Segmentation of medical images performs a critical role in the diagnosis and treatment of diseases. Traditional convolutional neural network (CNN) methods are deficient in acquiring global information, and although U-Net enhances the segmentation effect by local improvement, it is still difficult to meet the high accuracy requirement. In this article, an improved hybrid architecture UNETR++  based on Transformer and U-Net is proposed, which combines convolution and self-attention mechanisms for better integration of global and local information. The architecture presents the EPA (Efficient Pairwise Attention) module, enabling feature interactivity in both spatial and channel dimensions, effectively improving segmentation accuracy along with a significant reduction in the amount of parameters and complexity of computation. In the cardiac MRI image segmentation experiments on the ACDC dataset, the Dice coefficients of UNETR++ on myocardium, right ventricle, and left ventricle segmentation reached 90.61%, 91.89%, and 96%, respectively, which are better than other 3D segmentation of medical images methods. Experimental results indicate that UNETR++ achieves efficient and accurate segmentation through the combination of Transformer and convolution, which provides more reliable image analysis support for the clinic.