In this paper, the U-Net model was constructed to segment the multi-modal MRI images by introducing the Attention mechanism, and the image characterization characteristics of each mode in the MRI images were used to train and learn, and the brain tumor core and substructure were segmented from normal brain tissue. Based on the selection of the BraTS2021 data set for training, the sample data set is used as the training set, verification set, and test set according to the ratio of 8:1:1. The SimpleITK library in Python is used to read the NIFTI file image and preprocess it with random clipping, center clipping, random flipping, Gaussian noise, contrast transformation, brightness transformation, data type conversion, etc., and extract the image features of the model in the training process. The experimental results show that the model established based on U-Net has good segmentation performance and high convergence speed, the highest accuracy rate of tumor enhancement area, core area, and overall area is 78.94%, 78.81%, and 88.75%, respectively, and the lowest loss function is 0.1653. The method has a high segmentation accuracy rate. It is less time-consuming and simple to operate and can provide a reference for MRI brain tumor image segmentation.