A Comparison of Deep and Conventional Regression Methods for MRI-Based Estimation of Survival Time in GBM Patients

To evaluate the capability of conventional and deep regression methods in predicting the overall survival time of patients with glioblastoma multiform (GBM) based on magnetic resonance imaging (MRI).We obtained a dataset consisting of 118 GBM patients from The Cancer Imaging Archive (TCIA). Deep features were extracted from a pre-trained, Convolution Neural Network (CNN) and used for regression models. We developed eight conventional and three CNN-based regression methods to predict survival times. The conventional methods included regression with no regularization, Least Absolute Shrinkage and Selection Operator (LASSO), Ridge, Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Multilayer Perceptron (MLP), K-Nearest Neighbor (KNN). The CNN methods included LeNet, GoogLeNet, and ResNet. The 5-fold cross-validation was used to validate the prediction performance of these models. We calculated the mean absolute error (MAE), concordance index (c-index) and determination of coefficient (R2-score) as the performance metrics.The MAEs of the 11 aforementioned regression models were 286±19.86, 275±10.1, 276±7.1, 187±14.1, 179±5.5, 205±10.2, 198±8.7, 211±8.2, 175±9.5, 117±4.5, and 168±11.2, respectively. Based on the Kaplan Meier analysis, the p-values of XGBoost, DT and GoogLeNet, were lesser than 0.05. Their accuracy in classifying the patients into low and high risk was 86.9%, 88.5%, and 92.7%, respectively.The CNN-based methods (specially GoogLeNet) demonstrated better performance in terms of MAE than conventional methods. The ensemble methods like RF and XGBoost and MLP approaches achieved the best performance in terms of MAE among the conventional methods.