Evaluating AI in breast cancer screening: a complex task
Menée à partir de données portant sur 170 230 clichés mammographiques, cette étude évalue, par rapport à l'interprétation des radiologues, la performance d'un algorithme utilisant les technologies de l'intelligence artificielle pour diagnostiquer avec précision un cancer du sein
Today, artificial intelligence (AI) seems to be making an impact in almost every field of research. Breast cancer screening is one of them. Digital mammography is the current standard modality for early breast cancer detection, but as with every screening tool, it has its limitations. Sensitivity in a given screening round is estimated at 75–85%,1 meaning that a substantial fraction of cancers go undetected. In recent screening studies, digital breast tomosynthesis has found up to 40% more cancers than digital mammography,2, 3, 4 suggesting that previously reported sensitivities for digital mammography are too high. Additionally, the specificity of screening mammography is estimated at 90–95%, which in a screening modality where more than 99% of examined women are healthy means that the vast majority of detections are false positives. Every false positive means that a woman needs to be recalled, re-examined (often with additional tests such as ultrasound imaging), and often biopsied, which, apart from the additional health-care expenditure, can lead to anxiety and other negative psychosocial consequences.
The Lancet Digital Health , commentaire en libre accès, 2019