Joint Learning of Segmentation and Overall Survival for Brain Tumor based on U-Net

Junmo Kwon, Hyunjin Park

Abstract: The prognosis, hence survival, of patients with brain tumors is highly dependent on the size and grade of the tumor. Thus, joint learning of brain tumor segmentation and overall survival of patients with brain tumors can benefit each other. In this work, we explored the feasibility of prediction of patients' overall survival through U-Net guided by the information of brain tumor segmentation. We evaluated the proposed model on the multimodal brain tumor segmentation (BraTS) 2017 challenge dataset. We achieved the mean Dice score of 0.595 for brain tumor segmentation on the test set. The Pearson correlation coefficient for overall survival prediction on the test set was 0.243, indicating promising results for both brain tumor segmentation and overall survival prediction.