To Study The Value Of PET/CT Radiomics In Predicting EGFR Mutation In NSCLC Patients And To Analyze PD-L1 Expression

Objective:The diagnostic model of ¹⁸F-FDG PET/CT radiomics to predict EGFR mutation was constructed and validated by comparing radiomics between EGFR mutation and wild type in diagnosed untreated non-small cell lung cancer,and then the diagnostic efficacy of the model to predict EGFR mutation was evaluated by applying ROC curves,efficacy validation and DCA curves,aiming to explore a possible non-invasive method to obtain EGFR mutation results.The aim was to explore a possible method to obtain EGFR mutation results non-invasively.In addition,Patients with EGFR mutations combined with PD-L1 expression were also analyzed in relation to radiomics combined with clinical features and EGFR-TKI efficacy and other variables to summarize valuable clinical diagnosis and treatment information.Methods:One hundred and seventy patients with NSCLC who visited the Affiliated Hospital of Inner Mongolia Medical University between January 2020 and December 2022 were collected,and 85 cases each with EGFR mutation and EGFR wild type were finally screened according to the inclusion criteria,and each group was divided into a training set（n=119）and a validation set（n=51）in a ratio of 7:3.The differences in clinical features between EGFR mutant and wild type were first analyzed in the training set and validated in the validation set,and the clinical feature prediction model was established in the training set using binary logistic regression;then the open source software LIFEx 7.2.0 was used to outline the tumor lesions and extract 246 imaging histological features,and 15 imaging histological features were extracted in the training set using the LASSO algorithm and validated by the open source software Python.The output of the model is in the form of a Rad-score formula,i.e.,each enrolled patient will get a Rad-score for predicting EGFR mutations.Finally,a composite model（imagingomics+clinical characteristics model）is established using logistic regression analysis.The predictive ability of these three prediction models was evaluated using ROC curves,efficacy validation,and clinical decision curves（DCA）.A subgroup analysis of 142 patients with PD-L1 testing was performed to explore the differences in imaging histological characteristics,clinical characteristics and EGFR-TKI efficacy between patients with EGFR mutations combined with positive PD-L1 expression and those with negative PD-L1 expression.Results:Two differential features of gender and pathology type were screened in the training set for building the clinical feature prediction model,and the same method was validated by the validation set,and the results of both were 100%（2/2）consistent.Fifteen imaging histological features with statistically significant differences were screened by the LASSO algorithm,including 9 shape features,2 histogram features and 4 texture features,and then the XGboost algorithm was used to build the imaging histological prediction model（Rad-score）and calculate the Rad-score of all enrolled patients,and the results showed that the Rad-score of EGFR mutant type was significantly higher than EGFR wild type,i.e.,Rad-score higher than0.907 indicates EGFR mutation;11 differential imaging histological features were validated in the validation set,and their compliance rate was 73.3%（11/15）;correlation analysis showed that the correlation coefficient interval between 15 imaging histological features and2 clinical features was（-0.156,0.183）,and there was no correlation（P>0.05）i.e.,changes in clinical features did not cause changes in imaging histological features.A composite（clinical+imaging histology）model was established by binary logistic regression of the 2 clinical characteristics of gender and pathology type with the Rad-score,and the validation set was100%（3/3）consistent with it.The diagnostic efficacy of the composite,imaging histology and clinical models for differential diagnosis of EGFR mutation or not was good,with AUCs of 0.995,0.972 and 0.76 in the training set,and the results in the validation set were consistent with 0.978,0.958 and 0.764.The DCA curves showed a high risk threshold（probability of the model predicting an outcome of EGFR mutation）ranging from 0 to 100%.The clinical benefit of the composite model was higher than the other two models;the validation of efficacy using EGFR-TKI showed no significant difference（P<0.05）between the three prediction models and the gold standard（pathology and blood ct DNA test results）.Subgroup analysis showed that three differential imaging features were extracted between EGFR19Del deletion and EGFR21 L858R mutation,namely MORPHOLOGICAL＿Sphericity,INTENSITYHISTOGRAM＿Intensity Histogram Maximum Grey Level and GLCM＿Normalised Inverse Difference Moment.the percentage of positive PD-L1 expression was higher in the EGFR wild-type group than in the EGFR mutant group（55.1%vs.30.1%）;univariate analysis showed that squamous carcinoma was a significant predictor of EGFR mutation combined with PD-L1 expression;2 differential imaging features were screened by LASSO algorithm:Maximum Histogram Gradient Grey Level（MHGGL）and GLSZM＿Small Zone High Grey Level Emphasis（SZHGLE）.The AUC values for differential diagnosis of EGFR mutation combined with positive PD-L1 expression and negative PD-L1 expression were 0.758,0.729,and 0.604,respectively,using ROC curves for the 3 features,i.e,SZHGLE,MHGGL,and pathological type;the AUCs of the 2 imaging histological features were higher than those of the pathological type,but the diagnostic efficacy was average for both.For 69patients with first-line efficacy analysis with EGFR-TKI（72.5%of generation EGFR-TKI）,m PFS1 in the PD-L1 positive group（3.7 months）<EGFR mutation overall population（10months）<PD-L1 negative group（11 months）.Conclusion:1.All three prediction models based on ¹⁸F-FDG PET/CT imaging histology and clinical characteristics could predict EGFR mutation or not,and the composite model based on imaging histology combined with clinical characteristics had the best diagnostic efficacy,in which the AUC of the composite model,imaging histology model and clinical model for differential diagnosis of EGFR mutation or not were 0.995,0.972 and 0.760 in descending order.The diagnostic efficacy obtained by DCA curve and efficacy validation method is consistent with this.2.¹⁸F-FDG PET/CT radiomics analysis showed that the EGFR mutation group had 15differential radiomics features compared with the EGFR wild type in NSCLC,suggesting that EGFR mutated tumor lesions exhibited small volume,irregular shape,and more heterogeneous texture;2 intensity histogram features were extracted from PET images showing that EGFR mutated lung cancer lesions were highly metabolism,but conventional metabolic indexes such as SUVmax,MTV,and TLG did not show predictive ability for EGFR mutation or not.3.There were 3 differential radiomics features among the mutant subtypes of EGFR.Compared with EGFR21L858R mutation,the imaging histological features of EGFR19 Del mutation tended to show a near spherical morphology and high heterogeneity of texture.4.EGFR mutations were negatively correlated with PD-L1 expression.patients with NSCLC with EGFR mutations combined with PD-L1 expression had the following characteristics:imaging histology showed a high degree of dispersion and poor texture homogeneity,a greater tendency to develop squamous carcinoma,poor treatment effect on EGFR-TKI,and short survival time.