Radiomics based Deep Fully Connected Neural Network (R-DNN) for Prognostication of Lung Cancer

329 Objectives: Baseline positron emission tomography with fluorodeoxyglucose (FDG-PET) based radiomics are of increasing interest for lung cancer prognostic studies. However, many different strategies can be considered in particular to solve the issue of features selection for building a prognostic signature. In this study a radiomics-based Fully Connected Deep Neural Network (R-DNN) was developped to predict the prognosis without any prior need for a feature selection method. Methods: One hundred and thirty eight patients with stage I-III non-small cell lung cancer treated with radio-chemotherapy were retrospectively included in this analysis. Patients were classified according to median overall survival (OS) into high-risk or low-risk. Radiomics features following the IBSI standardization guidelines were extracted from automatically delineated metabolic tumor volumes (MTV) on baseline PET scans. A Fully Connected Deep Neural Network (DNN with 5 hidden layers, 400 hidden units per layer, 50% drop-out and the ELU activation function) was directly trained exploiting all radiomics features for prediction. The trained DNN was evaluated with 10-fold cross-validation using the area under the curve (AUC) of the receiver operating characteristic (ROC). Its performance was compared to each individual radiomics features. Results: The AUC for each individual radiomics features ranged from 0.60 (Sensitivity (Se)=45% and Specificity (Sp)=78%) to 0.70 (Se=86%, Sp=53%). By comparison, when using all radiomics features as input to the DNN, the mean AUC was 0.79 (Se=75% and Sp=68%) in training and 0.71 (Se=71% and Sp=64%) in cross-validation. Conclusions: PET-derived radiomics with Fully Connected Deep Neural Network (R-DNN) algorithm provided high accuracy in prognostication of lung cancer patients with multiple radiomics and could eliminate the need for any prior radiomics feature selection.