MFS-Net: Multi-Stage Feature Fusion and Shape Fitting Network for Ultrasound Image Segmentation

Most ultrasound image segmentation methods employ a joint training strategy, where the model learns the edge and core regions indiscriminately and uniformly. However, they often overlook the distinct shapes of the edges and the sizes of the core regions. In this paper, we propose an innovative Multi-Stage Feature Fusion and Shape Fitting Network for Ultrasound Image Segmentation (MFS-Net) that includes a core segmentation stage, an edge enhancement stage, and a fusion correction stage. The first two stages focus on learning the spatial features of the core region and the texture features of the edge region, respectively. After two stages of learning, the core and edge features of the labeled area are then fused in the third stage. In the third stage, we proposed a shape correction loss function to correct the shape of the labeled regions. To validate the superiority of our proposed MFS-Net, we conducted extensive experiments on our self-constructed Parathyroid Segmentation Dataset (PUS) and three public datasets. The results demonstrate substantial improvements over mainstream methods across all four datasets.