Enhanced Medical Image Segmentation via Wavelet-Deformable Attention Networks

Abstract Medical image segmentation plays a crucial role in improving diagnostic accuracy and therapeutic strategies. Traditional methods often struggle with noise interference and blurred boundaries in medical images. This study introduces a novel segmentation network, termed Wavelet-Deformable Attention Network (WDANet), featuring a Wavelet-Enhanced Deformable Attention Decoder (WEDA). WEDA integrates a Wavelet Deformable Convolution Attention Block (WDCB) and a Multi-Scale Frequency space Fusion Block (MSFB) to enhance feature extraction and contextual awareness. The WDCB combines wavelet deformable convolution with a convolutional attention mechanism, while the MSFB jointly models multi-scale spatial information and multi-frequency channel representations. Experiments on the ACDC and Synapse datasets demonstrate that WDANet achieves average Dice scores of 92.53% and 85.67%, respectively, outperforming existing state-of-the-art methods. These results validate the effectiveness of WDANet in medical image segmentation applications.The code is publicly available at https://github.com/zhangxuan-thecastle/WDANet.