A Comparative Study Of Computer-aided Diagnosis System Based On Deep Learning In Differential Diagnosis Of Benign And Malignant Pulmonary Nodules

Objective:To evaluate the effectiveness of computer-aided diagnosis system based on deep learning in the diagnosis of benign and malignant pulmonary nodules and analyze the causes of misdiagnosis.Methods from October 2018 to February 2020,324 cases of pulmonary nodules found by chest CT and confirmed by operation and pathology in the storage system(PACS)of radiology department of our hospital were analyzed retrospectively.There were 244 malignant nodules and80 benign nodules.Log in to the computer-aided diagnosis system(Yitu＿AICare＿CT＿Chest,Yitu Company,China,referred to as "AI system"),all nodules can be detected and extracted one by one,and the size,density,location,imaging features and final benign and malignant diagnosis of each nodule can be recorded,and the final benign and malignant diagnosis will be compared with the pathological report.The accuracy,sensitivity,specificity,false positive rate and false negative rate of AI in differential diagnosis of benign and malignant pulmonary nodules were evaluated.Statistics of relevant clinical information and imaging signs data,using SPSS23.0 software for statistical comparison.Result:1.Among the pulmonary nodules confirmed by operation and pathology,the AI of all nodules were detected and extracted one by one.There were 324 non-calcified pulmonary nodules,205 malignant nodules and119 benign nodules.The diagnostic accuracy,sensitivity,specificity,false positive rate and false negative rate were 0.645,0.684,0.525,0.185 and 0.647 respectively.There was significant difference in the size of lesions between benign group and malignant group(Upright12060.5,Personality 4.654 and malignant 0.000).The nodule in benign group was smaller than that in malignant group(P < 0.05),and the size of nodule in benign group was smaller than that in malignant group.There was significant difference in the composition of sex between benign and malignant groups(P < 0.05).There were significant differences in density,lobulation sign,spiculation sign,pleural indentation sign and vacuole sign between benign and malignant groups,and the ROC curve of lesion size showed that AUC was 0.673.The best cut-off value is 9.5mm,the sensitivity is 73%,and the specificity is 56.2%.There were 127 nodules less than or equal to10 mm,87 malignant nodules and 40 benign nodules.19 nodules were diagnosed as malignant by AI,of which 10 nodules were true positive.The sensitivity and specificity were 0.115 and 0.775 respectively.2.The diagnostic sensitivity and false negative rate of AI were relatively high.Statistical analysis showed that there was significant difference in age between false negative group and true positive group.In patients less than 40 years old,AI determined 24 benign nodules and13 malignant nodules,while actually 3 nodules were benign and 34 nodules were malignant.In the comparison of benign and malignant nodules,the incidence of malignant nodules in females is higher than that in males,but in the comparison between false negative and true positive nodules,the result is just the opposite.in false negative nodules,the proportion of females is higher than that of males.There was a significant difference in lesion size between the false negative group and the true positive group(Uspeak 489.5,Zhumile 11.607,and0.000),and the size of the false negative group was smaller than that of the true positive group.There were significant differences in density,lobulation sign,spiculation sign,pleural indentation sign and vacuole sign between false negative group and true positive group(P < 0.05).Conclusion : AI system has 100% detection efficiency in the detection of pulmonary nodules.The diagnostic accuracy,sensitivity and specificity of lung cancer are not very ideal,but the misdiagnosis rate is low.Density,size,lobulation sign,spiculation sign,pleural indentation sign and vacuole sign are helpful for the differential diagnosis of benign and malignant tumors.The false negative rate is higher,which may be too dependent on clinical and imaging signs,and is more inclined to judge female nodules under 40 years old as benign,or less than 10 mm nodules as benign.