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

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Ultrasound imaging is crucial for computer-aided diagnosis but faces challenges like speckle noise and low contrast.
• Existing segmentation methods often struggle to balance accuracy and computational efficiency, limiting use in resource-constrained settings.

Purpose of the Study:

• To develop a lightweight and efficient segmentation network (LS²Net) for improved ultrasound image analysis.
• To address limitations of existing methods in achieving both high accuracy and fast inference speeds.

Main Methods:

• LS²Net employs wavelet convolution in the encoder to reduce speckle noise and capture global context.
• Pixel difference convolution enhances detail perception in low-contrast regions.
• A Residual Refinement Skip Connection (RRSC) module and Multi-Receptive Field Coarse-to-Fine Module (MRCFM) improve feature reconstruction and boundary delineation.

Main Results:

• LS²Net achieves state-of-the-art segmentation accuracy with superior computational efficiency across five ultrasound datasets.
• The network demonstrates competitive performance on other imaging modalities and strong cross-dataset generalization.
• Achieves a favorable balance between segmentation accuracy and inference speed.

Conclusions:

• LS²Net offers a practical and deployable solution for real-time clinical ultrasound applications, especially on portable devices.
• The model's efficiency, accuracy, and robustness make it suitable for resource-limited environments.
• Publicly available source code facilitates further research and application.