| Part 1:The value of CT Radiomics in the differentiation of solid lung metastases and benign nodulesObjectiveBy establishing a clinical-traditional radiological model,a Rad-score radiometric model,and a combined model of the former two,and comparing the diagnostic efficacy among the three,we explored the value of radiomics in the differentiation of solid lung metastases and benign nodules.MethodsA retrospective analysis was conducted on the data of 290 patients(519 lung nodules)in our hospital,of which 312 were lung metastasis(LM)and 207 were benign lung nodules(BN).Randomly split the dataset into training and validation cohorts at a ratio of 7:3.(1)Chi-square analysis,independent sample t-test,or Mann-Whitney U-test were used to statistically analyze general data(including clinical data and traditional CT imaging signs of lesions).The training cohort used univariate logistic regression analysis to screen statistically significant general data factors for constructing a clinical-traditional radiological model.(2)Two imaging diagnostic physicians independently draw regions of interest,use the Python software package PyRadiomics for feature extraction,use intraclass correlation coefficient(ICC),Mann Whitney U test,least absolute shrinkage and selection operator(LASSO)for feature screening and dimensionality reduction,and use extreme gradient elevation classifier to construct a Rad-score radiomics model using the obtained key dimensionality reduction features.(3)Combine the general information in the training cohort with the extracted radiomics features,and based on the ICC>0.75 standard,eliminate the radiomics features with poor consistency.After data preprocessing,use the T-test to filter out statistically significant differences.After using the min-max standardization method to standardize the above verified feature data,select the LASSO algorithm to filter and reduce the dimension of the data,and calculate the Rad-score of each lesion,Establish a joint model using an extreme gradient lifting classifier.(4)Validate the above three models in a validation cohort.The predictive effectiveness of the three models in the training and validation cohorts was evaluated using the area under the ROC curve,and the corresponding area under the ROC curve(including the calculation of 95%confidence interval),accuracy,sensitivity,specificity,accuracy,positive predictive value,negative predictive value,and F1 score were calculated.The goodness of fit between the diagnostic results of different models and the actual results was evaluated using a correction curve and a Hosmer Lemeshow test.The decision curve analysis(DCA)method was used to evaluate the net income level of different models under different threshold probabilities in the training cohort and the validation cohort.The difference in AUC values between different models was compared using the Delong test.(5)A nomograph prediction model is established based on the model with the best efficiency among the three models.P<0.05 considered the difference to be statistically significant.ResultsIn clinical data and CT imaging indicators,there were statistically significant differences between LM and BN groups in longest diameter,lobulation sign,shape,and cavity sign(P<0.05).A clinical-traditional radiological model was constructed by incorporating the above factors,with morphological irregularity being a protective factor for predicting LM,and the rest being risk factors for LM.After screening and dimensionality reduction,the remaining 49 features of CT imaging were used to construct an imaging omics model.The joint model was established based on 21 key features after screening and dimensionality reduction,including 4 general clinical traditional imaging factors and 17 radiomics features.In the clinical-traditional radiological model,the area under the ROC curve(AUC)in the training cohort and the validation cohort were 1.00(95%CI,0.93-1.00)and 0.89(95%CI,0.87-0.94),respectively.The AUC of the pure imaging omics model training cohort and the validation cohort constructed by Rad-score were 0.99(95%CI,0.93 to 1.00)and 0.95(95%CI,0.91 to 0.98),respectively.A combined model consisting of 17 radiomics features,clinical(previous history of malignant tumors),and CT signs(cavity sign,lobulation sign,and shape)was included.The AUC of the training cohort and the validation cohort were 1.00(95%CI,0.95-1.00)and 0.98(95%CI,0.97-1.00),respectively.The DeLong test showed that there were statistical differences among the AUC of the three models.Conclusions1.Previous history of malignant tumors can be a clinical factor predicting LM,and the longest diameter,lobulation sign,shape and cavity sign can be used as predictive CT signs.2.The Rad-score radiomics model has good predictive value for lung metastases.3.Among the three prediction models,the joint model has the highest prediction efficiency.Part 2:Preliminary Study on CT Radiomics for differential diagnosis of lung adenocarcinoma and non-adenocarcinoma metastatic tumorsObjectiveBy establishing a traditional-radiological model,a Rad-score radiometric model,and a combined model of the former two,and comparing the diagnostic efficacy among the three,this study explores the value of CT Radiomics in distinguishing lung adenocarcinoma from non-adenocarcinoma metastatic tumors.MethodsA retrospective analysis was conducted on the data of 61 lung cancer patients(205 metastatic nodules)in our hospital,of which 120 were lung adenocarcinoma(LUAD)metastatic nodules and 85 were non-adenocarcinoma(NLUAD)metastatic nodules.Split the dataset into training queues and validation queues at a ratio of 7:3.