| Objective: To retrospectively analyze the benign and malignant nature of pulmonary nodules on chest CT by using artificial intelligence deep learning algorithm.Trying to establish an auxiliary,intelligent aided diagnosis model in qualitative diagnose of pulmonary nodules for the clinic.Methods: Retrospectively collected nine hundred and ninty-seven cases of pulmonary nodules detected on chest CT from three hospitals in Dalian(mainly from Affiliated Zhongshan Hospital of Dalian University)from January 2013 to December 2017.The total number of pulmonary nodules was 1177.All pulmonary nodules were confirmed by surgical pathology or clinical treatment,763(64.8%)were malignant nodules and 414(35.2%)were benign nodules.There were 410 males(41.1%),587 females(58.9%),an age range from 23 to 91 years old,with average age of(61.3±10.7)years old.First,two target nodules were marked by two trained resident physicians,957 nodules were randomly selected as training groups,and the remaining 220 nodules were defined as test groups.Then a trained model is established by using a combination of convolutional neural network and long-term memory type recurrent neural network.In order to improve the generalization ability of the model,data augmentation was performed in the training group.Finally,the trained model was used in the test group,to test the accuracy,sensitivity,and specificity of the model in differential diagnosis between benign and malignant pulmonary nodules detected by the CT,and the corresponding area under the ROC curve(AUC)was calculated.In addition,the senior radiologist who was blind to all the nodules pathology results diagnosis(perform a benign and malignant diagnosis)all the other 220 nodules in the test group.SPSS 20.0 statistical software was used for statistical analysis and processing.The ROC curve was used to analyze the effectiveness of artificial intelligence model and radiologist in the prediction and diagnosis of benign and malignant pulmonary nodules.The AUC represents the diagnostic efficacy;the test level was set as ?=0.05.The difference was considered statistically significant at P<0.05.Results: In this study,the feature extraction of all 1177 pulmonary nodules was performed by using a deep learning algorithm,and a 3D U-Net model based on ResNet was established and improved.In the 140 malignant lung nodules of the test group,the model successfully predicted that 130 pulmonary nodules were malignant,accounting for 92.9%,and 10 were misdiagnosed as benign,accounting for 7.1%;in 80 benign pulmonary nodules,the model was successful,56 pulmonary nodules were predicted to be benign,accounting for 70.0%;24 were misdiagnosed as malignant,accounting for 30.0%.If the malignant lung nodule was a positive standard,the accuracy of the model was 84.5%,the precision was 84.4%,the corresponding sensitivity and specificity were 92.9% and 70.0%,the false positive rate was 30.0%,and the recall was 92.9.%,FScore is 0.88.The corresponding true positive rate and false positive rate were obtained by adjusting the discriminating threshold parameters,and plotted as ROC curves.The area under the curve was calculated and the AUC value of the classifier was 0.84.When the threshold of benign and malignant pulmonary nodules was determined to be 0.52,the diagnostic rate was highest,reaching 84.5%.In this study,senior radiologist performed a differential diagnosis of 220 pulmonary nodules in the test group.Of the 140 malignant pulmonary nodules,118 was correctly diagnosed as malignant,accounting for 84.3%,and 22 were misdiagnosed as benign nodules,which accounted for 15.7%.In the 80 benign lung nodules,60 nodules were accurately diagnosed as benign pulmonary nodules,accounting for 75.0%;20 nodules were misdiagnosed as malignant nodules,accounting for 25.0%.Similarly,malignant lung nodules were positive criteria.The diagnostic accuracy of radiologists was 81.0%,precision was 85.5%,corresponding sensitivity and specificity were 84.3%,75.0%,false positive rate was 25.0%,recall rate.84.3%,F-Score was 0.85.Conclusion:1.This study confirmed that the learning algorithm of artificial intelligence based on depth,forecast by 3D U-Net model improved pulmonary nodules on larger sample size for the differential diagnosis of benign and malignant is feasible and effective,and has a high diagnostic accuracy and sensitivity.2.This study shows that the prediction accuracy rate was 84.5% in diagnosis benign and malignant pulmonary nodules,especially in malignant lung nodules diagnosis.The prediction accuracy is up to 92.9%,which is higher than the radiologist of ten years work experience,is expected to be an effective aided diagnosis system in the future prediction and differential diagnosis of benign and malignant lung nodules. |
PDF Full Text Request