| Part One Ability of 18F-FDG PET/CT radiomic features to distinguish solid lung adenocarcinoma from tuberculosisObjective:To develop a predictive model by 18F-FDG PET/CT radiomic features and to validate the predictive value of the model for distinguishing solitary lung adenocarcinoma from pulmonary tuberculosis.Methods:From January 2015 to October 2019,a total of 235 patients who underwent 18F-FDG PET/CT in The Fourth Hospital of Hebei Medical University or Hebei General Hospital and met the inclusion criteria and exclusion criteria were screened and collected the data retrospectively.All the patients were confirmed by surgical resection,image-guided biopsy pathology or follow-up.The enrolled patients were finally included 131 cases of solitary lung adenocarcinoma and 104 cases of solitary pulmonary tuberculosis,and all the enrolled patients were performed 18F-FDG PET/CT before any treatment or invasive operation.According to the ratio of 7:3,the enrolled patients with solitary lung adenocarcinoma or pulmonary tuberculosis were randomly divided into two sets:the training set and validation set,and the number of patients were 163 and 72,respectively.The PET/CT images with the format of DICOM were input to LIFEx 4.0 freeware which was applied to extract the radiomic features in the same volume of interest(VOI)on CT and PET images respectively.According to the VOI,LIFEx 4.0 freeware automatically processed and extracted 92 radiomic features for every enrolled patient,including 47 features based on PET and 45 ones originated from CT.Based on P<0.05 as significant difference,independent t-test,Mann-Whitney U test or chi-squared test in SPSS 21.0 were applied to screen the features from the 92 radiomic features and 3 clinical parameters between solitary lung adenocarcinoma and pulmonary tuberculosis in the training set.Continuous variables were analyzed by Mann-Whitney U test or independent t-test,and categorical variables were applied to chi-squared test.The clinical parameters include gender,age and the result of T cell spot test of tuberculosis infection(SPOT.TB).In order to avoid multicollinearity and over-fitting phenomenon,LASSO(Least absolute shrinkage and selection operator algorithm)was employed to further screen the optimal?for reducing the number of radiomic features.The optimal?represented the best radiomic features subset with the least number and the highest prediction value.Multivariate binary logistic regression was used to develop a formula according to the best radiomic features subset in the training set,and the score of radiomic features(Rad-score)was calculated based on the formula for every enrolled patient.The score of radiomic features(Rad-score)was performed to develop the radiomic model,and the clinical model was built by the clinical parameters which were the ones with statistical difference.The combination of the radiomic model and clinical model by multivariate binary logistic regression was applied to build the complex model.The performance of the three models to differentiate between solitary pulmonary tuberculosis and solid lung adenocarcinoma were assessed by Receiver operating characteristic(ROC)curve and AUC(the area under the ROC curve)in the training and validation set,respectively.Medcalc19.0 software was used to compare the AUC of different models.Nomogram was established,which could obtain the predictive probability of every patient.The apparent calibration curves were drawn by predicted probability based on model against actual probability related to solid lung adenocarcinoma,and the bias corrected curves were produced with 1000 bootstrap resamples.The goodness of fit was examined by the Hosmer-Lemeshow Test.The scanner comparability from the different PET/CT equipment was compared by Mann Whitney U test in SPSS 21.0,including the best radiomic features subset and the score of radiomic features(Rad-score)from the two PET/CT.It was statistically significant when the value of P was lower than 0.05 for the difference.Results:235 enrolled patients were randomly divided into training set and validation set,and the number of patients were 163 and 72,respectively.In the training set,there were 91 cases of lung adenocarcinoma and 72 cases of pulmonary tuberculosis,while the validation set included 40 cases with lung adenocarcinoma and 32 cases with pulmonary tuberculosis.Age and gender were no statistical differences not only in the training and validation sets(P=0.180 and 0.281)but also in the pulmonary tuberculosis and lung adenocarcinoma subset(P=0.621 and 0.505).T-SPOT.TB was significantly difference between