Domain Generalization for Mammography Detection via Multi-style and Multi-view Contrastive Learning

Lesion detection is a fundamental problem in the computer-aided diagnosis scheme for mammography. The advance of deep learning techniques have made remarkable progress this task, provided that training data are large and sufficiently diverse terms image style quality. In particular, diversity may be majorly attributed to vendor factor. However, collection mammograms from vendors as many possible very expensive sometimes impractical laboratory-scale studies. Accordingly, further augment generalization capability model various with limited resources, new contrastive developed. Specifically, backbone network firstly trained multi-style multi-view unsupervised self-learning embedding invariant features vendor-styles. Afterward, then recalibrated downstream task lesion specific supervised learning. proposed method evaluated four one unseen public dataset. experimental results suggest our approach can effectively improve performance on both seen domains, outperforms state-of-the-art (SOTA) methods.