Objective To explore the effectiveness of the FA-UNet model in improving the accuracy of lesion segmentation in three-dimensional medical images by incorporating the global position attention (GPA) block and frequency aware (FA) block. Methods This paper proposes the FA-UNet model, which builds upon the original UNet architecture by adding the GPA-block and FA-block in the bridge layer. This modification allows the model to effectively learn positional information in three-dimensional medical images without significantly increasing computational burden. Experiments were conducted using a 3D CT thymus lesions dataset, and segmentation performance was evaluated using metrics such as the Dice coefficient. Results Compared to traditional UNet, FA-UNet achieved significant performance improvements with the same computational resources, with the Dice coefficient for lesion segmentation reaching 86.16%, and the performance improvement corresponding to 3.44 times the additional computational cost. Conclusion By incorporating the GPA-block and FA-block, FA-UNet can improve segmentation accuracy while optimizing computational efficiency, making it an effective method for 3D medical image segmentation.