S²DENet: Shallow suppression and deep enhancement network for general ultrasound image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Ultrasound image segmentation is crucial for clinical diagnosis but challenged by noise and ambiguous boundaries.
• Current methods use uniform feature extraction, leading to inefficiency and over-parameterization.
• This approach struggles to balance noise suppression and feature enhancement across network depths.

Purpose of the Study:

• To develop an efficient network architecture for ultrasound image segmentation.
• To address limitations of uniform feature extraction in deep learning models.
• To improve accuracy and computational efficiency in resource-constrained environments.

Main Methods:

• Proposed S2DENet, a network with depth-differentiated noise suppression (shallow layers) and semantic feature enhancement (deep layers).
• Introduced Multi-order Differential Convolution (MDiffConv) for high-frequency feature capture and differential self-attention (DiffSA) for noise mitigation.
• Implemented depth-differentiated strategies to reduce learning complexity and enhance feature representation.

Main Results:

• S2DENet achieved state-of-the-art performance on nine out of ten public ultrasound datasets.
• The model boasts minimal parameters (0.05M/0.15M), over 99% reduction compared to existing methods.
• Real-time inference at 80+ FPS was maintained on high-end GPUs with competitive accuracy.

Conclusions:

• S2DENet offers a novel, efficient, and accurate solution for ultrasound image segmentation.
• The depth-differentiated approach optimizes performance for clinical applications.
• This work paves the way for practical deployment of advanced segmentation models in clinical settings.