HWASeg : Hierarchical Wavelet‐Assisted Network for Multi‐Organ Segmentation

In the multi‐organ segmentation task of medical images, convolutional neural networks (CNNs) are commonly employed for feature extraction. Traditional convolutional layers primarily focus on extracting information from local neighborhoods, which allows them to learn high‐frequency, detailed features in images, like edges and textures. However, this focus limits the network's ability to capture low‐frequency information, such as global features that represent the overall shape and size of organs. To overcome this limitation, this paper presents a novel hierarchical wavelet‐assisted multi‐organ segmentation network, named HWASeg. This network aims to extract both global structural features and local detailed information by utilizing convolution in the frequency domain. We have designed a Wavelet Hierarchical Fusion Module (WHFM) to enhance the representation of low‐frequency and high‐frequency information via a hierarchical multi‐level decomposition strategy. Furthermore, to boost the segmentation performance, we have developed a Progressive Depthwise Bottleneck Module (PDBM). This module extracts discriminative features by integrating multiple depthwise dilated convolutions. The experimental results demonstrate that HWASeg achieves good performance on multi‐organ abdominal segmentation (Synapse) and Automated Cardiac Diagnosis Challenge (ACDC) datasets.