Segmentation of Brain Tumors from MRI Images Using Deep Neural Networks

Brain tumor is formed by the uncontrolled growth of cancerous (malign) or non-cancerous (benign) unhealthy cells in the brain. In the present world brain tumor is a very dangerous disease and the main reason for many deaths. Magnetic Resonance Imaging (MRI) is mostly used to produce medical images for the brain tumor analysis. This paper has two objectives, first one is to identify if the MRI images has or not brain tumor using Convolutional Neural Network algorithm based on ResNet50 architecture. For second objective, image segmentation, we propose a generalized focal loss function based on Tversky index. Compared to the commonly used Dice loss, the proposed loss function achieves better precision and recall when training on small structures such as lesions.