Breast Ultrasound Image Classification Based on Dynamic Adversarial Adaptive Network with Triplet Loss
Ying Te Wu, Kaicheng Lin, Hao Huang, Faming Li, Bo Xu · 2025
In medical imaging, the acquisition and annotation of ultrasound image data require the expertise of medical professionals, resulting in substantial costs and posing challenges to constructing large-scale annotated datasets. Faced with these difficulties we construct a dynamic adversarial adaptive network model with Triplet Loss (TriLDA). The model captures invariant features of domain through the generator and adjusts the distributions across domain images and target domain images by employing global and local domain discriminators, and integrates the Triplet Loss function within the classifier layer to optimize training. This approach enhances TriLDA's performance to distinguish between classes by increasing the feature distance between samples of different categories while minimizing the distance within samples of the same category. Experimental evaluations demonstrate that the TriLDA model achieves superior performance compared to standard unsupervised domain adaptation models, including DAN, MADA, and DAAN, in the task of breast ultrasound image classification.