AI for enhancing efficiency and effectiveness of population breast cancer screening
Mené en Suède sur 105 934 femmes (âge médian : environ 54 ans), cet essai évalue la non-infériorité, du point de vue du taux de cancer de l'intervalle, de la sensibilité et de la spécificité, d'une mammographie de dépistage assistée par l'intelligence artificielle par rapport à une mammographie avec double lecture standard
Breast cancer screening has long been a target use case for artificial intelligence (AI). Developments in computational power coincided with the adoption of digital mammography in the early 2000s and the generation of large, data-rich, archival imaging datasets suitable for algorithm training and validation. As such, the labour-intensive task of mammography screen-reading to identify suspicious lesions was a model application of AI automation, with potential for improving screening effectiveness and simultaneously reducing screening workload burden.1 Consequently, the last decade has seen rapid growth in AI algorithm development.2 Against a backdrop of disappointing outcomes from early adoption of pre-contemporary AI software for mammographic interpretation,3 evaluation of more recent, deep learning-based AI models has progressed swiftly to inform evidence-based implementation: from studies of AI accuracy in selected cancer-enriched datasets,4 to retrospective studies reporting screening metrics using representative cohorts of consecutive participants,5,6 and now the first randomised controlled trial (RCT) of AI embedded into a population breast cancer screening programme.
The Lancet , commentaire, 2026