3D EMSU‐Net: A framework for automatic segmentation of brain tumors

Abstract

Glioma is the most common tumor in the brain’s central nerve cells, which is extremely dangerous clinically. Glioma’s accurate surgical localization and diagnosis both rely on the segmentation result of the tumor area in brain Magnetic Resonance Imaging (MRI) images. Recently, deep learning methods have been widely used in the tasks of semantic segmentation of brain tumor images. However, traditional 2D Convolutional Neural Networks (CNNs) only consider the feature information of images rather than the spatial features, which is hard to learn the contextual information between adjacent MRI slices of brain tumors. Simultaneously, 3D CNNs have attracted more and more attention due to their concentration on the correlation among slices. Inspired by 3D U-Net [1], this paper proposes an efficient 3D EMSU-Net to segmentation multi-modal brain tumor images with attention mechanism [2] and Efficient Multi-Scale Feature Extraction Component (EMS), which pays more attention to the feature information related to brain tumor areas to achieve over-performance segmentation of brain tumors. The proposed network is applied to brain tumor segmentation tasks and verified by BRATS 2018 dataset in experiments. The results demonstrated that 3D EMSU-Net is superior to other existing methods.