Accuracy Of Artificial Intelligence In The Diagnosis Of Lung Tumors

Objective: To evaluate the accuracy of artificial intelligence(AI)in the diagnosis of lung tumors by using the method of systematic review / Meta-analysis to comprehensively search the published literature.Methods: The foreign language databases such as Pub Med/MEDLINE,Embase(through OVID),Cochrane Library and IEEE were searched by computer,and the Chinese databases such as CNKI,Wan Fang and VIP were comprehensively searched for the reports on the diagnostic accuracy of AI in lung tumors published from 1946 to December 2020.The literatures were screened according to the inclusion and exclusion criteria,and the quality of the included literatures was evaluated by QUADAS-2 quality evaluation chart.2 × 2 diagnostic data were extracted,and the complete data were calculated by Review Manager 5.3.Stata SE 15.0 software and Meta Di Sc 1.4 were used to analyze whether there was heterogeneity among the studies,and to find out the causes of heterogeneity.The Sensitivity(Sen),Specificity(Spe),Positive Likelihood Ratio(+LR),Negative Likelihood Ratio(-LR),Diagnostic Odds Ratio(DOR)and Diagnostic Score(DS)were calculated by combining the data with Stata SE 15.0.And drawing the Forest Map and the Summary Receiver Operating Characteristic Curve(SROC),to calculate the Area Under the Curve(AUC),the Fagan Diagram,the Likelihood Ratio Point Map.Results: A total of 42 sets of data were included for the study of artificial intelligence to distinguish between benign and malignant of Ground Glass Opacity(GGO)of the lung.The results showed that the combined value of Sen was 0.90(95%CI: 0.89,0.92),the combined value of Spe was 9.0(95%CI:0.88,0.92),the +LR merge value was9.0(95%CI: 7.2,11.3),the-LR merge value was 0.11(95%CI: 0.09,0.13),DOR was83(95%CI: 60,116)and AUC was 0.96(95%CI: 0.93,0.97).The results suggest that AI model is more accurate in the diagnosis of solid pulmonary nodules.The Fagan diagram shows that the prior probability is 50%,and the post-test probability is 93%when the AI model is positive for solid pulmonary nodules,and 8% when the results are negative.The results of likelihood ratio dot map show that the ability of excluding diagnosis by using artificial intelligence model to detect GGO in clinic is stronger.To summarize the accuracy and specificity of AI model in the diagnosis of lung tumors,the accuracy of diagnosing solid pulmonary nodules(0.90)was higher than that of diagnosing pulmonary GGO(0.92),and the specificity(0.90)was higher than that of GGO(0.93).Conclusion: The artificial intelligence model has high accuracy in the diagnosis of lung tumors and has high clinical diagnostic value.The accuracy of artificial intelligence model in the diagnosis of solid pulmonary nodules is higher than that of ground glasso pacity.The methods of label marking,image preprocessing and feature learning will affect the accuracy of artificial intelligence model in the diagnosis of lung tumors,while the selection of learning image database and verification database usually does not affect the accuracy of the model.