Wavelet-ASPP U-Net: Enhancing ultrasound image segmentation with multi-scale context and frequency-aware features

Ultrasound imaging is one of the most economical and non-invasive imaging modalities in clinical diagnosis, and ultrasound image segmentation is of great value for disease diagnosis, treatment planning, and image-guided intervention. However, accurate ultrasound image segmentation remains challenging because of low contrast, blurred boundaries, speckle noise, and indistinct features of lesions. To address these challenges, we proposed a novel network architecture, named Wavelet-ASPP U-Net (WAU-Net). Built upon the classical U-Net architecture, WAU-Net integrates a wavelet decomposition module and an Atrous Spatial Pyramid Pooling (ASPP) module at the bottleneck layer to jointly model frequency-domain details and multi-scale spatial context. Specifically, discrete wavelet decomposition separates bottleneck feature maps into low-frequency structural components and high-frequency detail components. The high-frequency subbands preserve edges, textures, and fine details, which are important for anatomical boundary recognition. ASPP is further applied to the high-frequency subbands to enhance multi-scale edge and texture representation, enabling more accurate localization of lesions with varying sizes and ambiguous boundaries. Experiments on three public ultrasound datasets, namely BUS-BRA, TN3K, and DDTI, demonstrate that WAU-Net consistently improves segmentation performance over the standard U-Net. WAU-Net achieves Dice scores of 0.8872, 0.7889, and 0.7590 on BUS-BRA, TN3K, and DDTI, respectively. Additional evaluation using IoU, accuracy, sensitivity, specificity, AUC, HD95, computational complexity, and statistical significance analysis further verifies the effectiveness and practical efficiency of the proposed method. All source code and data-splitting settings are publicly available at https://github.com/apple1986/WAU-Net to support reproducibility.