A dual-branch encoder context-aware fusion network for ultrasound image segmentation

Accurate segmentation of lesion regions in ultrasound images remains a challenging task. Recent research has focused on integrating Transformers and CNNs to leverage their complementary strengths. However, most existing methods employ coarse fusion strategies that often lead to the loss of critical local details, such as lesion boundaries. Additionally, these methods fail to fully leverage the Transformer’s capability for global context modeling, thereby limiting their effectiveness in enhancing comprehensive feature representation. To this end, we propose a dual-branch encoder context-aware fusion network (DECF-Net) for automatic and robust lesion segmentation. The network introduces a parallel dual-branch encoder architecture to simultaneously capture global information and maintain sensitivity to the low-level context. We present a progressive feature extraction (PFE) module suitable for the Transformer branch, which aims to effectively suppress clutter noise and emphasize local features. In order to facilitate the interaction and fusion of feature information between different branches, we further introduce a supplementary feature fusion (SFF) module. In addition, we present a spatial channel attention bridge (SCAB) module to enhance the features of skip connections, which can extract multi-stage and multi-scale context information. Experimental results show that DECF-Net exhibits competitive segmentation performance in both qualitative and quantitative evaluation.