Abstract:Brain tumor MRI image segmentation is an important part of brain tumor diagnosis and treatment, aiming at the problem that the U-Net network structure has limited the size of the image feature receptive field and the gap in the context information caused by the low segmentation accuracy, this paper proposes a brain tumor MRI image segmentation algorithm that integrates multi-scale features. Firstly, a multiscale aggregation module (MAM) is designed to replace the conventional convolutional layer in the original U-Net network, increase the depth and width of the network, and capture the boundary details of the feature map. Secondly, a context atrous spatial pyramid (CASP) module is introduced in the encoding path and decoding path to expand the receptive field of the network and enhance the ability of the network to obtain multi-scale features. Finally, a multi-level aggregation attention module (MAA) is designed at the junction of the codec path, so that the network model pays attention to the effective features of the image segmentation area and eliminates background noise. The results show that the mean cross-union ratio, Dice coefficient, sensitivity, specificity, accuracy and other indicators of the proposed algorithm are 91.39%, 92.81%, 89.14%, 99.95% and 95.78%, respectively.