Study On The Prediction Of Tumor Mutation Burden Status In Non-small Cell Lung Cancer By Radiomics Models Based On CT Imaging

Objective: To explore the value of preoperative CT-based radiomics model in predicting the status of tumor mutation burden(TMB)in non-small cell lung cancer(NSCLC)by establishing radiomics model,CT features model,joint model,combined model and comparing their diagnostic efficiency.Methods: The preoperative CT and clinical data of 400 patients with lung cancer confirmed by pathology in the first affiliated Hospital of Guangzhou Medical University from February to September 2019 were collected retrospectively,who underwent lung CT scan and TMB tests.After screening of inclusion and exclusion criteria,163 cases with high TMB expression and low TMB expression in the ratio of1:2 were included,which were randomly divided into training group(n = 13)and verification group(n = 50)in the ratio of 7:3.Anong the clinical data of the subjects,age,gender,smoking history,clinical stage and preoperative CEA were included;and for the imaging signs of CT,lung lobe where the tumor was located,the maximum diameter of the tumor,lobulation sign,spiculation sign,pleural indentation sign and tumor density were included.Univariate analysis and multivariate logistic regression analysis were used to compare the clinical data and CT features between the two groups to establish the clinical data and CT features model.ITK-SNAP software was used to manually outline and segment all the layers of all lesions in the high resolution CT images,and to draw the region of interest(ROI)along the lesion outline,then import the original images and ROI files of all patients into the AK soft ware(Artificial Kit,GE Company,the United States)to extract 1316 imaging features from the CT images of each lesion.In the AI Statistical Learning Intelligent Platform(IPMStatistics,IPMs),Univariate analysis and the Least absolute shrinkage and selection operator(LASSO)were used to reduce the dimension of the extracted radiomics features to establish the radiomics model,and then Radscore was calculated.The prediction factors of radiomics features and Radscore were used to establish the joint model using Logistics regression model.The combined model was established by incorporating Radscore,CT features and clinical predictors.The area under ROC curve(AUC)was used to evaluate the predictive efficacy of the four models for TMB expression in non-small cell lung cancer.De Long test was used to compare the predictive efficacy of the four models.The decision curve analysis was used to evaluate the benefits of different threshold probabilities of each model.Calibration curve is used to evaluate the predictive efficiency of nomogram.Results: In clinical data and CT features,there were significant differences in sex,age,lesion density and lesion maximum diameter > 3cm between high and low TMB expression groups.After reducing the dimension by LASSO,the remaining four features of CT imaging features were used to establish the radiomic model.In the above models,the areas under the ROC curve of training group and testing group were 0.787(95%CI: 0.695-0.879)and 0.779(95%CI:0.634-0.923),respectively.The CT features model was established by incorporating the CT features including lesion density and maximum diameter > 3cm.The AUC of training group and testing group were 0.725(95%CI:0.628-0.822)and 0.783(95%CI: 0.682-0.876),respectively.The joint model was established by radiomics features and CT features,the AUC of training group and testing group was 0.802(95%CI:0.704-0.899)and 0.881(95%CI:0.635-0.915),respectively.The combined model was established by incorporating clinical predictive factors including age,gender,CT features and radiomics features.The AUC of training group and testing group were 0.863(95%CI:0.787-0.940)and 0.909(95%CI: 0.825-0.993)respectively.De Long test showed that the AUC of the combined model was statistically higher than the imaging model,CT features model and image merging model.Conclusions:1.Gender and age are clinical factors for predicting TMB expression status in non-small cell lung cancer;while lesion density and lesion maximum diameter can be used as CT features for prediction.2.The radiomics model has a high predictive value in TMB expression status of NSCLC.3.Among the four prediction models,the combined model has the highest prediction efficiency.