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

Ultrasound image segmentation serves as a cornerstone of clinical diagnosis, yet remains a formidable challenge due to inherent image artifacts such as speckle noise and ambiguous boundaries. Existing approaches typically employ uniform feature extraction strategies across all network layers, disregarding the fundamental disparities between noise-dominated shallow stages and semantically-rich deep stages. This monolithic strategy compels networks to concurrently learn noise suppression and feature enhancement, often resulting in over-parameterization and critical compromises in computational efficiency, particularly within resource-constrained environments. To address these limitations, we propose S²DENet, an efficient network architecture that employs noise suppression in shallow layers and semantic feature enhancement in deep layers. Specifically, we design Multi-order Differential Convolution (MDiffConv) to enhance high-frequency feature capture, implementing suppression strategies in shallow layers and enhancement strategies in deep layers. Simultaneously, we introduce a Differential Self-Attention (DiffSA) mechanism to mitigate noise artifacts while preserving structural integrity, with deep layers transitioning to standard self-attention for semantic feature amplification. This depth-differentiated design, combined with differential mechanisms, enables the model to focus on specific tasks at different network stages, thereby reducing learning complexity and improving feature representation efficiency. Experimental results on ten public ultrasound datasets demonstrate that S²DENet achieves an exceptional balance between efficiency and accuracy, attaining state-of-the-art (SOTA) performance on nine public datasets. With merely 0.05M/0.15M parameters (a reduction exceeding 99% compared to SOTA methods), S²DENet maintains real-time inference at 80+ FPS on an NVIDIA RTX A8000 GPU while achieving competitive or superior performance. S²DENet not only establishes a new paradigm for ultrasound image segmentation but also opens new possibilities for deployment in clinical practice. Code is available at https://github.com/PXinTao/S2DENet.