Prediction Of Non-Small Cell Lung Cancer(NSCLC) EGFR Mutation,ALK Rearrengment Status And Acquired EGFR T790M Mutation Based On Radiomics And Machine Learning

Purpose:1.To determine the role of CT texture features and CT characteristics in predicting epidermal growth factor receptor(EGFR)mutations and echinoderm anaplastic lymphoma kinase(ALK)rearrangements in non-small cell lung cancer(NSCLC).2.To develop an appropriate machine learning model for predicting anaplastic lymphoma kinase(ALK)rearrangement status in non-small cell lung cancer(NSCLC)patients using computed tomography(CT)images and clinical features.3.To investigate the deep learning model(DLM)combining CT radiomic features and clinicopathological information for predicting EGFR mutation status in non-small cell lung cancer(NSCLC)patients.4.To evaluate the clinical features and CT radiomics deep learning prediction acquired T790M mutation status of non-small cell lung cancer(NSCLC)after epidermal growth factor receptor(EGFR)-tyrosine kinase inhibitor(TKI)therapy failure.Methods:1.We retrospectively analyzed Two hundred and forty-two patients who were diagnosed with non-small cell lung cancer and had EGFR mutations and ALK rearrangements tested.All patients underwent chest CT scanning.Conventional morphological features of tumors including location,size,shape,margin,density,air bronchogram,pleural thickening,pleural retraction and pleural effusion were analyzed.12 CT texture features of tumors were analyzed.Logistic regression analysis was performed to determine the predictive value of conventional features and texture parameters for EGFR mutations and ALK rearrangements.2.We retrospectively analyzed 193 patients with NSCLC(154 in the training cohort,39 in the validation cohort),68 of whom tested positive for ALK rearrangements and 125 of whom tested negative.From the nonenhanced CT scans,157 radiomic characteristics were extracted,and 8 clinical features were collected.Five machine learning(ML)models were assessed to find the best classification model for predicting ALK rearrangement status.A radiomic signature was developed using the least absolute shrinkage and selection operator(LASSO)algorithm.The predictive performance of the models based on radiomic features,clinical features,and their combination was assessed by receiver operating characteristic(ROC)curves.3.We retrospectively analyzed 735 patients of non-small cell lung cancer with complete reports of EGFR gene testing.Collected the clinical pathological features,analyzed clinical features,then manually segmented ROI and extracted radiomic features,and deep learning features.The cases were randomly divided into training,validation and test set.We carried out feature screening,then applied the logistic regression to develop sole models,and fused models to predict EGFR mutation status.The efficiency of models was evaluated by ROC and PRC curves.4.We retrospectively analyzed 77 patients with NSCLC who underwent rebiopsy after first-line or second-line EGFR-TKI therapy was conducted.Rebiopsy methods included CT guided lung biopsies,endobronchial US-or bronchofibroscopy-guided biopsies,US-guided lymph node biopsies,pleural fluid analysis,other solid organ biopsies,and cerebrospinal fluid analysis.Deep learning predicted acquired EGFR T790M mutation models were constructed from pre-treatment chest CT images.Progression-free survival(PFS),the duration from the start of TKI therapy to rebiopsy were calculated.Results:1.In the 242 patients,there were 128 patients with EGFR(n=64 ex19del and n=64 L585R)and 37 patients with ALK rearrangements,77 patients with EGFR mutation and ALK rearrangements negative.ALK rearrangements was associated with young age,large lesion size,smooth margin,solid lesions and texture parameters including Variance,Skewness,Kurtosis,Contrast,Entropy,Homogeneity,Correlation,Sum-Average,Variance2,Dissimilarity(P=0.000-0.015).Ex19del was associated with spiculate,pleural retraction,Auto-Correlate(p=0.002-0.004).L858R was associated with age,smooth margin,Auto-Correlate(P=0.025-0.046).Multiple logistic regression analysis showed clinical feature,conventional CT features and texture parameter predictors ALK rearrangement,the AUC of ROC was 0.901(P<0.0001).2.The support vector machine(SVM)model had the highest AUC of 0.914 for classification.The clinical features model had an AUC=0.805(95%CI 0.731-0.877)and an AUC=0.735(95%CI 0.566-0.863)in the training and validation cohorts,respectively.The CT image-based ML model had an AUC=0.953(95%CI 0.913-1.0)in the training cohort and an AUC=0.890(95%CI 0.778-0.971)in the validation cohort.For predicting ALK rearrangement status,the ML model based on CT images and clinical features performed better than the model based on only clinical information or CT images,with an AUC of 0.965(95%CI 0.826-0.882)in the primary cohort and an AUC of 0.914(95%CI 0.804-0.893)in the validation cohort.3.We successfully established Model clinical,Model radiomic,Model CNN(based on clinical features,radiomic and deep learning features respectively),Model radiomic+clinical(combining clinical features and radiomic features),and Model CNN+radiomic+clinical(combining clinical features,radiomics,and deep learning features).Among the prediction models,ModelCNN+radiomic+clinica]showed the highest performance AUC=0.801,followed by Model CNN,and then Modelradiomic+clinical.Further analysis showed that ModelCNN+radiomic+cunical effectively improved the efficacy of Modelradiomic+clinical and showed better efficacy than Model CNN.Among distinguishing ex19del and L585R mutations,the highest AUC value was 0.775 in Model CNN+radiomic.4.At rebiopsy,45(58.4%)patients were T790M mutation positive.Deep learning predictive TM790 mutation the area under the receiver operating characteristic curves(AUC)were 0.779.Conclusion:1.Our analyses revealed that CT texture features and CT characteristics of lung adenocarcinomas harboring ALK rearrangements were significantly different,compared with those with EGFR mutations.These differences may be related to the molecular pathology of these diseases.2.Our findings revealed that ALK rearrangement status could be accurately predicted using an ML-based classification model based on CT images and clinical data.3.Either deep learning or radiomic signature-based models can provide a fairly accurate non-invasive prediction of EGFR expression status.The model combined both features effectively enhanced the performance of radiomic models and provided marginal enhancement to deep learning models.Collectively,fusion models offer a novel and more reliable way of providing the efficacy of currently developed prediction models,and have potential for the optimization of noninvasive EGFR mutation status prediction methods.4.Deep learning based on chest CT may be useful for predicting acquired EGFR T790M mutations in NSCLCs after first-or second-generation EGFR TKI resistance in NSCLC patients.