Automatic Brain Tumour Segmentation using fully Convolution Network and Transfer Learning

Abstract

Brain tumor segmentation is a challenging issue due to the heterogeneous appearance, shape, and intensity of tumors. In this paper, we present an automatic method for brain tumor segmentation in Magnetic Resonance Imaging (MRI) using deep neural networks (DNN). Transfer learning and fully convolution network (FCN) have been utilized to achieve robust tumor segmentation using VGG-16 network. The proposed architecture of the VGG-16 network includes the encoder and decoder networks with a classification layer to generate the pixel-wise classification. Comparison results demonstrate that the proposed method achieved state-of-the-art results with a global accuracy of 0.97785 and 0.89 dice score in terms of whole tumor segmentation on images from the BRATS2015 database.