Abstract: In this study, we propose a multiscale efficient brain tumor segmentation network using ShuffleNet to obtain a more accurate segmentation of brain tumor regions when using limited hardware platform resources. First, we design a deep feature extraction network based on ShuffleNet and added multiple parallel convolution layers and mixed receptive fields to enhance the multiscale information extraction capability of the network. Next, we apply a depthwise separable convolution to reduce the number of network parameters. Finally, we propose a weighted mixed loss function to alleviate the impact of data category imbalance on brain tumor segmentation and improve the stability of network segmentation. Furthermore, we employ the BraTS2019 dataset for training and validation and conduct clinical testing on the clinical patient dataset BraTS2021. The experimental results indicate that the proposed deep lightweight network reduces the number of network parameters and computations, has higher segmentation accuracy, and performs better in enhancing tumor region segmentation.