A Novel Brain Tumor Segmentation Method Based on Im-proved Spatial Attention Mechanism and Multi-path Deep Neural Network

Abstract

Preoperative Magnetic Resonance Image (MRI) brain tumor diagnosis is an effective technical approach. To accurately segment tumor regions, we propose a novel brain tumor segmentation method based on improved Spatial Attention mechanism and Multi-path neural network (SAMPU-Net). Firstly, we propose a multipath input method to extract feature information of different scales by using convolution kernels of different sizes, so as to fully extract MRI feature information. Secondly, we improve the spatial attention mechanism by adding convolution layer of pyramid structure to it to obtain the features of different receptive fields. In the convolution layer of this pyramid structure, the larger the convolution kernel is, the more global features will be extracted; conversely, the smaller the convolution kernel is, the more local features will be extracted. Thirdly, we use more multi-mode MRI information to segment the brain tumor images. In practical application, due to the fuzzy tumor regions in some MRIs, we use the method of restricted contrast adaptive histogram equalization to perform local enhancement of images. The proposed model and several other mainstream segmentation methods are trained and tested on the BraTS2019 public dataset. Experimental results indicate that using our method, the Dice coefficient of tumor core and tumor enhancement region is increased by 2.4% and 1.3% respectively, and our proposed method has better segmentation effect than other methods.