Segmentation of Brain Tumor Tissues in Multi-channel MRI Using Convolutional Neural Networks

Unmanned segmentation of brain tumors is one of the hardest tasks to be solved in Computer Vision. In this work, we focus on Convolutional Neural Network model to segment tumorous cells in MRI brain scans. The inputs to the network are multi-channel MR image intensity information extracted from patches around each point to be predicted. The pre-processing steps are employed to precise the magnetic field bias and then intensity values are normalized using Z-score technique. The training was done for both HGG and LGG and the network was optimized with SGD in which the gradients are calculated using Nesterov Accelerated Gradient. The obtained results are promising for the complete tumor, the core tumor and the enhancing tumor segmentation. The propounded model achieved a dice score of 0.86, 0.62 and 0.65 for complete, core and enhancing tumor.