| Part Ⅰ:A comparative study of pulmonary nodule detection efficiency of artificial intelligence assisted diagnostic system and double reading filmObjective:To investigate the matching rate and clinical significance of pulmonary nodules detection by artificial intelligence assisted diagnostic system.Methods:The chest CT images of 586 patients with complete clinical data from January 2016 to November 2017 in our Hospital were collected.Compare the output results of 12Sigma intelligent diagnostic system for pulmonary nodules with the consistent results of double reading.analyze the sensitivity and positive predictive rate of pulmonary nodule detection.Results:A total of 2751 pulmonary nodules were found in 586 patients.The total number of nodules detected by 12Sigma pulmonary nodule intelligent diagnostic system was 3259.Among them,1850 were solid nodules,502 were mixed ground glass nodules,458 were calcified nodules,449 were pure ground glass nodules,519 were false positive,11 were false negative,the sensitivity was 99.6%(2740/2751),and the positive predictive rate was 84.1%(2740/3259),false positive rate 15.9%(519/3259),missed diagnosis rate 0.4%(11/2751);2559 cases were found by double reading,false negative 192 cases,sensitivity 93.0%(2559/2751),positive predictive rate is 100%(2559/2559),missed diagnosis rate 7.0%(192/2751).The missed diagnosis rate of each nodule in 12 Sigma pulmonary nodule intelligent diagnostic system wasrespectively:0.34%(5/1486)for solid nodules,0.81%(4/493)for mixed ground glass nodules,0.47%(2/425)for pure ground glass nodules;missed diagonosis rate of various nodules by Double Reading was as follows respectively:solid nodules were 4.37%(65/1486),mixed ground glass nodules were 6.09%(30/493)and pure ground glass nodules were 22.82%(97/425).Conclusion:The false positive rate of AI-assisted diagnosis system for pulmonary nodules is high,the false negative rate is low,and the missed diagnosis rate of pure ground glass nodules is low(2/425,0.47%);the false positive rate of double reading is low,the false negative rate is high,and the missed diagnosis rate of pure ground glass nodules is high(97/425,22.82%).With referrence to the "second comment" from the output of AI-assisted diagnosis system for pulmonary nodules,film-reading can effectively reduce the missed diagnosis rate of pulmonary nodules,especially those of pure ground glass nodules.The second part A preliminary study on value of artificial intelligence assisted diagnostic system for pulmonary nodules in the qualitative diagnosis of pulmonary nodulesObjective:To analyze the output of artificial intelligence assisted diagnosis system for pulmonary nodules and to explore the value of quantitative parameters in assisted diagnosis.Methods:588 cases CT image of pulmonary nodules with complete pathological data from January 2016 to November 2017 in our Hospital were collected.The detection results of 12 Sigma pulmonary nodule intelligent diagnosis system were analyzed.The diameter of the nodules,CT value and"malignant probability" of the system output were extracted to compare with the pathological results.The difference between groups were studied by variance analysis.Possible cut points for differential diagnosis were determined by the Receiver operating characteristic(ROC).Results:588 cases of pulmonary nodules,pathological results thereof:31 cases of squamous cell carcinoma,42 cases of granuloma,68 cases of inflammatory lesions,447 cases of adenocarcinoma;12 Sigma pulmonary nodule intelligent diagnosis system output:including calcified nodules 25 cases,78 cases of ground glass nodules,176 cases of mixed ground glass nodules,solid nodules 309 cases;pathological classification and nodule relationship:Chi-square test The results showed that there were significant differences in different pathological types of nodules(P<0.001).The proportion of various nodules in adenocarcinoma is:solid nodules 41.2%(184/447),including calcified nodules 4.7%(21/447),mixed ground glass nodules 37.4%(167/447),pure ground glass nodules 16.8%(75/447)respectively.Solid nodules were found only in squamous cell catcinomas,while Solid nodules were predominant in inflamatory lesions and granulomas.Comparison between groups:There were significant statistical differences(P<0.001)in diameter,CT value and malignant probability between benign group(inflammatory granuloma)and malignant group(squamous carcinoma adenocarcinoma).Using ROC curve analysis,the cut point of using diameter of nodules and CT value,the "probability of malignancy" in differentiating malignant from benign lesions were 11.58mm,-312.1HU,68%.The area under the curve(AUC)was 0.553,0.697,0.500,respectively.The sensitivity was 73.6%,83.6%,94.5%;the specificity was 43.1%,55.4%,8.7%;and the accuracy was 48.8%,60.7%,24.8%respectively.The subtype of adenocarcinoma:pre-invasive lesion,micro-invasive adenocarcinoma,invasive adenocarcinoma.The inter-group diameter,CT value,malignant probability all have statistical difference(P<0.05).ROC curve analysis:nodule diameter,CT value,"malignant probability" of invasive adenocarcinoma group and non-invasive group(pre-invasive lesions and micro-invasive adenocarcinoma)the best cut point was 12.24 mm,-363.2 HU,94%,corresponding ROC area under curve was 0.929,0.895,0.860;sensitivity was 87.6%,80.3%,82.8%,respectively.The heterosexual rates were 86.4%,89.3%,74.8%and accuracy were 87%,84.5%,79.0%,respectively.Conclusion:The accuracy of differentiating benign and malignant lesions from benign lesions was low in combination of three indexes of nodule diameter,CT value and "malignant probability" of pulmonary nodules outputed by 12 Sigma pulmonary nodule intelligent diagnosis system.The accuracy of differentiating invasive adenocarcinoma from non-invasive group(preinvasive lesion/microinfiltrated adenocarcinoma)was moderate,and it was of moderate value for assistant doctors in the diagnosis of adenocarcinoma and adenocarcinoma classification. |
PDF Full Text Request