Segmentation of gliomas in magnetic resonance images using recurrent neural networks

In our work we focus on automatic segmentation of high-grade gliomas (HGG) from magnetic resonance images (MRI). The results of segmentation have great impact on treatment of patients and consequently on the length of their life. In this paper a new approach of automatic glioma segmentation based on recurrent neural units is proposed. We use the Long Short-Term Memory units (LSTMs) which are able to extract latent features of brain structure by global contextual information. Unlike convolutional neural networks, where global context is gained by combination of local features, LSTMs have the potential to capture the global context at once. We use a region-based classification using the 3D Hilbert space-filling curve. To evaluate this method, the HGG data from the International Multimodal Brain Tumor Segmentation (BraTS-17) Challenge 2017 are being used. Our method achieved a dice score 0.62, 0.77, 0.64, on validation dataset of BraTS-17, for enhancing tumor, whole tumor and tumor core, respectively.

I. Introduction

A Glioma is one of the most common types of a brain tumor. It is specific for the heterogeneity of its origin, shape, size, position and aggressiveness. One of the reasons of heterogeneity is an origin of a glioma. Glioma arises from glial cells in the brain, which are divided into four different types of cells and make up to 90% of the whole brain volume. With an addition of different types of glial cell mutations a glioma has a wide range of origins [1] .