pulmonary tuberculosis and lung adenocarcinoma subset both in the training(P<0.000)and the validation set(P=0.018).The positive rate of T-SPOT.TB in pulmonary tuberculosis subsets was much higher than that in lung adenocarcinoma subset in both the training and the validation set.However,The positive rate of T-SPOT.TB was no statistical differences between the training set and validation set(P=0.476).T-SPOT.TB was applied to build the clinical model by univariate logistic regression.Mann-Whitney U test was applied to screen the features from the 92radiomic features between the two diseases in the training set,and then eventually obtained 61 candidate-features.LASSO(Least absolute shrinkage and selection operator algorithm)was employed to further screen the optimal subset from the 61 candidate-features and 10-fold cross-validation was performed to select the best?in the LASSO.The 1standard error of?min(the min of?)was the optimal?,which can choose the optimal subset of radiomic features.The value of-3.093 was eventually chosen for the optimal?.According to the optimal?,61 radiomic candidate-features were decreased to9 ones with nonzero coefficients in the training set.The 9 best radiomic features were developed a formula by multivariate binary logistic regression in the training set to calculate the score of radiomic features(Rad-score).A radiomic model was built by the formula,and the score of radiomic features was calculated for every enrolled patient.T-SPOT.TB was applied to build the clinical model by univariate logistic regression.Clinical parameterswhich was selected index according to p less than 0.05 and Rad-score were combined together to build a complex model by multivariate binary logistic regression.For differentiating solitary pulmonary tuberculosis from solid lung adenocarcinoma,the AUC of radiomic model was 0.861 and that of clinical model was 0.686.The AUC of radiomic model was higher than that of clinical model(0.686)in the training set(P=0.000).The AUC of radiomic model in the validation set was 0.889 while the AUC of clinical model was 0.644.the AUC of radiomic model was higher than that of clinical model(P=0.000).The AUC of complex model was 0.889 and 0.909 in the training set and the validation set,respectively.And the AUC of complex model was the highest one of the three models either in the training or validation set.The differences between the complex model and the clinical model were significant in the training and the validation sets(both P<0.000)while the complex and the radiomic models were no significantly different in the training(P=0.118)and validation set(P=0.321).According to the good predictive ability of complex model to differentiate solitary pulmonary tuberculosis and solid lung adenocarcinoma,a nomogram that could demonstrate the proportion of every factor and predictive value of every patient was generated from the training set.The calibration curves were drawn and shown that the predicted probability agreed well with the actual probability in the training set and validation set.The goodness of fit examined by Hosmer-Lemeshow test were no significantly different,which demonstrated good fits for the predicted and actual probability both in the training set(?2=6.855,P=0.552)and validation set(?2=8.492,P=0.387).The 9 radiomic features and Rad-score stratified according to the scanners(PHILIPS or GE),and there were no significantly different in the training and validation set(All the P>0.05).In the training set,the AUC of radiomic model were 0.910 and 0.841 in the PHILIPS subset and the GE subset,respectively.The AUC of radiomic model were 0.961 and 0.862 in the PHILIPS subset and the GE subset in the validation set.Although the absolute value of AUC about the radiomic model on PHILIPS device was slightly higher than that of GE either in the training set or in the validation set,there were no statistical significance.The value of P was 0.411 in the training set and 0.461 in the validation set,respectively.Conclusion:Radiomic model based on 18F-FDG PET/CT represents important guiding value in differentiating solitary lung adenocarcinoma from tuberculosis.The radiomic model integrates the macro and micro features of multimodality fusion imaging,which provides important information for clinical diagnosis and identification of two kinds of diseases from different perspectives.Part Two Value of 18F-FDG PET/CT metabolic parameters in differen-tiating between solid lung adenocarcinoma and tuberculosis Objective:To establish a predictive model based on the conventional parameters of PET and CT in 18F-FDG PET/CT,and to