Clinical Application Of Artificial Intelligence Assisted Diagnosis System For Pulmonary Nodules

Part I:Evaluation of the Ability of Artificial Intelligence Assisted Diagnosis System to Detect Small Pulmonary NodulesObjective:To evaluate the ability of artificial intelligence assisted diagnosis system to detect pulmonary nodule;to compare the sensitivity and false positive rate of artificial intelligence auxiliary diagnosis system in detecting pulmonary nodules in chest CT images with thickness of 5mm and 1mm.Methods: Collected from January 2019 to March 2019 all in our hospital during the chest CT scanning,and at the same time with artificial intelligence assisted diagnosis system for 5mm and 1mm layer thickness prediction of patients a total of 8493 people,the subjects were 1000 people who met the experimental conditions,the number of pulmonary nodules jointly determined by at least one attending physician and at least two doctors combining images of 5mm layer thickness and 1mm layer thickness as the "gold standard",the sensitivity,positive predictive value,misdiagnosis rate and missed diagnosis rate of the artificial intelligence assisted diagnosis system for detecting small pulmonary nodules were analyzed,The sensitivity and false positive rate of pulmonary nodules detection in chest CT images with thickness of 5mm and 1mm were compared between the artificial intelligence assisted diagnosis system for pulmonary nodules.Results: In 1000 lung CT examinations,A total of 1,527 pulmonary nodules were detected by the "gold standard".A total of 4087 small pulmonary nodules were identified by the artificial intelligence assisted diagnosis system recognizing 5mm layer thickness images,including 2583 false positive nodules and 23 false negative ones.The sensitivity,positive predictive value,misdiagnosis rate and missed diagnosis rate were 98.5%,36.8%,63.2% and 1.5% respectively for the detection of pulmonary nodules by the artificial intelligence assisted diagnosis system in 5mm layer thickness images.The artificial intelligence assisted diagnosis system recognized 5mm layer thickness images,and the sensitivity of detection of nodules with d ≤3mm,3mm< d ≤6mm and 6mm< d≤10mm were 97.1%,99.5% and 100%,respectively;The sensitivity of solid nodules,mixed grinding glass nodules and pure grinding glass nodules were 98.9%,100% and 97.8%,respectively.artificial intelligence assisted diagnosis system in 1mm layer thickness images detection lung nodules a total of 6400,including false positives nodules in 4884,false negatives nodules for 11;Artificial intelligence assisted diagnosis system identification,1 mm thick reconstruction image detection of lung nodules sensitivity,positive predictive value,the misdiagnosis rate and missed diagnosis were 99.3%,23.7%,76.3% and0.7% respectively;The sensitivity of the artificial intelligence assisted diagnosis system for detecting 1mm layer thickness images was 98.9%,95.5%,100% and99.2%,100% and 99.4% respectively for nodules with d≤3mm,3mm<d≤6mm,6mm<d≤10mm,solid nodules,mixed grinding glass nodules and pure grinding glass nodules.Statistical analysis results showed that: the sensitivity of the reconstructed thickness image to the detection of pulmonary nodules was higher than that of the5 mm thickness image(99.3%>98.5%),there is statistical significance difference(χ~2=4.283 P=0.038);the false positive rate of 1mm layer thickness image was higher than 5mm layer thickness image(76.3%>63.2%),there is statistical significance difference(χ~2=209.136 P﹤0.01).Based on the clinical significance of nodules with diameter ≤3mm is small,after removing nodules with diameter ≤3mm,the detection rate of small pulmonary nodules with 5mm layer thickness is consistent with that with 1mm reconstructed layer thickness image.Conclusion:1.The artificial intelligence assisted diagnosis system has a high overall sensitivity to the detection of small pulmonary nodules and is more suitable for the screening of lung health examination in normal population.2.The artificial intelligence assisted diagnosis system to 1 mm,5 mm layer thickness image lung nodules prediction results show that: 1 mm layer thickness on the detection of lung nodules is higher than 5 mm layer thickness,but 1 mm layer thickness false positive rate is higher.3.Based on the clinical significance of nodules with diameter ≤3mm is small,after removing nodules with diameter ≤3mm,there was no significant difference in the detection rate of pulmonary nodules with 5mm layer thickness and 1mm reconstruction layer thickness.PartⅡ:Clinical Application of Artificial Intelligence in Diagnosing Benign and Malignant Pulmonary NodulesObjective: To analyze the value of artificial intelligence auxiliary diagnosis system in differentiating benign and malignant pulmonary nodules.Methods: Analysis of our hospital between January 2019 and June 2020 complete pathological data of 104 pulmonary nodules chest CT images,taking pathology as the "gold standard",they were divided into benign and malignant groups,contrast analysis of lung nodules of artificial intelligence assisted diagnosis system and image physician of lung nodules of benign and malignant diagnosis sensitivity,specificity,accuracy,positive predictive value and negative predictive value,to evaluate the clinical value of artificial intelligence assisted diagnosis system in the qualitative diagnosis of pulmonary nodules.Results: A total of 104 pulmonary nodules with pathological results in 104 cases were collected,the benign nodules 25,79 malignant nodules.The detection rate of benign and malignant pulmonary nodules by AI assisted diagnosis system for pulmonary nodules was 100%,the accuracy,sensitivity,specificity,positive predictive value and negative predictive value of pulmonary nodules were 76.0%,91.1%,28.0%,80.0% and 50.0%,respectively,104 pulmonary nodules were detected and diagnosed by radiologists,and the total detection rate was 100%,the accuracy,sensitivity,specificity,positive predictive value and negative predictive value of pulmonary nodules were 90.3%,96.2%,64.0%,89.4% and 84.2%,respectively.The two diagnostic methods were compared with the pathological results,the results showed:two types of diagnosis for lung nodules detection rate of 100%,the accuracy,sensitivity,specificity,positive predictive value and negative predictive value of the diagnosis of pulmonary nodules by radiologist were all higher than that of the AI assisted diagnosis system for pulmonary nodules,the sensitivity of radiologists in diagnosing malignant pulmonary nodules was higher than artificial intelligence assisted diagnosis system for pulmonary nodules(96.2% > 91.1%),there was no statistically significant difference comparing(P=0.754>0.05),the specificity of the former was higher than the latter(64.0%>28.0%),there is statistical significance difference(P=0.027<0.05),and the area under the ROC curve were 0.596 and 0.801.Conclusion: The sensitivity of artificial intelligence diagnostic system for pulmonary nodules to qualitative diagnosis of malignant nodules was similar to that of radiologists;but compared with radiologists,the specificity of the AI diagnostic system for pulmonary nodules is low,the reference value for differential diagnosis of benign nodules is poor,and the system needs to be further improved.