Require stage-shift or interval-cancer endpoints for AI in cancer screening
AI for screening should be judged on whether it finds dangerous cancers earlier and misses fewer, not just on whether it agrees with radiologists on old images.
AI in mammography, lung CT and colonoscopy is evaluated on retrospective detection metrics that reward finding more lesions regardless of clinical significance, which risks overdiagnosis. The proposal requires, for adoption in organised screening programmes, evidence on interval cancer rates, stage distribution of detected cancers and recall rates from prospective studies (randomised or well-designed stepped implementations), with post-implementation monitoring of the same endpoints via registry linkage.
- 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.
- 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.
- Most lethal cancers are found late · Screening exists for only a few cancers. Pancreatic, ovarian, liver, oesophageal and most lung cancers are found when cure is unlikely.
Pages like this
not linked directly; found by shared links- IdeaSet each woman's mammogram interval from her last mammogram, using AI risk
Shares MASAI: AI-supported mammography screening finds more cancers with half the radiologist workload, Mammography & tomosynthesis, AI in radiology, Overdiagnosis and false alarms.
- IdeaRoll out AI-supported mammography nationally as a stepped-wedge trial
Shares MASAI: AI-supported mammography screening finds more cancers with half the radiologist workload, Mammography & tomosynthesis, AI in radiology, Most lethal cancers are found late.
- Key paperNELSON: volume-based CT screening reduces lung cancer deaths with fewer false alarms
Shares Low-dose CT lung screening, Stage shift, AI in radiology, Overdiagnosis and false alarms.
- IdeaAI malignancy scores to end repeat scans and biopsies for benign lung nodules
Shares AI in radiology, Overdiagnosis and false alarms, AI that is built but not validated or deployed.
- IdeaMolecular indolence classifiers bundled with every screening programme
Shares Stage shift, AI in radiology, Overdiagnosis and false alarms, Most lethal cancers are found late.
- PersonKatie Couric
Shares Mammography & tomosynthesis, Stage shift, Most lethal cancers are found late.
- IdeaA dedicated fund for randomised trials of cancer AI with patient outcomes
Shares Mammography & tomosynthesis, AI in radiology, AI that is built but not validated or deployed.
- TrialNLST & NELSON (low-dose CT screening)
Shares Low-dose CT lung screening, Overdiagnosis and false alarms, Most lethal cancers are found late.