evaluate the value of the model in differentiating solitary lung adenocarcinoma from pulmonary tuberculosis.Method:A total of 235 patients were enrolled in this study,who underwent 18F-FDG PET/CT and met the inclusion criteria and the exclusion criteria in The Fourth Hospital of Hebei Medical University or Hebei General Hospital from January 2015 to October 2019.All the enrolled were confirmed by surgical resection,image-guided biopsy pathology or follow-up,which included 131 patients with solitary lung adenocarcinoma and 104 cases of solid pulmonary tuberculosis.All the enrolled patients were performed the18F-FDG PET/CT examination before any treatment resulting from the two diseases or invasive operation.The data of 18F-FDG PET/CT were imported into the Life X 4.0 free software system,which was applied to extract the conventional parameters based on PET and CT,respectively.The parameters from PET included the min of standardized uptake value(SUV min),the max of standardized uptake value(SUV max),the mean of standardized uptake value(SUV mean),the standard deviation of standardized uptake value(SUV std),total lesion glycolysis(TLG)and metabolic tumor volume(MTV),while the parameters based on CT were the min of CT value(CT min),the max of CT value(CT max),the mean of CT value(CT mean),the standard deviation of CT value(CT std).Two sample independent t test or Mann Whitney U test in SPSS 21.0 was used to compare the conventional parameters based on PET and CT,and the parameters with statistical difference between solitary lung adenocarcinoma and pulmonary tuberculosis were found by the above statistical methods.When the continuous variables were completely in accord with homogeneity of variance and normal distribution,two sample independent t test was used,and the Mann Whitney U test was applied when the parameter did not conform to any of the above items.The conventional parameters with statistical difference from the PET and CT were considered to be independent variables.And the parameters based on PET and CT were used to establish PET model,CT model and the combined model of PET and CT named as PET/CT model by multivariate binary logistic regression with the method of forward respectively.Receiver operating characteristic(ROC)carves of PET model,CT model and PET/CT model were drawn by SPSS21.0.The area under the ROC curves were calculated and compared to evaluate the predictive ability of the three models.Medcalc19.0 software was used to compare the AUC of different models.It was statistically significant when the value of P was lower than 0.05 for the difference.Result:In term of gender and age,there was no significant difference,and the value of P was 0.505 and 0.620 respectively.Univariate analysis showed that the PET parameters including SUV min,SUV peak,TLG,MTV between solitary lung adenocarcinoma and pulmonary tuberculosis were no significant difference(all the P>0.05),while the CT parameters consisted of CT min and CT mean were not statistically significant(all the P>0.05).In term of SUV max,SUV mean,SUV std which were PET parameters and CT max,CT std from CT parameters had significant difference(P<0.05).Among the parameters of PET,the SUV max,SUV mean and SUV std in the solitary pulmonary tuberculosis group was 5.610,3.112 and 0.801 respectively,and those parameters were 7.470,4.214,1.241 in the lung adenocarcinoma group.The value of SUV max,SUV mean and SUV std in the solitary pulmonary tuberculosis group were significantly lower than those parameters in the lung adenocarcinoma group and the value of P was 0.034?0.014?0.001 respectively.As for CT related parameters,CT max and CT std in the solitary pulmonary tuberculosis group were 163.101 and 236.423,while those of 72.903 and207.641 in the lung adenocarcinoma group.The value of P about CT max and CT std was 0.000 and 0.014 respectively.Multivariate binary logistic regression showed that SUV max,SUV std,CT max and CT std were statistically significant in differentiating solitary pulmonary tuberculosis from lung adenocarcinoma(P<0.05),while SUV mean was not significant in differentiating solitary pulmonary tuberculosis from lung adenocarcinoma(P=0.857).According to the value of odds ratio(OR)about the four significant parameters,the order from strong to weak was SUV max,CT std,CT max,SUV std,and the value was 2.299,1.008,1.003,0.002 respectively.SUV max,CT std,CT max,SUV std were built the PET/CT model by Multivariate binary logistic regression.The area under ROC curve of the PET/CT model was 0.816,and the 95%confidence interval was from 0.762 to 0.869.The area under ROC curve of CT