WMCAF-Net: A Wavelet-Enhanced Mamba and Cross-Scale Adaptive Fusion Network for Breast Ultrasound Image Segmentation

Breast ultrasound image segmentation remains challenging because of low contrast, speckle noise, and blurred lesion boundaries. Although Mamba can model long-range dependencies efficiently, directly processing noisy ultrasound features may introduce unreliable contextual responses. Moreover, conventional skip connections can transmit redundant background information and hinder accurate boundary recovery. To address these issues, we propose WMCAF-Net, a Wavelet-Enhanced Mamba and Cross-Scale Adaptive Fusion Network. The Wave-Mamba module decomposes features into low- and high-frequency components using wavelet transform. Low-frequency components are processed by a visual state-space module to capture stable global structures, while high-frequency components are enhanced by directional convolutions to strengthen boundary and texture representation. The Cross-Scale Adaptive Fusion Module improves encoder–decoder feature fusion through channel-response recalibration and adaptive spatial selection. On BUSI, WMCAF-Net achieves 75.47% IoU and 83.47% Dice, surpassing the strongest competing methods by 2.08 and 1.97 percentage points, respectively. Under five-fold cross-validation on Dataset B, it achieves 72.89% IoU and 82.27% Dice, surpassing the strongest competing methods by 1.53 and 2.57 percentage points, respectively. Furthermore, bidirectional cross-dataset evaluation was conducted to analyze the segmentation performance of WMCAF-Net when transferred across different ultrasound datasets.