Abstract
METHODS
In this retrospective cohort study, 42 236 consecutive 2D mammograms from 42 100 women attending the Dutch population-based breast cancer screening between Sept 1, 2016, and Aug 31, 2018 were processed by an AI-based cancer detection system (Transpara version 1.7.0, ScreenPoint Medical). Verified outcomes from the Netherlands Cancer Registry on screen-detected cancers, interval cancers, and later-in-future-detected breast cancers were available with 4-year follow-up. We compared sensitivity, specificity, and recall-rate between single human reading, double human reading, stand-alone AI reading, and combined single human reading with AI. Furthermore, we assessed potential differences in performance regarding breast density, tumour size, lymph-node positivity, and invasiveness between cancers identified by single human readers and AI alone.
BACKGROUND
Breast cancer screening programmes have shown to reduce mortality, but current methods face challenges such as limited mammographic sensitivity, limited resources, and variability in radiologist expertise. Artificial intelligence (AI) offers potential to improve screening accuracy and efficiency. This study simulated different screening scenarios, evaluating the performance of population-based breast cancer screening when using an AI system as a stand-alone reader or second reader.
FUNDING
MARBLE.
FINDINGS
After follow-up, 580 mammograms (579 woman) were labelled positive: 291 screen-detected cancers, 102 interval cancers, and 187 future breast cancers. Double human reading recalled 1244 mammograms (2·9%, 291 screen-detected cancers) and combined single human reading with AI recalled 2112 mammograms (5·0%, 282 screen-detected cancers, 29 interval cancers, 38 future breast cancers), improving the sensitivity by 8·4% (95% CI 5·7-11·2, p<0·0001). No significant difference in performance between combined single human reading with AI across density categories was found. AI-detected future breast cancers and interval cancers missed by human readers were more often invasive cancers (26·7%) or tumours larger than 20 mm in diameter (16·6%) by the time of eventual detection compared with the average screen-detected cancers.
INTERPRETATION
Evaluating screening mammograms with one human reader and AI leads to increased breast cancer detection compared with double human reading, independent of breast density. However, an effective arbitration process is needed as the recall rate increases. AI-identified breast cancers that are missed by human readers seem larger and more often invasive by the time they are eventually detected, confirming the clinical relevance of these cases, recognisable by AI at an earlier stage.