Efficient scheme to perform semantic segmentation on 3-D brain tumor using 3-D u-net architecture

Glioma is the most common type of brain tumor with varying level of malignancies and projection. Designing personalized therapy and to foresee response towards the therapy needs better understanding of tumor biology and diversification between tumors. Using different computational methods, accurate segmentation of tumors within the brain on MRI is the primary stride towards the understanding of tumor biology. The goal of current study is to draw an algorithm for MRI image segmentation of pre-treatment brain tumors and to evaluate its performance. In our research, we designed and implemented a novel 3D U-Net architecture for segmentation of sub-regions including edema, necrosis and enhancing tumor which are radiologically detectable. The group variance between tumor and non-tumorous spots is addressed by presenting weighted patch extraction scheme from tumor border regions. In its framework, context is captured using a contracting path and precise localization is performed by symmetric expanding path. In our study, the architecture based on Deep Convolutional Neural Network (DCNN) is trained on Brain Tumor Segmentation (BraTS) dataset of 750 patients among which 484 scans were labelled and 267 scans were used as training dataset. 3D patches were extracted from the dataset to train the system and results were assessed in terms of Specificity, Sensitivity and Dice Score. Our proposed system achieved Dice scores of 0.90 for whole tumor, 0.85 for tumor core, and 0.77 for enhancing tumor on dataset which shows potential of accurate intra-tumor segmentation of patch-based 3D U-Net architecture.