Herein, U-Net structure was improved to segment magnetic resonance (MR) images of brain tumors to address the loss of information in image segmentation in the full convolutional neural network and low segmentation accuracy caused by fixed weights. Based on the attention module in the U-Net contraction path, the weights were distributed to different size convolutional layers, which is beneficial to information usage for image space and context. Replacing the original convolution layer with the residual compact module can extract more features and promote network convergence. The brain tumor MR image database provided by BraTS (The Brain Tumor Image Segmentation Challenge) is used to validate the proposed new model and evaluate the segmentation effect using the Dice score. The accuracy of 0.9056, 0.7982, and 0.7861 was obtained in the total tumor region, core tumor region, and tumor enhancement, respectively, demonstrating that the proposed U-Net structure can enhance the accuracy and efficiency of MR image segmentation.