HCMNet: A Hybrid CNN-Mamba Network for Breast Ultrasound Segmentation for Consumer Assisted Diagnosis

Accurate segmentation of Breast UltraSound (BUS) images is crucial for enhancing clinical diagnostic efficiency and improving consumers’ experience. The recently proposed Mamba architecture addresses the limitations of Convolutional Neural Networks (CNNs) in global feature extraction and overcomes the quadratic computational complexity of Transformer-based models by introducing 2D Selective Scan (SS2D) with linear complexity. However, it may still face challenges in effectively capturing fine-grained local details. To address the issue, we propose a Hybrid Network (HCMNet) that integrates CNN and Mamba, leveraging CNN’s strengths in local feature extraction while harnessing Mamba’s capabilities in capturing global features. Additionally, we propose a Wavelet Feature Extraction Module (WFEM) that integrates both low-frequency and high-frequency features, significantly enhancing feature representation and effectively mitigating the loss of spatial information during encoder downsampling. In the skip connections, our proposed Adaptive Feature Fusion Module (AFFM) dynamically integrates encoded features with wavelet features, preserving critical information while effectively suppressing feature redundancy and noise interference. Comparative experiments on the BUSI and UDIAT Dataset B show improvements in IoU by 0.99% and 1.07% respectively over state-of-the-art methods. For reproducibility, the code is available at https://github.com/XYQ1517/HCMNet.