MSRA-Net: A multi-scale and region-aware network for breast cancer ultrasound image segmentation

Automated analysis of breast ultrasound images holds significant potential to improve the accuracy of early breast cancer diagnosis, enabling physicians to rapidly and precisely identify lesion areas and providing timely, scientifically grounded decision support for clinical treatment. However, the inherent challenges of breast ultrasound images—such as speckle noise, blurred lesion boundaries, and heterogeneous gray-scale distributions—make accurate lesion extraction difficult for traditional segmentation methods. Although deep convolutional neural networks (CNNs) have achieved remarkable progress in medical image segmentation, their limited local receptive fields often result in insufficient modeling of long-range spatial dependencies, hindering their ability to effectively handle the complex and variable morphology of breast lesions. To address these challenges, this study proposes a novel multi-scale and region-aware network (MSRA-Net) for breast cancer ultrasound image segmentation. In the encoder stage, the model incorporates a Multi-Scale Feature Extraction Module (MFEM), which leverages wavelet convolution (WTConv) with a large receptive field to efficiently capture morphological features of lesions at multiple scales. In the decoder stage, the model innovatively integrates a Global Region-Aware Block (GRAB) and a Boundary Feature Enhancement Block (BFEB). The GRAB employs Space-Adaptive Channel Reduction Attention (SCRA) to focus on the global features of lesions, while the BFEB enhances boundary depiction accuracy by separating and processing low-frequency and high-frequency features. Extensive experiments on three breast cancer ultrasound datasets, BUSI, BUS-BRA, and BUET_BUSD, demonstrate that the proposed network significantly outperforms state-of-the-art medical image segmentation methods for breast ultrasound lesion segmentation. Furthermore, ablation studies validate the effectiveness of the individual modules and underscore the robustness and clinical utility of the proposed approach.