Related Experiment Video
Updated: Jul 10, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Mammography Breast Cancer Screening Triage Using Deep Learning: A UK Retrospective Study
Sarah E Hickman1, Nicholas R Payne1, Richard T Black1
1From the Department of Radiology, University of Cambridge School of Clinical Medicine, Box 218, Level 5, Cambridge Biomedical Campus, Cambridge CB2 0QQ, UK (S.E.H., N.R.P., Y.H., A.N.P., M.N., F.J.G.); University of Cambridge School of Clinical Medicine, Cambridge, UK (M.I.A, A.S.); Department of Radiology, Barts Health NHS Trust, The Royal London Hospital, London, UK (S.E.H.); Department of Radiology, Addenbrooke's Hospital, Cambridge University Hospitals NHS Foundation Trust, Cambridge, UK (R.T.B., A.N.P., F.J.G.); EPSRC Cambridge Mathematics of Information in Healthcare Hub, University of Cambridge, Cambridge, UK (Y.H.); Peel & Schriek Consulting, London, UK (S.H.); Department of Radiology, Norfolk and Norwich University Hospital, Norwich, UK (B.K., A.J.); and University of East Anglia, Norwich Research Park, Norwich, UK (B.K.).
Deep learning (DL) algorithms can effectively triage mammograms, identifying normal results to reduce radiologist workload and flagging potential cancers. This adaptive workflow demonstrated superior sensitivity and noninferior specificity compared to traditional double reading.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Mammography is crucial for early breast cancer detection.
- The high volume of normal mammograms leads to repetitive, resource-intensive reading tasks.
- Deep learning (DL) offers potential solutions for optimizing screening workflows.
Purpose of the Study:
- To evaluate the efficacy of deep learning (DL) algorithms in triaging mammograms.
- To determine if DL can identify normal mammograms for workload reduction.
- To assess if DL can flag cancers that might be overlooked by human readers.
Main Methods:
- Retrospective study of 78,849 mammograms from two UK Breast Screening Program sites.
- Three commercial DL algorithms were tested using two triage scenarios (rule-out and rule-in).
- Performance was assessed using sensitivity and specificity, comparing DL-adapted workflows to routine double reading.
Main Results:
- DL algorithms achieved high sensitivity (0.0%-0.1% missed screening-detected cancers) in rule-out triage.
- In rule-in triage, DL algorithms identified interval and subsequent-round cancers at rates of 4.6%-8.2% and 5.2%-6.1%, respectively.
- The combined DL-adapted workflow (scenario C) showed superior sensitivity (2.7% increase) and noninferior specificity (-0.9% difference) versus double reading.
Conclusions:
- Deep learning-adapted triage workflows can significantly improve mammography screening efficiency.
- DL algorithms enhance the efficacy of breast cancer screening by optimizing workload and detection rates.
- These findings support the integration of DL into routine mammography screening practices.
