Asymmetry and Architectural Distortion Detection with Limited Mammography Data

Detection of the asymmetry (AS) and architectural distortion (AD) on mammograms is important for early breast cancer diagnosis. However, this is a challenging task because there are very limited mammography data containing these two lesions. In this paper, we tackle this problem by presenting a novel transfer learning framework of Supervised mass-Transferred Pre-training (STP) followed by Supervised Constrained Contrastive Fine-tuning (S $$\mathrm C^2$$ F). While STP can leverage the commonly available mass data to help with detecting the rarely available AS and AD as pre-training, S $$\mathrm C^2$$ F can depart the mass, AS, and AD in the embedding space as far as possible with a carefully designed constrained contrastive loss. In addition, a novel detection network - AsAdNet, is proposed for the AS and AD detection. The validation results on the largest-so-far AS and AD dataset show state-of-the-art (SOTA) detection performance.