LS$^{2}$Net: A Lightweight Segmentation Network for Ultrasound Imaging via Synergy of Large and Small Receptive Fields.

Abstract

Ultrasound imaging, with its advantages of portability, real-time feedback, and non-invasiveness, has become an indispensable modality in computer-aided diagnosis. However, inherent challenges in ultrasound images, such as speckle noise, low contrast, and blurry boundaries, significantly hinder segmentation accuracy. Existing methods attempt to alleviate these issues by deepening network architectures, enlarging convolutional kernels, or introducing self-attention mechanisms to enhance feature representation. Nevertheless, these strategies often compromise computational efficiency, making it difficult to achieve a desirable balance between segmentation accuracy and inference speed, thereby limiting their applicability in resource-constrained clinical scenarios such as portable ultrasound devices. To address these challenges, this paper proposes a lightweight and efficient segmentation network, termed LS²Net. In the encoder, to mitigate the impact of speckle noise, LS²Net incorporates wavelet convolution to transform spatial features into the frequency domain for targeted processing, which not only suppresses noise but also enlarges the receptive field and captures global contextual information. Meanwhile, pixel difference convolution is employed to capture fine-grained details under low-contrast conditions, enhancing the perception of local textures. Furthermore, we design a Residual Refinement Skip Connection (RRSC) module, which leverages the discrepancy between down-sampled and up-sampled features to preserve informative components while filtering redundant information, thereby facilitating more effective feature reconstruction. In the decoder, a Multi-Receptive Field Coarse-to-Fine Module (MRCFM) is introduced to further integrate multi-scale contextual information through hybrid receptive fields, enabling more precise delineation of object boundaries. LS²Net demonstrates superior computational efficiency and state-of-the-art (SOTA) performance across five ultrasound datasets, achieving a favorable balance between segmentation accuracy and inference speed. In addition, it delivers competitive performance on other imaging modalities, highlighting its versatility. Moreover, LS²Net exhibits strong generalization capability in cross-dataset evaluations, validating its robustness and scalability. Overall, LS²Net provides a practical and deployable solution for real-time clinical applications, particularly suitable for portable or resource-limited devices. The source code is publicly available at: https://github.com/CYYJL/LS2Net.