2026/07/01 by Tianren Wu, Rundong Zhao, Zijun Hou +8
paper · doi:10.1002/ima.70411
ABSTRACT Computer‐aided diagnosis as a second opinion has become increasingly common in the field of medical image segmentation. Clinical ultrasound has become an important tool for disease screening due to its convenience, cost‐effectiveness, and safety. However, tumors in breast ultrasound images exhibit characteristics such as diverse morphology, blurred edges, and similar intensity distributions, making it difficult for traditional segmentation methods to accurately locate lesions. To address this issue, our proposed position‐focused hybrid structured network (PFHNet) integrates transformer and graph convolutional network (GCN) components to jointly guide position information for breast ultrasound image segmentation. PFHNet consists of two parts: a convolution branch (CB) and a transformer branch (TB), where CB focuses more on detailed features and TB emphasizes global semantic information. In CB, we design a detail feature supplementary self‐attention module (DSAM) and a hybrid convolutional decoder (HCD). DSAM uses a self‐attention mechanism to model multi‐dimensional information in shallow and deep features while establishing long‐range dependencies. HCD exploits the flexible spatial node aggregation capability of graph convolution so that the decoder can accurately focus on the precise location of the lesion while recovering detailed information. In TB, we propose a position feature guidance module (PFGM) to guide the CB decoding process by highlighting the tumor position in the feature map using deformable convolution and grid sampling. We conduct rigorous evaluations on two public datasets (BUSI and Dataset B), and the results demonstrate that our model outperforms nine other state‐of‐the‐art segmentation methods.