Multi-dimensional Deep Learning Model Based On Radiomics For Prognosis Prediction Of Non-small Cell Lung Cancer Patients Receiving Bevacizumab

Part ? Objectives: To construct and validate radiomics signature to predict the survival of NSCLC patients receiving bevacizumab and chemotherapy.Methods: NSCLC patients receiving bevacizumab and chemotherapy were enrolled in analysis,then radiomics features were extracted.Least absolute shrinkage and selection operator Cox(LASSO-Cox)was performed to identify prognostic radiomics features,and generate radiomics signature(Radscore).Results: Total of 195 NSCLC patients were included in analysis,and characteristics were balanced between patients in training and validation cohort.Radscore consisted of 5 selected features and was significantly associated with PFS(C-index 0.65)in training cohort.The Radscore was further validated in validation cohort(C-index 0.60).Conclusion: Radiomics exhibited a favorable effect to predict the survival of NSCLC patients receiving bevacizumab and chemotherapy.Part ? Objectives: To construct and validate a deep neural network(DNN)system to predict the survival of NSCLC patients receiving bevacizumab and chemotherapy,and infer the underlying biological property of the DNN survival prediction system.Methods: The features of NSCLC patients receiving bevacizumab and chemotherapy were included in high-dimensional feature set,and a low-dimensional feature set was selected with feature engineering.The study trained and optimized DNN by adjusting hyperparameter,then generated Deep Surv and N-MTLR survival prediction system.DNN was compared with CPH and machine learning prognostic model,then external validated in prospective dataset.DNA methylation sequencing and functional analysis was performed to infer the underlying biological property.Results: Total of 272 NSCLC patients receiving bevacizumab and chemotherapy were enrolled in analysis.The study trained DNN with Re Lu activation and He?uniform weight initialization,and dropout and L2 regularization was performed to control overfitting.Deep Surv and N-MTLR survival prediction system was trained after 1000 iteration and optimized by adjusting hyperparameter.The performance of highdimensional Deepsurv was the best with the c-index of 0.712,which was better than CPH(0.665)and RSF model(0.68),and was further validated in prospective dataset with the c-index of 0.73.All patients were divided into high and low risk group by DNN survival prediction system.There were significant differences in DNA methylation promotor region between two groups,and corresponding genes included SOX9,MAP2K2,AKT1 etc.,which were significantly enriched in the biological process related to cell cycle regulation,cell adhesion and tumor immunity.And patients in high risk group were related to high level of Foxp3 and PD-L1.Conclusion: The DNN is applicable to predict the survival of NSCLC patients receiving bevacizumab and chemotherapy.And predictive effect of DNN survival prediction system was related to the tumor immune microenvironment.Part ? Objectives: To construct a multi-dimensional prognostic model based on the radiomics to predict the survival of NSCLC patients receiving bevacizumab and immunotherapy.Methods: NSCLC patients receiving bevacizumab and immunotherapy were enrolled.The study constructed and compared prognostic model by integrating multidimensional features using CPH,RSF,DNN algorithm,and respective nomogram was performed.Results: 39 NSCLC patients receiving bevacizumab and immunotherapy were included.EGFR mutation status,chemotherapy and NLR2 were independent prognostic factors for PFS.The Delta Rad was generated with 2 selected features,and the c-index was 0.670.The performance of multi-dimensional CPH model was the best with the c-index of 0.717 compared with DNN and RSF model.Conclusions: Delta Rad could assist RECIST criteria in early evaluation of efficacy.There was a best performance of multi-dimensional CPH model based on radiomics to predict the survival of NSCLC patients receiving bevacizumab and immunotherapy.