The Performance Of Deep-learning-based Artificial Intelligence On Detection Of Pulmonary Nodules In Chest CT

First partThe Performance of Deep-learning-based Artificial Intelligence on Detection of Pulmonary Nodules in Chest CTObjectiveDeep learning based Artificial intelligence(AI)presents as the state of the art in the area of nodule detection,however,a validation with clinical data is necessary for further application.Therefore,the aim of this study is to evaluate the performance of AI in the detection of malignant and non-calcified nodules in chest CT.Materials and MethodsTwo hundred chest computed tomography(CT)data were randomly selected from a self-built nodule database from Tianjin Medical University General Hospital.The medical report was completed by a senior physician reviewing the initial report of a low-grade physician under daily work.The detection of non-calcified nodules in the medical report was recorded as a radiologist's nodule test.Both the pathology confirmed lung cancers and the nodules in the process of follow-up were included.Under the premise of the reference deep learning model and the radiologist's nodule test results,two experienced radiologists first read the chest CT image,combined with multiplanar reconstruction(MRP)and maximum density projection.(MIP),three-dimensional reconstruction and other technologies,the final result of two people as a true nodule,and record the location,size and density of the nodule.The McNemar test was used to determine whether there was a significant difference.The sensitivity,false negative rate and false positive rate were used to evaluate the nodule detection efficiency of artificial intelligence software and physicians.The McNemar test was used to determine whether there was a significant difference in sensitivity between the two.The Wilcoxon rank sum test was used to determine whether there was a significant difference in the false positive rates between the two.A P value of less than 0.05 indicates a statistical difference.Results1.The sensitivity of AI and the radiologist to detect malignant and benign nodules both were 100%.2.The sensitivity of AI total pulmonary nodules was significantly higher than that of radiologists(99.1%vs 43%,X2=483.20,P<0.001).Compared to the radiologists,the additional nodules detected by AI were solid and ground-glass density nodules with diameters<5 mm and 5 mm-10 mm.For subsolid nodules,the detection rate of AI was slightly higher,but there was no significant difference in the detection rate.The detection rate of AI was higher than that of radiologists in nodules at different locations.3.The false negative rate of AI was 0.90%,and 8 cases of nodules were missed,all of which were nodules with diameters<5 mm.The false negative rate of radiologists was 56.69%.4.A total of 504 nodules were missed,of which 346 were<5 mm in diameter.200 cases of chest multi-slice spiral CT included 889 non-calcified nodules,including 133 lung cancer nodules and 442 nodules less than 5 mm.The number of false positive nodules in AI was 993,and the false positive rate was 4.9 nodules per CT.If a false positive nodule with a diameter<5 mm is excluded,the false positive rate is reduced to 1.5 per case.The total number of false positives of radiologists was 3 nodules,which were related to vascular structure.The false positive rate was 0.015 CT per case.The false positive rate of AI was significantly higher than that of radiologists(4.9 vs 0.015),and the difference was statistically significant(P<0.05).Conclusion1.Artificial intelligence software based on deep learning can achieve no missed diagnosis of malignant pulmonary nodules.2.The deep learning based artificial intelligence software has better sensitivity than the radiologist,and can reduce the false positive rate after eliminating the small nodules.The second partThe Performance of two Deep-learning-based Artificial Intelligence on Detection of Pulmonary Nodules in Chest CTObjectiveTo compare the effects of two deep learning models in the detection of malignant and non-calcified nodules in thoracic CT.Materials and MethodsA total of 133 non-calcified pulmonary nodules were obtained from 128 patients with malignant nodules in the first part.One hundred twenty eight images were imported into two artificial intelligence software,and the detection of the nodules of the artificial intelligence software were recorded.The first software was based on 3D-convolutional neural network(3D-CNN).The second software was developed on the basis of the first one,adding the number of training sets and depth residual network(ResNet),fusing the models.The second model was the result of the first software optimization.The method of developing a gold standard true nodule was same as before,and the location,size and density of each nodule are recorded.The McNemar test was used to determine whether there was a significant difference in sensitivity between the two.The Wilcoxon rank sum test was used to determine whether there was a significant difference in the false positive rates between the two.A P value of less than 0.05 suggests a statistical difference.Results1.The sensitivity of the second software for detecting malignant nodules was slightly higher than that of the first software(100.00%vs 98.50%,P=0.500),but the difference was not statistically significant.2.The sensitivity of total lung nodule detection in second software was significantly higher than that in first software(99.18%vs 75.82%,X2=34.32,p<0.001).Compared with the first software,the additional nodules detected by the second software are nodules with diameter<5mm,5?10mm,solid and ground glass density nodules,non-calcified nodules connected with pleura,peripheral and central.For>10mm and some solid nodules,the detection rate of the second software was slightly higher,but there was no significant difference in detection rate.For the hilar nodules,both the first software and the second software can be detected.3.The false negative rate of the second software was 24.18%,and 148 cases of nodules were missed,of which 127 were<5 mm nodules,the false negative rate of the first software was 0.82%,and 5 cases of nodules were missed,all of which were<5 mm in diameter.4.One hundred twenty eight cases of chest CT included a total of 133 malignant nodules.The number of false positive nodules in the second software was 606,and the false positive rate was 3.0 nodules per CT.The number of false positive nodules in the first software was 216,and the false positive rate was 1.1 nodules per CT.The false positive rate of the second software was higher than that of the first software(3.0 vs 1.1),and the difference was statistically significant(P<0.05).Conclusion1.The sensitivity of two deep learning models have no significant difference on the detection of malignant nodule.2.The two deep learning models have different sensitivity and false positive detection rate on the detection of non-calcified nodules The optimized model increased both sensitivity and false positive rate for non-calcified nodules.