Radiomic features extracted from MR images, along with other clinical covariates, have been shown to facilitate glioma patient prognostication via survival analysis. In this study, we apply the DeepSurv and neural network-parameterized probability mass function (PMF-NN) models to the survival analysis of glioma patients using the age and extracted radiomic features from multiparametric MRIs in the Brain Tumor Segmentation Challenge 2018 (BraTS 2018) dataset. First order and texture features were calculated for the whole tumor and peritumoral brain zone (PBZ) regions in both T1-weighted contrast-enhanced (T1-wce) and T2-weighted fluid attenuated inversion recovery (T2-FLAIR) images. To the best of our knowledge, this work is the first to apply the DeepSurv and PMF-NN models to radiomics-based survival analysis of glioma patients. Whole tumor shape features were also calculated from the delineated masks of the whole tumor. Pearson correlation analysis showed age, 4 first-order, 10 shape, and 17 texture features are correlated with the overall survival of glioma patients. These correlated features were used as input for the developed DeepSurv and PMF-NN models. The best performing DeepSurv model in our study obtained a 1050-day integrated Brier score (IBS) and C-index of 0.122 and 0.670, respectively. For the best performing PMF-NN in terms of IBS, a 1050-day IBS of 0.125 and a C-index of 0.632 were obtained. Thus, our DeepSurv and PMF-NN models provide clinically informative predictions. Our developed models outperformed the Cox proportional hazards (CPH) with supervised principal component analysis (SPCA) from a separate study in terms of IBS.

No abstract

No abstract

No abstract

No abstract