Spatial Attention-based Efficiently Features Fusion Network for 3D-MR Brain Tumor Segmentation

Accurate segmentation of brain tumors is critical for gliomas diagnosis and treatment planning. A large number of studies have shown that 3D convolutional neural networks (CNN) can use 3D convolution to capture the correlation between adjacent slices of brain magnetic resonance images (MRI). However, these methods often failed in the insufficient fusion of different layer features which have resolution and semantic difference and resulted in model information confusion. To this end, this paper presents a 3D spatial attention based efficient feature fusion network (SA2FNet) to achieve more accurate dense volumetric segmentation. Spatial attention mechanism methods are introduced in this paper to solve the problem of missing segmentation details caused by model information confusion. Then the deep supervision is employed to prevent the gradient explosion/disappearance problem of CNN and speed up the convergence, which significantly improves the performance. The experiment results indicate the SA2FNet has better performance in segmenting the brain tumor detail.