Artificial Intelligence to Support Independent Assessment of Screening Mammograms—The Time Has Come
Menée à l'aide de clichés mammographiques réalisés auprès de 739 patientes atteintes d'un cancer du sein et auprès de 8 066 témoins en bonne santé (âge : de 40 à 74 ans), cette étude compare les performances de trois logiciels disponibles dans le commerce et basés sur les technologies de l'intelligence artificielle pour aider les radiologues à interpréter les clichés mammographiques et améliorer la détection des lésions cancéreuses
Screening mammography is our best method currently available to detect breast cancer early, when it can be cured. However, global access to high-quality, affordable screening mammography is constrained by the limited supply of radiologists subspecialized in breast imaging to interpret each individual examination. The need for interpretation of each mammogram by a subspecialist not only increases costs and limits access to screening but also adds the element of human error to even the most advanced screening programs. Owing to well-documented human error and variation, there is no “diagnostic accuracy” of screening mammography but rather a wide range of performance outcomes based on the individual radiologist interpreting the mammogram. In a study of more than 1.6 million modern, all-digital screening mammograms, investigators of the Breast Cancer Surveillance Consortium found a wide range of interpretive performance across radiologists, with more than 40% of certified, specialized radiologists failing to meet recommended recall rates. Recognition of these challenges supported early efforts to develop deep learning models to assist humans in mammographic interpretation. However, the outcomes have been mixed, with wide variation in quantity and quality of data used for model development, variable methods to train, test, and internally and externally validate models developed, and inconsistent use of peer-reviewed publications to share discoveries.
JAMA Oncology , éditorial, 2019