Attention Gated Multi-Axial Fusion Network for Brain MRI Lesion Segmentation

Fast and accurate segmentation of brain stroke lesions is a very essential part for timely and proper diagnosis of brain stroke. Various deep learning based methods have been deployed for lesion segmentation from the Magnetic Resonance Imaging (MRI) data. Many of these methods are utilizing 2D segmentation network based. However 2D networks cannot leverage three-dimensional relationships in the MRI and suffer from slice level class imbalance problem. In this paper, we propose an attention gated multi axial fusion network (AGFN) that uses 2D network to segment along 3 different axes and then fuse the information from all 3 axes to generate final segmentation. AGFN is a shallow network with fewer parameters than most of the 3D networks and it uses attention mechanism for robust feature extraction making it efficient at integrating 3D information for segmentation. Along with methods for dealing with class imbalance problem, extensive experimental results show that our method improves performance of the constituent 2D network on the ATLAS dataset.