Thirty percent of interval breast cancers, diagnosed between routine screening mammograms, have a poorer prognosis than screen-detected cancers. Deep learning algorithms can estimate short-term risk from negative mammograms to guide supplemental imaging or screening intervals, but comparative validation on complete national screening data is lacking. We retrospectively evaluated four risk algorithms (Mirai, iCAD, Transpara, and Google) using 112,621 negative mammograms from two UK NHS Breast Screening Programme sites with different mammography systems (Philips, GE) over one screening round (2014-2017) with five-year follow-up, including 1225 future cancers. There was a distinct ranking in discriminative ability; overall AUCs ranged 0.65-0.72, only one algorithm significantly differed between systems. For interval cancers, AUCs ranged 0.67-0.77. Within the highest 4.0% of risk scores, top algorithms identified ~20% of future cancers, including ~27% of interval cancers, doubling at the 14.0% threshold. These differences highlight the need for multi-algorithm prospective trials and potential fine-tuning to improve generalisation across unseen systems.
Rothwell et al. (Sun,) studied this question.