model was lower than that of PET/CT model,and there was statistically significant in differentiating solitary pulmonary tuberculosis from lung adenocarcinoma(the value of P was 0.005).There were no significantly different between the PET model and CT model about the area under the ROC curve,the value of P was 0.4265.The area under ROC curve of PET/CT model was higher than that of PET model,and there was statistically significant(the value of P was 0.027).The goodness of fit according to PET/CT model,Nagelkerke R2were 0.388,and the Chi square value resulting from Hosmer-Lemeshow test was 5.938.SUV max and SUV std were built the PET model by binary logistic regression.The area under ROC curve of the PET model was 0.771,and the 95%confidence interval was from 0.712 to 0.830.The goodness of fit about PET model,Nagelkerke R2were 0.293,and the Chi square value of Hosmer-Lemeshow test was 10.370.CT std and CT max were established the CT model and the area under ROC curve was 0.739 with the 95%confidence interval from 0.675 to 0.803.The goodness of fit of CT model,Nagelkerke R2were 0.222,and the Chi square value from Hosmer-Lemeshow test was 3.754.According to the result of Hosmer-Lemeshow test(P>0.05),the three model fitted well.In term of SUV max,SUV std,CT max,CT std,there was no significant difference between PHILIPS group and GE group(P>0.05).The area under ROC curve of the PET/CT model,PET model and CT model in the group of PHILIPS was0.783,0.703,0.802,while those in the group of GE was 0.743,0.701 and0.795 respectively.And there was no significant difference between PHILIPS group and GE group(all the P>0.05).Conclusion:The prediction model originated from the convention parameters of 18F-FDG PET/CT is helpful to differentiate between solitary lung adenocarcinoma and pulmonary tuberculosis.The PET/CT model combined with PET parameters(SUV max,SUV std)and CT parameters(CT max,CT std)integrates all the advantages of the two imaging methods,and its predictive ability is significantly better than that of pure PET and CT model,which reflects the advantages of multimodality imaging.And it is good for the clinical diagnosis of pulmonary tuberculosis and lung adenocarcinoma with multimodal macroscopic parameters that can be recognized by naked eyes.Part Three Value of 18F-FDG PET/CT image texture parameters in dis-tinguishing solid lung adenocarcinoma from tuberculosisObjective:Solitary pulmonary tuberculosis is easily misdiagnosed as lung cancer,and it is very difficult to distinguish the two diseases when they have similar clinical and radiological manifestations.The purpose of this study was to analyze the texture parameters based on 18F-PET/CT from solitary pulmonary tuberculosis and solid lung adenocarcinoma in order to accurately differentiate them.Method:235 patients who met the inclusion criteria and the exclusion criteria were enrolled in this study.All of the enrolled patients underwent18F-FDG PET/CT in The Fourth Hospital of Hebei Medical University or Hebei General Hospital from January 2015 to October 2019.The number of cases with solid lung adenocarcinoma or pulmonary tuberculosis was 131 and104 respectively.All the enrolled patients were confirmed by surgical resection,image-guided biopsy pathology or follow-up and performed18F-FDG PET/CT scan before all the treatment or puncture biopsy.Patients with solitary lung adenocarcinoma and pulmonary tuberculosis were randomly divided into the training group and the verification group according to the ratio of 7:3.According to the texture parameters extracted by the training group,the prediction model is established by optimization,and the prediction ability and stability of the model established by the training group are verified in the verification group.82 texture parameters of 18F-FDG PET/CT were extracted by LIFEx 4.0 free software system.In the training group,two sample independent t test or Mann Whitney U test in SPSS 21.0 was used to select the parameters with statistical differences between solid lung adenocarcinoma subset and pulmonary tuberculosis subset from 82 extracted texture parameters.Artificial neural network(ANN)was applied to obtain the standard weight values of each texture parameter with statistical difference in order to further screen the best combination of texture parameters.60%standard weight value was the best cutoff value,and the parameters with standard weight value greater than 60%truncation value are the best combination of texture parameters.Multivariate binary logistic regression was used to establish three model according to the different combination of the best