Evaluation Of Diagnostic Value Of Different AI Software For Pulmonary Nodules

Objective:To evaluate the detection efficiency of lung nodules based on three kinds of artificial intelligence software based on deep learning and the diagnostic value of CT quantitative indicators in benign and malignant thoracic nodules and ground glass nodules of lung adenocarcinoma.The research contents include:Part one: Evaluate the detection efficiency of two AI software for pulmonary nodules and the diagnostic value of different AI software combined for pulmonary nodules.Part two: To explore the diagnostic value of three AI software quantitative CT indicators in evaluating benign and malignant pulmonary nodules.Part three: To explore the diagnostic value of three AI software quantitative CT indicators in evaluating the invasion of ground-glass nodule lung adenocarcinoma.Part one1.Materials and methodsA total of 92 patients(483 nodules)with pulmonary nodules underwent physical examination in the Affiliated Hospital of Yan ’an University from June2021 to October 2021 were selected.The nodule detected by AI software was evaluated by imaging physicians and the number and type of nodule were recorded.Visual evaluation was performed manually by two senior radiographers and used as the gold standard for nodular identification.The detection rate,false positive rate and false negative rate of the two AI software were calculated,and the diagnostic value of the two AI software for nodules was evaluated.Chi-square tests and Fisher’s exact tests are applied to compare the differences between different software and the gold standard.Finally,the diagnostic value of the combined AI software for pulmonary nodules was evaluated.2.Results2.1 Distribution of different nodulesA total of 483 nodules were detected by chest CT in 92 patients,including295 solid nodules(61.1%),6 PSN(1.24%),68 m GGNS(14.1%)and 114 calcified nodules(23.6%).2.2 Comparison of 483 pulmonary nodules detected by two AI softwareThe detection rate of nodule in software A was 92.1%,and there was a statistical difference between software A and manual reading,and the consistency between software A and manual reading was general(Kappa value=0.213).The detection rate of nodules in software B was 87.0%,and there was a statistical difference between software B and manual scanning.The consistency between software B and manual scanning was weak(Kappa value=0.150).Software A had a higher false positive rate and software B had a higher false negative rate.2.3 Comparison of the detection of different types of nodules by two AI softwareThe detection rates of solid nodules and ground glass nodules were statistically different between the two kinds of software.There was A statistical difference between software A and manual scanning in the detection of solid nodules and calcified nodules.There was statistical difference between software B and manual reading in the detection of glass grinding nodules.2.4 Joint detection of pulmonary nodules by two AI softwareThe combination of the two AI software increased the detection rate of different types of pulmonary nodules,especially solid nodules and ground glass nodules.The combination of the two software made no statistical difference in the detection rate of different types of nodules compared with manual examination,and the combination of the two software showed no statistical difference in the detection number of nodules compared with manual examination.The consistency between the combination of the two software and manual examination was better than that of the single software(Kappa value=0.439).3.Brief SummaryThere are differences in the detection of pulmonary nodule types between the two AI software.Therefore,it is necessary to further optimize the algorithm of AI software,strengthen its homogeneity for nodule recognition,and improve the management of software data set,so as to facilitate clinical follow-up and decision-making of nodule.The combination of the two AI software diagnostic methods is a feasible optimization for accurate identification of early lung cancer.Part two1.Materials and methodsA retrospective analysis was performed on 135 isolated pulmonary nodules with pathological findings after surgical resection in Affiliated Hospital of Yan’an University from December 2019 to October 2022,including 104 cases of lung adenocarcinoma,31 cases of lung benign nodules(11 cases of tuberculous granulation tissue,3 cases of inflammatory pseudotumor,7 cases of hamartoma,7 cases of inflammatory nodules,and 1 case of Langerhans giant cell hyperplasia).1 case of sclerosing cell tumor of lung,1 case of neuroendocrine tumor).The original CT image data were imported into the lung nodule AI auxiliary software provided by three medical companies(A,B,C),and quantitative indicators(including average CT value,average diameter,maximum CT value,etc.)were recorded respectively.Clinical baseline data and CT quantitative indicators were tested for difference.The quantitative CT indicators with P < 0.05 were analyzed by univariate and multivariate binary logistic regression to find the sensitive indicators that could predict benign and malignant pulmonary nodules.Receiver operating characteristic(ROC)curve was drawn,and area underthecurve(AUC)was used to evaluate the diagnostic efficacy of different indicators on benign and malignant pulmonary nodules.2.Results2.1 General clinical data of benign and malignant nodulesThere were statistically significant differences in age and family tumor history between the benign and malignant nodules group(P < 0.05),and the mean age of the benign and malignant nodules group was(58.039±9.236)years and(52.455±11.987)years,respectively.There were no significant differences in gender,smoking history and BMI between the two groups(P>0.05).2.2 Logistic regression analysis and evaluation of diagnostic effectiveness of three AI software in differentiating benign and malignant nodulesLogistic regression analysis was performed for the quantitative