Brain Tumor Image Segmentation Network Based on Dual Attention Mechanism.

Due to the different shape, location and size of brain tumors, and their appearance is highly heterogeneous, which makes it difficult to establish effective brain tumor image segmentation rules. In view of the excessive depth and the lack of connection between global and local feature information of current medical image segmentation network, which leads to the reduction of image segmentation accuracy, we propose an improve brain tumor image segmentation network based on depth residuals and dual attention mechanisms. Firstly, inspired by the residual network, we replaced the traditional convolutional block with deep residual block, which not only can more feature information be extracted, but also network degradation can be suppressed and convergence can be accelerated. Secondly, the introduction of dual attention mechanism in each skip connection can fuse richer context information, making the model more focused on the features that need to be segmented in the tumor region, while suppressing the irrelevant regions. Then, the loss function of our model is the combination of cross entropy function and dice similarity coefficient function. Finally, MRI images from the BraTS2019 dataset are used to train and test, and the Dice coefficient is used to evaluate the segmentation accuracy. The average Dice coefficient of the whole tumor area WT, the core tumor area TC and the enhanced tumor area ET are 0.8214, 0.8408 and 0.7448, respectively. The experiment results show that our model can effectively improve the segmentation accuracy of brain tumor images compared with other deep segmentation models.