665,and CT value 0.712,excluding the type parameters.The critical values respectively were 8.9mm,87.5mm2,1287.55mm3,-477.15Hu,differences were statistically significant,P≤0.001.Binary Logistic regression was performed for the four quantized parameters.The independent risk factors of malignancy were area and CT value,P<0.05,and the prediction accuracy of the model was 74.4%.Conclusion:1.The AI system has auxiliary detection value for the screening of all nodules、nodules of different types、locations and sizes.The sensitivity is higher than that of doctors’ manual reading,and the number of misjudgement is significantly higher than that of doctors’ manual reading.Misjudgement of AI system and missed diagnosis of physicians are mainly micro-nodules less than 3mm.The working mode of " AI initial screening+physician exclusion of false positive nodules" is advocated.2.The AI system has a reference value for the assessment of pulmonary nodules’malignant risk,which can identify a small number of malignant nodules and exclude a small number of benign nodules.3.Among the characteristic parameters extracted by AI system.CT value and largest area are independent risk factors of malignant nodules,which have high diagnostic value.And the maximum diameter and volume have moderate diagnostic value.Each parameter has a specific critical value,which can help to distinguish the benign and malignant nodules. |
PDF Full Text Request