SDFS-Net: A Breast Ultrasound Segmentation Network combining Static–Dynamic Receptive Field and Frequency-Selective Modeling

Abstract

Breast ultrasound (BUS) plays a critical role in breast cancer diagnosis. However, the large-scale variability of breast tumors poses challenges to receptive field modeling in segmentation methods. Existing approaches are often constrained in their receptive field design, making it challenging to coordinate dynamic and static receptive fields effectively. Moreover, benign and malignant breast tumors exhibit substantial differences in morphological characteristics and structural complexity. By leveraging frequency-domain analysis, this paper finds that these differences are clearly reflected in their spectral distributions. Nevertheless, most methods neglect such spectral heterogeneity. To address these challenges, this paper proposes a novel breast ultrasound segmentation network, termed SDFS-Net. The network adopts the Static-Dynamic Receptive Field (SDRF) block as its visual backbone, which integrates static and dynamic receptive fields through a dual-branch design for complementary and collaborative modeling. Additionally, we design a Frequency-Selective Modeling (FSM) module to selectively model the spectral information of different types of breast tumors. On three breast ultrasound datasets (PACE 2025, BUI, and BUS_UC), SDFS-Net achieves Dice scores of 84.04%, 87.15%, and 92.43%, outperforming the recent state-of-the-art method H-vmunet by 6.91%, 0.29%, and 0.52%, respectively, validating the effectiveness of static-dynamic receptive field collaboration and frequency-selective modeling.