ideasIdea

AI second reads to stop borderline lesions being upgraded to cancer

Whether a lesion is called precancer or cancer varies between pathologists, and over time the bar has drifted lower. AI reference reads could hold the line.

Inter-observer disagreement is high for DCIS versus atypia, Gleason pattern 3 versus 4, and melanocytic lesions; diagnostic drift inflates incidence. Propose AI reference classifiers calibrated to historical outcome-linked cohorts, used as mandatory second reads for borderline categories, with discordance triggering expert review.

Hypothesis

AI second reads reduce upgrade rates of borderline lesions by at least 20% and reduce inter-laboratory variation without increasing subsequent invasive cancer.

Rationale

Digital pathology AI already matches expert Gleason grading; anchoring to outcome-linked training sets counters drift.

What would test it

Multi-laboratory study comparing diagnosis rates with and without AI second read, with five-year outcome linkage.

Maturity

early clinical

Who has to act

clinic

Cost to try

Medium ($1M to $50M)

Years to first evidence

Bottlenecks it attacks

Overdiagnosis and false alarms · Finding more cancer is not the same as saving lives. Screening also finds cancers that would never have hurt anyone, and treats them.
AI that is built but not validated or deployed · Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients.

Connected

7top

Pages like this

not linked directly; found by shared links

AI second reads to stop borderline lesions being upgraded to cancer

Connected

Pages like this

fronts

technologies

companies

bottlenecks