Protocol for evaluating the fitness for purpose of an artificial intelligence product for radiology reporting in the BreastScreen New South Wales breast cancer screening programme
Files
(Published version)
Date
2024
Authors
Warner Smith, M.
Ren, K.
Mistry, C.
Walton, R.
Roder, D.
Bhola, N.
McGill, S.
O'Brien, T.A.
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Journal article
Citation
BMJ Open, 2024; 14(5):1-7
Statement of Responsibility
Conference Name
Abstract
Introduction: Radiologist shortages threaten the sustainability of breast cancer screening programmes. Artificial intelligence (AI) products that can interpret mammograms could mitigate this risk. While previous studies have suggested this technology has accuracy comparable to radiologists most have been limited by using ‘enriched’ datasets and/or not considering the interaction between the algorithm and human readers. This study will address these limitations by comparing the accuracy of a workflow using AI alongside radiologists on a large consecutive cohort of examinations from a breast cancer screening programme. The study will combine the strengths of a large retrospective design with the benefit of prospective data collection. It will test this technology without risk to screening programme participants nor the need to wait for follow-up data. With a sample of 2 years of consecutive screening examinations, it is likely the largest test of this technology to date. The study will help determine whether this technology can safely be introduced into the BreastScreen New South Wales (NSW) population-based screening programme to address radiology workforce risks without compromising cancer detection rates or increasing false-positive recalls.
Methods and analysis: A retrospective, consecutive cohort of digital mammography screens from 658 207 examinations from BreastScreen NSW will be reinterpreted by the Lunit Insight MMG AI product. The cohort includes 4383 screen-detected and 1171 interval cancers. The results will be compared with radiologist single reading and the AI results will also be used to replace the second reader in a double-reading model. New adjudication reading will be performed where the AI disagrees with the first reader. Recall rates and cancer detection rates of combined AI–radiologist reading will be compared with the rates obtained at the time of screening.
Ethics and dissemination: This study has ethical approval from the NSW Health Population Health Services Research Ethics Committee (2022/ETH02397). Findings will be published in peer-reviewed journals and presented at conferences. The findings of this evaluation will be provided to programme managers, governance bodies and other stakeholders in Australian breast cancer screening programmes.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2024 Author(s) (or their employer(s)). Re-use permitted under CC BY-NC. No commercial re-use. This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. (https://creativecommons.org/licenses/by-nc/4.0/.)