Ultrasound (US) imaging is a critical diagnostic tool in medical applications, but accurate segmentation remains challenging due to inherent noise, low contrast, and irregular anatomical textures. To address these challenges, we propose WF-UNet, an enhanced U-Net architecture that introduces two novel components: The WaveFusion Block and the Shuffle Group Attention Gate (SGAG) Block. The WaveFusion Block leverages wavelet transformations to capture high-frequency and multi-scale edge details, integrating them with spatial features to preserve edges in noisy images. The SGAG Block combines attention mechanisms and shuffle convolutions to enhance feature selection, suppress irrelevant features, and reduce computational costs. These modules are applied to the shallow layers of the network, where high-resolution spatial details are abundant. Experiments on three clinical datasets (BUSI, TN3K, and ISIC17) demonstrate WF-UNet’s superior performance over state-of-the-art (SOTA) methods. WF-UNet achieves the highest Dice scores (72.43%, 79.66%, 94.06%) and IoU values (64.59%, 69.90%, 89.24%) while significantly reducing HD95 and improving recall. These results highlight WF-UNet’s robustness in handling diverse and challenging segmentation tasks, particularly in noisy ultrasound images, while maintaining strong generalization across different medical imaging modalities. Our code is available at: https://github.com/viviennebla/WaveFusion .