A Muti-Scale Perception Attention Network for Breast Lesions Segmentation in Ultrasound Images
Ruilin Zhang, Dinghao Guo, Zhijian He, Zhangyu Liu · 2025
Breast tumor segmentation is a critical task in both clinical diagnostics and computer-aided diagnosis (CAD). However, challenges such as variable tumor morphology, similar intensity distributions, and blurred boundaries make accurate segmentation difficult. To address these challenges, the novel ZpjNet model is proposed for breast lesion segmentation. Specifically, we introduce a multi-scale perception attention module$(\mathbf{C Z H})$. The module consists of a method that combines multiscale feature extraction with a hybrid attention mechanism, replacing the traditional convolution operation. Compared with traditional convolution, CZH module use multi-scale feature extraction to obtain feature information from different receptor fields, and use an attention mechanism which combined by spatial attention mechanism and channel attention mechanism to suppress irrelevant features, lock key regions. Different from existing attention mechanisms, CZH module can better retain original spatial information to help the network better cope with tumor lesion segmentation. Extensive experiments demonstrate that ZpjNet outperforms other methods in breast tumor segmentation, achieving Intersection over Union (IoU) and Dice Similarity Coefficient (Dice) values of$\mathbf{6 8. 9 4 \%}$and$\mathbf{8 1. 5 1 \%}$on the BUSI dataset, respectively. Furthermore, the CZH module can be flexibly integrated into existing network frameworks, offering strong scalability.