Unsupervised Domain Adaptation for Cross-domain Histopathology Image Classification

Unsupervised domain adaptation (UDA) methods have made remarkable progress in histopathological image analysis and various cancer diagnosis domains. However, most cur-rent research focuses on transfer between single-source domains. The distribution of features between different cancer types is far away, and a well-trained model in one field may not be able to generalize well to data in other fields. To address the domain shift problem, this paper proposes a multi-source unsupervised domain adaptation method with domain mixing bridging. Using multiple source and target domains, feature representations of all domains are extracted, and latent relationships are captured. Afterward, the complementary information of the hybrid bridging intermediate domain is integrated to align the feature distribution. Addi-tionally, we introduce a domain adversarial adaptation module to generate domain invariant features. We experimented on three different cancer pathology image datasets and achieved an average accuracy of 92.94% classification performance. It is proved that compared with the existing deep transfer learning technology, the method in this paper has a better effect. Code will be available at: https://github.com/Ww994/MHDAN.