indicators with significant differences between the groups,including the average CT value and the maximum CT value of software A;The average CT value of software B;The average CT value and average diameter of software C were significantly different(P<0.05).ROC curve analysis showed that the average CT value of software B had the highest diagnostic efficiency,with an AUC of0.746(0.661～0.881).3.Brief SummaryThis study found that some quantitative indicators of different AI software,including average CT value,maximum CT value and average diameter,had diagnostic value for the differentiation of benign and malignant nodules,among which the average CT value of software B had the best diagnostic efficacy.Part three1.Materials and methodsA retrospective analysis was performed on 66 GGN nodules patients confirmed by surgery and pathology as lung adenocarcinoma in the Affiliated Hospital of Yan ’an University from December 2020 to October 2022,including24 males(36.4%)and 42 females(63.6%),aged 25-76 years with a median age of 58 years.Surgery and pathology confirmed 17 cases of prodromal lesions(including 2 cases of atypical adenomatous hyperplasia,15 cases of in situ adenocarcinoma),25 cases of microinvasive adenocarcinoma,and 24 cases of invasive adenocarcinoma.The original CT image data were imported into the lung nodule AI auxiliary software provided by three medical companies(A,B,C),and quantitative indicators(including average CT value,average diameter,3D diameter,etc.)were recorded respectively.Clinical baseline data and CT quantitative indicators were tested for difference.Quantitative CT indicators with P < 0.05 were analyzed by univariate and multivariate binary logistic regression to find sensitive indicators that could predict the aggressiveness of ground-glass nodular lung adenocarcinoma.Receiver operating characteristic(ROC)curve was plotted,and area underthecurve(AUC)was used to evaluate the diagnostic efficacy of different indicators on pathologic invasiveness of lung adenocarcinoma.2.Results2.1 General clinical baseline data of patientsThe mean ages of preglandular disease,microinvasive adenocarcinoma group and invasive adenocarcinoma group were(58.133±8.467),(53.909±12.095)and(57.313±8.950)years,respectively.There were no significant differences in age,gender,BMI,smoking history and family history of tumor among the three groups(P > 0.05).2.2 Logistic regression analysis results of Software A in differentiating grinding glass nodular lung adenocarcinoma pathological classification3D long diameter and mean CT value were independent predictors of distinguishing glandular prodromal disease(AAH/AIS)from microinvasive adenocarcinoma(MIA).Mean CT value was an independent predictor of distinguishing between microinvasive adenocarcinoma(MIA)and invasive adenocarcinoma.2.3 Logistic regression analysis results of differentiating grinding glass nodular lung adenocarcinoma pathological classification by Software BMean diameter was an independent predictor of distinguishing glandular prodromal disease(AAH/AIS)from microinvasive adenocarcinoma(MIA);Mean CT value was an independent predictor of distinguishing between microinvasive adenocarcinoma(MIA)and invasive adenocarcinoma.2.4 Logistic regression analysis results of differentiating grinding glass nodular lung adenocarcinoma by Software CMean CT value was an independent predictor of distinguishing glandular prodromal disease(AAH/AIS)from microinvasive adenocarcinoma(MIA).Mean diameter was an independent predictor of distinguishing between the microinvasive adenocarcinoma group(MIA)and the invasive adenocarcinoma group.2.5 The invasive ROC results of ground-glass nodular lung adenocarcinoma were distinguished by quantitative indexes of different AI softwareThe 3D diameter of software A showed the highest diagnostic efficacy in differentiating glandular prodromal lesions from microinvasive adenocarcinoma,with an AUC of 0.805(0.577～1.000)and a critical value of 16.230 mm.The average CT values of the three software could distinguish glandular prodromal lesions from microinvasive lesions,and there was no difference in diagnostic efficacy.The average CT value of software A had the greatest diagnostic efficacy in differentiating microinvasive adenocarcinoma from invasive adenocarcinoma,with an AUC of 0.802(0.595～1.000)and a truncation value of-334.59 HU.The average diameter of software B/C could distinguish microinvasive adenocarcinoma from invasive adenocarcinoma,and there was no difference in diagnostic efficacy between the two.3.Brief summaryQuantitative CT indexes of different AI software have certain diagnostic value in differentiating the invasion degree of GGN lung adenocarcinoma,among which diameter indexes(3D long diameter,average diameter)and density indexes(average CT value)are more sensitive indexes.To provide some reference for the selection of clinical diagnosis and surgical treatment.Conclusions1)There were differences in the detection and homogeneity of pulmonary nodules with different AI softwares.Therefore,in the later stage,it is necessary to further optimize the algorithm of artificial intelligence software and improve the management of software data set.The joint diagnosis method of different AI software is a feasible optimization scheme for accurate identification of early lung cancer.2)Some quantitative indexes of AI software: average CT value,maximum CT value and average diameter have diagnostic value for differential diagnosis of benign and malignant nodules,and provide more sensitive indicators for clinical identification of benign and malignant nodules,among which the average CT value is the most valuable.3)The CT quantitative index of AI software has certain diagnostic value in differentiating the invasive degree of GGN lung adenocarcinoma,among which diameter index(3D long diameter,average diameter)and density index(average CT value)are more sensitive indexes.To provide a certain reference basis for clinical diagnosis and selection of surgical treatment.