Related Experiment Video
Updated: Oct 16, 2025
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Machine Learning for Workflow Applications in Screening Mammography: Systematic Review and Meta-Analysis
Sarah E Hickman1, Ramona Woitek1, Elizabeth Phuong Vi Le1
1From the Department of Radiology (S.E.H., R.W., G.C.B., J.W.M., F.J.G.) and Department of Medicine (E.P.V.L., Y.R.I., C.M.L.), University of Cambridge School of Clinical Medicine, Box 218, Cambridge Biomedical Campus, Cambridge, CB2 0QQ, England; Department of Radiology, Addenbrooke's Hospital, Cambridge University Hospitals National Health Service Foundation Trust, Cambridge, England (R.W., F.J.G.); Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria (R.W.); Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Cambridge, England (A.I.A.R.); and Norwich Medical School, University of East Anglia, Norwich, England (J.W.M.).
Machine learning (ML) algorithms show promise in mammographic screening, matching or surpassing human reader performance for detection and improving efficiency. However, further prospective validation is needed to confirm these findings from retrospective studies.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Advances in computing power and data availability have spurred the development of machine learning (ML) for mammography.
- ML techniques are increasingly being applied to mammographic imaging analysis.
Purpose of the Study:
- To evaluate the performance of standalone machine learning (ML) applications within screening mammography workflows.
- Assessing the efficacy of ML algorithms used independently of human readers.
Main Methods:
- A systematic literature search was conducted across multiple databases (Ovid Embase, Medline, Cochrane, Scopus, Web of Science) for studies from January 2012 to September 2020.
- Included studies were assessed for quality using established tools ( குவாலிட்டி அசஸ்மென்ட் ஆஃப் டயக்னாஸ்டிக் அக்யூரசி ஸ்டடீஸ் 2, பிரிடிக்ஷன் மாடல் ரிஸ்க் ஆஃப் பயாஸ் அசஸ்மென்ட் டூல்).
- A primary meta-analysis calculated pooled estimates for the area under the receiver operating characteristic curve (AUC) for both ML algorithms and human readers.
Main Results:
- Fourteen articles detailing 15 studies (8 detection, 7 triage) were included.
- Triage studies indicated that ML could read 17%-91% of normal mammograms while missing 0%-7% of cancers.
- The meta-analysis (185,252 cases, >39 readers) showed pooled AUCs of 0.89 for ML algorithms and 0.85 for human readers, with comparable sensitivity and specificity.
Conclusions:
- Standalone ML algorithms in mammography screening can achieve or exceed human reader performance and enhance efficiency.
- Current evidence stems from a limited number of retrospective studies.
- Rigorous, independent, external prospective testing is essential to validate ML algorithm performance at pre-defined thresholds.
