| Part 1 A Study on the Differentiation of Benign and Malignant Lung Nodules by Artificial Intelligence using Deep Learning MethodsObjectives:To investigate the value and advantage of the artificial intelligence lung nodules automatic system based on deep learning of DenseNet network in differentiating benign from malignant lung nodules,and compared with the artificial reading.Materials and methods:To collect 510 patients with chest of the First Affiliated Hospital of Soochow University January 2015 to December 2017.There were 160 benign lesions and 350 malignant lesions.The experiments were divided into an artificial reading group and an artificial intelligence group.The artificial reading group refers to four radiologists(those with less seniority who have worked on chest CT diagnosis for two years and senior physicians who have been engaged in chest CT diagnosis for 5 years)used double-blind methods to diagnose the benign and malignant results,mainly by these perspectives of pulmonary nodules size,density,morphology,margin and air bronchogram.In the artificial intelligence group,510 patients with pulmonary nodules were input into the DenseNet network system,and the nodules were extracted automatically through network learning,and the nodules were classified to obtain the benign and malignant diagnosis of pulmonary nodules.Data between the two groups were statistically analyzed by Chi-square test.Results:Benign and malignant lung nodules were assessed in 510 cases respectively in the primary and advanced physician group(AR group)and the artificial intelligence group(AI group).For the 160 cases of benign pulmonary nodules,the diagnostic accuracy in the AI group was significantly higher than that in the AR group(p=0.000<0.05),and the difference was statistically significant.There was no significant difference in diagnostic accuracy between the primary AR group and the advanced AR group(p=0.651>0.05).For the diagnosis of 350 cases of malignant pulmonary nodules,the accuracy rate in the advanced AR group was higher than that in the primary AR group.It had the highest accuracy in diagnosing malignant lung nodules in the AI group.Comparison between the primary AR group and the advanced AR group,and the primary AR group and the AI group showed statistically significant differences(p<0.05).The difference between in the advanced AR group and AI group was not statistically significant(p>0.05).On this basis,according to the diameter of pulmonary nodules,510 cases of pulmonary nodules were further sub-divided into three ranges,namely,diameter of?10mm,diameter of>10mm?20mm and diameter of>20mm.For lung nodules with different diameters,the diagnostic accuracy in the AI group was higher than that in the AR group(except for malignant lung nodules with diameter of>20mm,the diagnostic accuracy in the AI group was equivalent to that in the advanced AR group).However,the physicians group showed a certain degree of difference in the diagnosis of benign and malignant lung nodules within different diameters,that is,the accuracy rate of benign pulmonary nodules in the advanced AR group was lower than that in the primary AR group(85.92%vs.64.79%),and there was a statistical difference between the two groups(p<0.05).Within the range of diameter of>10mm<20mm,the diagnostic accuracy rate of benign pulmonary nodules in the advanced AR group was slightly higher than that in the primary AR group(51.67%vs.43.33%),with no significant statistical difference(p>0.05).However,within the range of>20mm in diameter,there was no statistical difference in the diagnostic accuracy between the primary and the senior group for malignant pulmonary nodules(p>0.05).Conclusions:The application of artificial intelligence(DenseNet network deep learning)can effectively identify benign and malignant lung nodules.The DenseNet network deep learning can achieve higher diagnostic accuracy,and the application of artificial intelligence can effectively help the imaging physician to make more accurate and reliable diagnosis of pulmonary nodules.On the whole,the senior physicians were more accurate in the diagnosis of benign and malignant pulmonary nodules than the primary physicians.Overall,the primary physicians had the lowest diagnostic accuracy rate for both benign and malignant lung nodules.The diagnostic accuracy of pulmonary nodules in different diameter ranges is different in the physician.In contrast,in the artificial intelligence group(DenseNet network deep learning),it can achieve good and stable diagnostic accuracy.Part 2Research on the Diagnosis of Specific Lesions Related to Pulmonary Nodules by Artificial Intelligence Using Deep LearningObjective:The application of the artificial intelligence(DenseNet network deep learning)to pulmonary nodules should not be limited to screening for benign and malignant nodules,but should be clearly diagnosed for specific lesions related to pulmonary nodules.In this study,selected pulmonary nodules were classified according to pathological results to further explore the effectiveness of artificial intelligence in the classification of specific lesions associated with pulmonary nodules.Materials and methods:To collect 510 patients with chest of the First Affiliated Hospital of Soochow University from January 2015 to December 2017.There were 160 benign lesions,which were divided into three groups according to pathological types:inflammatory group,pulmonary lymph nodes,and atypical adenomatoid hyperplasia.There were 350 cases of malignant lesions,which were divided into three groups according to pathological types:in situ adenocarcinoma,micro-invasive adenocarcinoma and invasive adenocarcinoma.The CT images of 510 patients with pulmonary nodules were input into the DenseNet network system,and the nodules were extracted automatically through network learning,and the nodules were classified and classified through the features to obtain a clear diagnosis of pulmonary nodules.In combination with the known pathological results,the accuracy rate of definite diagnosis of pulmonary nodules by artificial intelligence in each group was calculated,and the statistical analysis was conducted by chi-square test.Results:510 cases of pulmonary nodules were diagnosed by artificial intelligence(AI).For 160 cases of benign pulmonary nodules,there was no statistically significant difference in the diagnostic efficacy of AI in each group according to the pathological classification(p>0.05).For 350 cases of malignant lung nodules,the diagnostic accuracy rate of AI for micro-invasive adenocarcinoma was significantly higher than that for invasive adenocarcinoma(p=0.005<0.05),while there was no clear statistical difference in the diagnostic accuracy rate between orthotopic carcinoma and micro-invasive adenocarcinoma(p=0.669>0.05).Conclusion:Artificial intelligence(DenseNet network deep learning)CT pulmonary nodules diagnosis system can effectively diagnose specific pulmonary nodules.The diagnostic efficacy of various benign pulmonary nodules was similar,The diagnostic accuracy rate of micro-invasive adenocarcinoma was the highest among all types of specific lesions related to malignant lung nodules,and the diagnostic efficiency of in situ adenocarcinoma was similar to that of invasive adenocarcinoma. |
PDF Full Text Request