Send to

Choose Destination
NPJ Breast Cancer. 2019 Nov 19;5:43. doi: 10.1038/s41523-019-0134-6. eCollection 2019.

Application of convolutional neural networks to breast biopsies to delineate tissue correlates of mammographic breast density.

Author information

1Division of Population Health Sciences, Royal College of Surgeons in Ireland, Dublin, Ireland.
2Division of Cancer Epidemiology and Genetics, National Cancer Institute, Bethesda, MD USA.
3Department of Pathology, Radboud University Medical Center Nijmegen, Nijmegen, the Netherlands.
4Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA USA.
5University of Vermont and University of Vermont Cancer Center, Burlington, VT USA.
6University of California, San Francisco, San Francisco, CA USA.
7University of Hawaii Cancer Center, Honolulu, HI USA.
8Department of Radiation Medicine, Hokkaido University Graduate School of Medicine, Sapporo, Hokkaido, Japan.
9The University of Texas MD Anderson Cancer Center, Houston, TX USA.
10Center for Cancer Research, National Cancer Institute, Bethesda, MD USA.
11Mayo Clinic, Jacksonville, FL USA.
Contributed equally


Breast density, a breast cancer risk factor, is a radiologic feature that reflects fibroglandular tissue content relative to breast area or volume. Its histology is incompletely characterized. Here we use deep learning approaches to identify histologic correlates in radiologically-guided biopsies that may underlie breast density and distinguish cancer among women with elevated and low density. We evaluated hematoxylin and eosin (H&E)-stained digitized images from image-guided breast biopsies (n = 852 patients). Breast density was assessed as global and localized fibroglandular volume (%). A convolutional neural network characterized H&E composition. In total 37 features were extracted from the network output, describing tissue quantities and morphological structure. A random forest regression model was trained to identify correlates most predictive of fibroglandular volume (n = 588). Correlations between predicted and radiologically quantified fibroglandular volume were assessed in 264 independent patients. A second random forest classifier was trained to predict diagnosis (invasive vs. benign); performance was assessed using area under receiver-operating characteristics curves (AUC). Using extracted features, regression models predicted global (r = 0.94) and localized (r = 0.93) fibroglandular volume, with fat and non-fatty stromal content representing the strongest correlates, followed by epithelial organization rather than quantity. For predicting cancer among high and low fibroglandular volume, the classifier achieved AUCs of 0.92 and 0.84, respectively, with epithelial organizational features ranking most important. These results suggest non-fatty stroma, fat tissue quantities and epithelial region organization predict fibroglandular volume. The model holds promise for identifying histological correlates of cancer risk in patients with high and low density and warrants further evaluation.


Cancer epidemiology; Cancer prevention

Conflict of interest statement

Competing interestsThe following authors have competing interests to disclose: Dr. Andrew Beck is an employee and equity holder of PathAI. Dr. Sally D. Herschorn is on the Medical Advisory Board of, an education coalition about breast density. Dr. Jeroen van der Laak is member of the scientific advisory board of Philips, the Netherlands, is a member of the scientific advisory board of ContextVision, Sweden, has received research funding from Philips, the Netherlands and research funding from Sectra, Sweden. Dr Nico Karssemeijer reported receiving holding shares in Volpara Solutions, QView Medical, and ScreenPoint Medical BV; consulting fees from QView Medical; and being an employee of ScreenPoint Medical BV. Remaining authors have no competing interests.

Supplemental Content

Full text links

Icon for PubMed Central
Loading ...
Support Center