GPG-Net: graph-prior-guided breast ultrasound image segmentation network

Abstract

Convolutional Neural Networks (CNNs) have demonstrated strong performance in breast ultrasound image segmentation. However, constrained by their inherent local receptive fields, they struggle to effectively capture global contextual information. Although existing studies have attempted to incorporate global feature extraction modules, most of them rely on simple feature-level concatenation or addition, failing to fully exploit the deep relationships between global structures and local features. To address this limitation, we propose a graph-prior-guided breast ultrasound image segmentation network (GPG-Net), which efficiently leverages global structural information while preserving local detail perception. Specifically, we design a graph-prior-guidance (GPG) module and introduce graph convolution network (GCN). By utilizing the global topological structure captured by the GCN as prior information, this module dynamically guides the feature extraction and fusion within the convolutional branch, achieving a deep synergy between global semantics and local details. In addition, a superpixel image augmentation (SIA) module is introduced to enhance the diversity of the training data. We validate the effectiveness and superiority of the proposed model on two public breast ultrasound datasets.