AsyDisNet: Scalable Mammographic Asymmetry and Architectural Distortion Detection With Angle-Based Quadruplet Loss

Early asymmetry (AS) and architectural distortion (AD) detection on mammograms are essential in breast cancer diagnosis. However, they are challenging as the prevalence of AS and AD is very low. This paper proposes an efficient AsyDisNet for the AS and AD detection. First, a novel angle-based quadruplet loss is proposed to detect the AS and AD with limited pixel-level labeled mammograms. Second, we scale the AsyDisNet with a novel semi-weakly supervised learning framework to boost the detection performance with a large number of mammograms with image-level labels extracted from medical reports. The validation on the two largest and privately collected datasets shows an average of $\sim ~10$ % improvement over State-of-the-Art baselines in terms of sensitivities under various false-positive-per-image (FPPI). Furthermore, the proposed AsyDisNet is scalable to the current Picture Archiving and Communication System (PACS) with incremental learning ability. The dataset will be made publicly available at https://github.com/ML-AILab/AsyDisNet.