texture parameters.The best texture parameters from PET or CT were built PET model or CT model respectively,while the combination of PET and CT texture parameters was established PET/CT model by multivariate logistic regression.Receiver operating characteristic(ROC)curve was drawn in SPSS21.0 resulting from the training group and verification group respectively.The predictive ability of the three models were evaluated by the receiver operating characteristic curve and the area under the ROC curve(AUC).Medcalc19.0software was used to compare the AUC of different models.It was statistically significant when the value of P was lower than 0.05 for the difference.Result:235 cases were randomly divided into training group and validation group according to the ratio of 7:3,the number of the two group were 163 and 72 cases.There were 91 cases with lung adenocarcinoma and 72cases with pulmonary tuberculosis in the training group,while in the validation group the cases of lung adenocarcinoma and pulmonary tubercu-losis were 40 and 32 respectively.In term of gender and age,there were no significant difference between the training group and validation group,and the value of P were 0.180 and 0.281.There was also no significant difference between the cases with lung adenocarcinoma and pulmonary tuberculosis,the values of P were 0.621 and 0.505.In the training group,56 of the 82 texture parameters were statistically different by two sample independent t test and Mann Whitney U test between the patients with lung adenocarcinoma and the cases with tuberculosis.The artificial neural network(ANN)calculated the standard weight values of 56 texture parameters,and the best texture parameters whose standard weight values were more than 60%was selected to reduce the number of best parameters.Finally,9 texture parameters were extracted from the 56 texture parameters.The best texture parameters based on PET and CT were used to establish PET model,CT model and the combined model of PET and CT named as PET/CT model by multivariate binary logistic regression with the method of forward.In the training group,the area under the ROC curve of PET model was 0.822,and the 95%confidence interval was from 0.759 to 0.886.There were no significantly different between PET model and CT model,and the value of P were 0.8547.The area under ROC curve was higher than PET model and CT model,and the difference was statistically significant(P=0.011 and 0.017).The goodness of fit according to PET/CT model,Nagelkerke R2were 0.415,and the Chi square value resulting from Hosmer-Lemeshow test was 5.389(P=0.884).The area under ROC curve of the PET/CT model was 0.878,and the 95%confidence interval was from 0.823 to 0.932.The goodness of fit according to PET/CT model,Nagelkerke R2were 0.530,and the Chi square value resulting from Hosmer-Lemeshow test was 3.695(P=0.715).The area under ROC curve of the CT model was 0.815,and the 95%confidence interval was from 0.748 to0.882.The goodness of fit according to PET/CT model,Nagelkerke R2were0.374,and the Chi square value resulting from Hosmer-Lemeshow test was10.323(P=0.243).According to the result of Hosmer-Lemeshow test(P>0.05),the three model fitted well.In the validation group,the area under ROC curve of the PET,PET/CT,CT model were 0.792(0.690?0.894),0.822(0.728?0.915),0.658(0.529?0.787),and the goodness of fit Nagelkerke R2were and the Chi square value of Hosmer-Lemeshow were 0.400,0.375,0.117and 5.651,14.061,8.755,respectively.The value of P were 0.686,0.080,0.363.All the P were more than 0.05 which indicted the three model fitted well.The area under ROC curve of CT model was lower than that of PET/CT model,and there was statistically significant in differentiating solitary pulmonary tuberculosis from lung adenocarcinoma(the value of P was 0.010).There were no significantly different between the PET/CT model and PET model about the area under the ROC curve,the value of P was 0.178.The area under the ROC curve of PET model and that of CT model.There was no significantly different between the PET model and CT model about the area under the ROC curve,the value of P was 0.071.In term of the 9 best texture parameters,there was no significant difference between PHILIPS cases and GE cases(P>0.05).Conclusion:The texture parameter model of 18F-FDG PET/CT is of great value to differentiate between solitary lung adenocarcinoma and tuberculosis.The constructed 18F-FDG PET/CT texture parameters regression model is instrumental in more accurate identification of solitary lung adenocarcinoma and tuberculosis from the microscopic perspective. |
PDF Full Text Request