Domain Adaptation Based on Adversarial-Learned Loss and Deep Correlation Alignment for Breast Cancer Diagnoses

Breast cancer is the most prevalent and fatal cancer among women globally, with breast ultrasound imaging often being the primary diagnostic method in clinical examinations. However, challenges such as difficulty in acquisition, limited annotation, and inconsistencies in data feature distribution result in low accuracy for both manual and traditional network model-assisted diagnoses. To address these issues, this study proposed a domain adaptation network based on adversarial-learned loss and deep correlation alignment (ALCOR_DA). Building upon the foundational framework of dynamic adversarial adaptation networks (DAANs), the proposed model introduces a novel loss function, the adversarial-learned loss. By leveraging the adversarial interaction between the generator and domain discriminator within the framework, a noise-confusion matrix is generated to refine the pseudo-labels produced by the classifier for target domain data. This process reduces the discrepancy between pseudo-labels and true labels, thereby improving the classification accuracy for unlabeled target domain data. Additionally, the model incorporates an adaptive layer and employs the deep correlation alignment algorithm to measure and align the feature distributions between the source and target domains. This alignment enhances the generalization capability of the model across different datasets. On public datasets, ALCOR_DA achieved an accuracy of 86.26%, representing a 4.65% improvement over the traditional DAAN model, underscoring its effectiveness in practical diagnostic scenarios.