Two-step u-nets for brain tumor segmentation and random forest with radiomics for survival time prediction

Abstract

In this paper, a two-step convolutional neural network (CNN) for brain tumor segmentation in brain MR images with a random forest regressor for survival prediction of high-grade glioma subjects are proposed. The two-step CNN consists of three 2D U-nets for utilizing global information on axial, coronal, and sagittal axes, and a 3D U-net that uses local information in 3D patches. In our two-step setup, an initial segmentation probability map is first obtained using the ensemble 2D U-nets; second, a 3D U-net takes as input both the MR image and initial segmentation map to generate the final segmentation. Following segmentation, radiomics features from T1-weighted, T2-weighted, contrast enhanced T1-weighted, and T2-FLAIR images are extracted with the segmentation results as a prior. Lastly, a random forest regressor is used for survival time prediction. Moreover, only a small number of features selected by the random forest regressor are used to avoid overfitting. We evaluated the proposed methods on the BraTS 2019 challenge dataset. For the segmentation task, we obtained average dice scores of 0.74, 0.85 and 0.80 for enhanced tumor core, whole tumor, and tumor core, respectively. In the survival prediction task, an average accuracy of 50.5% was obtained showing the effectiveness of the proposed methods.