Related Experiment Video
Updated: Jul 27, 2025
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
Impact of Different Mammography Systems on Artificial Intelligence Performance in Breast Cancer Screening
Clarisse F de Vries1, Samantha J Colosimo1, Roger T Staff1
1From the Aberdeen Centre for Health Data Science, Institute of Applied Health Sciences (C.F.d.V., M.B., L.A.A.), School of Medicine, Medical Science and Nutrition (S.J.C., R.T.S.), and Grampian Data Safe Haven (DaSH), Aberdeen Centre for Health Data Science, Institute of Applied Health Sciences (J.A.D.), University of Aberdeen, Polwarth Building, Foresterhill, Aberdeen AB24 3FX, Scotland; National Health Service Grampian (NHSG), Aberdeen Royal Infirmary, Aberdeen, Scotland (S.J.C., R.T.S., G.L.); Kheiron Medical Technologies, London, England (J.Y., D.D.); and School of Medicine, University of St Andrews, St Andrews, Scotland (D.J.H.).
Artificial intelligence (AI) in breast screening mammography shows promise but requires validation. AI performance varied with different thresholds and software versions, necessitating careful site-specific evaluation before widespread adoption.
Area of Science:
- Radiology
- Medical Imaging
- Artificial Intelligence
Background:
- Artificial intelligence (AI) tools are being developed to aid breast screening mammography.
- Evidence regarding the generalizability of AI algorithms to new clinical settings is limited.
- Evaluating AI performance in diverse settings is crucial for safe and effective implementation.
Purpose of the Study:
- To assess the performance and generalizability of a commercial breast screening AI algorithm in a U.K. regional screening program.
- To determine if AI performance is transferable to a new clinical site using prespecified and site-specific thresholds.
- To investigate the impact of software upgrades on AI algorithm performance and required thresholds.
Main Methods:
- Retrospective analysis of a 3-year dataset (April 2016-March 2019) from a U.K. regional screening program.
- Inclusion of 55,916 women aged 50-70 years attending routine screening.
- Assessment of a commercial breast screening AI algorithm using prespecified and calibrated decision thresholds, considering different mammography equipment software versions.
Main Results:
- A prespecified AI threshold led to high recall rates (48.3%), reduced to 13.0% after calibration, approaching the service level of 5.0%.
- Mammography software upgrades necessitated per-software version thresholds, increasing recall rates threefold.
- Using software-specific thresholds, the AI recalled 91.4% of screen-detected cancers and 34.1% of interval cancers.
Conclusions:
- AI performance in breast screening mammography is sensitive to decision thresholds and mammography equipment software versions.
- Validation of AI algorithms and their thresholds in new clinical settings is essential before deployment.
- Ongoing quality assurance is necessary to monitor AI performance consistency in real-world screening programs.
More Related Videos
08:32Using Computer-based Image Analysis to Improve Quantification of Lung Metastasis in the 4T1 Breast Cancer Model
Published on: October 2, 2020
15:48Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014