Developing a weakly supervised deep learning framework for breast cancer diagnosis with HR status based on mammography images

Mengyan Zhang; Cong Wang; Li Cai; Jiyun Zhao; Ye Xu; Jiacheng Xing; Jianghong Sun; Yan Zhang

doi:10.1016/j.csbj.2023.08.012

Developing a weakly supervised deep learning framework for breast cancer diagnosis with HR status based on mammography images

Comput Struct Biotechnol J. 2023 Aug 18:22:17-26. doi: 10.1016/j.csbj.2023.08.012. eCollection 2023.

Authors

Mengyan Zhang¹, Cong Wang¹, Li Cai², Jiyun Zhao¹, Ye Xu³, Jiacheng Xing⁴, Jianghong Sun³, Yan Zhang^{1

5}

Affiliations

¹ School of Life Science and Technology, Harbin Institute of Technology, Harbin 150001, China.
² The Fourth Department of Medical Oncology, Harbin Medical University Cancer Hospital, 150040, China.
³ Department of Radiology, Harbin Medical University Cancer Hospital, Harbin 150081, China.
⁴ Beidahuang Industry Group General Hospital, 150060 Harbin, China.
⁵ College of pathology, Qiqihar Medical University, Qiqihar 161042, China.

Abstract

The status of hormone receptors (HR) at the molecular level is crucial for accurate diagnosis and effective treatment of breast cancer. Meanwhile, mammography is an effective screening method for detecting breast cancer, which significantly improve survival. However, diagnosing the molecular status of breast cancer involves a pathological biopsy, which can affect the accuracy of the diagnosis. To non-invasively diagnose the hormone receptor (HR) status of breast cancer and reduced manual annotation, we proposed a weakly supervised deep learning framework BSNet which detected breast cancer with HR status and benign tumors. BSNet was trained on 2321 multi-view mammography cases from female undergoing digital mammography for the general population at Harbin Medical University Cancer Hospital in Heilongjiang Province during the period 2017-2018 and was validated on the external cohort. The average AUCs of BSNet on the test set and the external validation set were 0.89 and 0.92, respectively. BSNet demonstrated excellent performance in non-invasive breast cancer diagnosis with HR status, using multiple mammography views without pixel annotation. Furthermore, we developed a web server (http://bsnet.edbc.org) for easy use. BSNet described high-dimensional mammography of breast cancer subtypes, which helped inform early management options.

Keywords: Breast cancer diagnosis; Deep learning; HR status; Mammography images; Weakly supervision.