Brain tumor segmentation with cascaded deep convolutional neural network

Abstract

Cancer is the second leading cause of death globally and is responsible for an estimated 9.6 million deaths in 2018. Approximately 70% of deaths from cancer occur in low and middle-income countries. One defining feature of cancer is the rapid creation of abnormal cells that grow uncontrollably causing tumor. Gliomas are brain tumors that arises from the glial cells in brain and comprise of 80% of all malignant brain tumors. Accurate delineation of tumor cells from healthy tissues is important for precise treatment planning. Because of different forms, shapes, sizes and similarity of the tumor tissues with rest of the brain segmentation of the Glial tumors is challenging. In this study we have proposed fully automatic two step approach for Glioblastoma (GBM) brain tumor segmentation with Cascaded U-Net. Training patches are extracted from 335 cases from Brain Tumor Segmentation (BraTS) Challenge for training and results are validated on 125 patients. The proposed approach is evaluated quantitatively in terms of Dice Similarity Coefficient (DSC) and Hausdorff95 distance.