Complementary Attention Mechanism for Tumor Feature Integration in Mammography

This paper proposes a complementary attention mechanism for the complementary integration of tumor feature information between the MLO and CC views in mammography. Building upon self-attention and cross-attention mechanisms, we establish our complementary attention mechanism. The complementary attention mechanism learns how to find a balance between the self-attention and cross-attention mechanisms, rather than mechanically connecting the self-attention layer and the cross-attention layer. More specifically, our proposed method focuses on learning the complementary information between the two views and achieving a balanced representation by considering both self-attention and cross-attention.