EMD‐DAAN: A Wasserstein Distance‐Based Dynamic Adversarial Domain Adaptation Network Model for Breast Ultrasound Image Classification
Ying Te Wu, Hao Huang, Bo Xu · IET Image Processing · 2025
ABSTRACT Breast cancer is commonly diagnosed through ultrasound imaging as a primary method in clinical practice. However, the lack of large annotated datasets for breast ultrasound images, along with issues such as inconsistent edge and conditional distributions across different datasets, poses significant challenges to both manual and AI‐assisted diagnosis. To address these issues, this paper proposes a dynamic adversarial domain adaptation model based on Wasserstein distance (EMD_DAAN). The EMD_DAAN model enhances the existing dynamic adversarial domain adaptation framework by incorporating an adaptive layer, further aligning the feature distributions between the source and target domain datasets. The Wasserstein distance is employed to optimize this adaptive layer, minimizing the distributional discrepancy between the feature spaces of the two domains by constructing the least‐cost transport path. This approach improves the model's cross‐domain generalization ability and robustness to noise interference. Through dual feature alignment via the adaptive layer and adversarial learning, the model's classification performance on breast ultrasound images is significantly enhanced. Experimental results demonstrate that the EMD_DAAN model achieves an accuracy of 82.75% on breast ultrasound images, substantially outperforming typical adversarial domain adaptation models such as DAAN in terms of classification performance.