Renal tumors pose great harm to and seriously affect human health. Early detection and diagnosis of renal tumors can help patients' treatment and recovery. To efficiently segment kidneys and tumors from abdominal computed tomography (CT) images, this paper proposes a method based on 3D U²-Net. First, we upgrade the 2D U²-Net, adjust loss function, depth of network, and supervision strategy. To improve the feature expression ability of the decoder, we propose a residual feature enhancement module, which enhances the channel and spatial domain of the feature map at the decoder. To further improve the model's ability to extract global information, we propose a multi-head self-attention module based on global features, which calculates the long-term dependencies between all voxel points in the feature map and obtains more contextual information of 3D medical images. The method is tested on the official KiTS19 dataset and the results show that the average Dice value is 0.9008 and the parameter quantity is 4.60 MB. Compared with existing methods, our method can achieve better segmentation accuracy with small parameter quantity, and has great application value for small memory embedded system used for kidneys and tumors image segmentation.