M2NET: Two-Stage Multi-Label Breast Cancer Detection Networks

Mammograms are widely used for breast cancer screening, diagnosis, and follow-ups. The Breast Imaging Reporting and Data System (BI-RADS) provides standardized terminology, reporting structure, and classification for mammography findings. Certain characteristics, such as shape, margins, calcifications, asymmetry, and architectural distortion, can indicate different BI-RADS scores corresponding to malignancy. Leveraging multi-label learning, this paper introduces M2Net, a Two-stage Multi-label Breast Cancer Detection Network that simultaneously addresses lesion localization, lesion type and BI-RADS detection. Additionally, inspired by how radiologists examine mammograms, we propose the sliding windows approach to the training pipeline to enhance breast cancer diagnostic performance. M2Net outperforms single-label models in comprehensive lesion detection and precise BI-RADS detection in an In-house and the CBIS-DDSM dataset. This work represents a novel strategy that bridges AI and clinical practice for more accurate breast malignancy diagnosis.