Brain tumor segmentation has important value for radiotherapeutic planning and therapeutic effect evaluation. Due to the shape diversity, location instability, structural complexity, and diverse pathological symptoms in different patients, traditional manual segmentation is not only difficult, time-consuming and laborious, but also depends on the personal experience of professional physicians. Therefore, how to segment brain tumors efficiently, accurately and fully automatically has become a research hotspot. In this paper, we propose an improved brain tumor segmentation architecture named AMRUNet++ based on UNet++. First, we add attention gates(AGs) to filter the features propagated through each skip connection. Second, we replace all the original two convolutional layers with MultiRes block. Finally, we add 'regions of interest (ROI)' to the network input and concatenate it to the output. In addition, to solve the problem of insufficient training data for medical images, we use the mixup principle for data augmentation. Extensive experiments are carried out on the CE-MRI data set. Experimental results show that AMRUNet++ achieves a dice score gain of 0.0529 points over UNet++ and adding the mixup principle also increases dice score by 0.0158 points.