(1)Chi-square analysis,independent sample t-test,or Mann-Whitney U-test were used to statistically analyze general data(including clinical data and traditional CT radiological signs of the lesion).The training cohort used univariate logistic regression analysis to screen statistically significant general data factors for constructing a clinical-traditional radiological model.(2)Two imaging diagnostic physicians independently draw regions of interest,use the Python software package PyRadiomics for feature extraction,use intraclass correlation coefficient(ICC),Mann Whitney U test,minimum absolute shrinkage and selection algorithm(LASSO)for feature screening and dimensionality reduction,and use a random forest classifier to construct a Rad-score radiomics model using the key dimensionality reduction features obtained.(3)Combine the general information in the training cohort with the extracted radiological characteristics,and based on the ICC>0.75 standard,eliminate the radiological characteristics with poor consistency.After data preprocessing,use T-test to screen out statistically significant differences.After using the min-max standardization method to standardize the above verified feature data,select the LASSO algorithm to filter and reduce the dimension of the data,and calculate the Rad-score of each lesion,A joint model is established using a random forest classifier.(4)Validate the above three models in a validation cohort.The predictive effectiveness of the three models in the training and validation cohorts was evaluated using the area under the ROC curve,and the corresponding area under the ROC curve(including the calculation of 95%confidence interval),accuracy,sensitivity,specificity,accuracy,positive predictive value,negative predictive value,and F1 score were calculated.The goodness of fit between the diagnostic results of different models and the actual results was evaluated using a correction curve and a Hosmer-Lemeshow est.The decision curve analysis method(DCA)was used to evaluate the net income level of different models under different threshold probabilities in the training cohort and the validation cohort.The difference in AUC values between different models was compared using the Delong test.(5)A nomograph prediction model is established based on the model with the best efficiency among the three models.P<0.05 considered the difference to be statistically significant.ResultsIn clinical data and CT radiologic signs,the differences between NLUAD and LUAD metastatic tumor groups were statistically significant(P<0.05).The above factors were included in a multivariate logistic regression analysis to construct a traditional radiologic model,in which the longest diameter was a risk factor for predicting non-adenocarcinoma metastatic tumors,and the lobulation sign and spiculation sign were protective factors.After screening and dimensionality reduction,the remaining 22 CT radiomics features were used to construct a radiological model.The joint model was established based on 26 key features after screening and dimensionality reduction,including 2 traditional radiologic signs(longest diameter,spiculation sign)and 24 radiologic features.The area under the ROC curve(AUC)was 1.00(95%CI,0.93-1.00)for the training cohort and 0.66(95%CI,0.63-0.74)for the validation cohort in traditional radiological models,respectively.The AUC of the training cohort and the validation cohort in the Rad-score radiomics model constructed with pure radiomics features were 0.97(95%CI,0.95-1.00)and 0.75(95%CI,0.71-0.83),respectively.In the joint model,the AUC of the training cohort and the validation cohort were 0.93(95%CI,0.91-1.00)and 0.79(95%CI,0.75-0.86),respectively.The DeLong test showed that there were statistical differences among the AUCs of the three models.Conclusions1.The longest diameter,lobulation sign,and spiculation sign can be used as CT signs to distinguish between lung adenocarcinoma and non-adenocarcinoma metastatic tumors.2.Rad-score radiomics model has certain value in predicting the source of lung metastases(lung adenocarcinoma and non-adenocarcinoma).3.Among the three prediction models,the joint model has the highest prediction efficiency. |
PDF Full Text Request