Brain tumor detection and patient survival prediction using U-Net and regression model

Brain tumor segmentation is necessitated to ascertain the severity of tumor growth in a brain for possible treatment planning. In this work, we attempt the development of U-Net-based semantic segmentation of brain tumors. This network model is trained and tested on three MRI datasets: Brats 2018, Brats 2019, and Brats 2020. The trained U-Net yields the dice scores of 0.893, 0.837, and 0.753 on Brats 2018; 0.912, 0.891, and 0.808 on Brats 2019, and of 0.917, 0.894, and 0.811 on Brats 2020 for the complete tumor, tumor core, and enhancing tumor respectively. This paper also presents a novel formulation of a regression model based on an Information set to predict the survival rates of patients affected with a brain tumor. The weights of the regression model are learned using the pervasive learning model based on the pervasive information set. The overall survival rates of patients are predicted using the proposed regression model on the three datasets for which High-Grade Gliomas subjects are considered, and the model achieves the accuracies of 64.2%, 59.8%, and 60.5% on Brats 2018, 2019, and 2020 datasets respectively.