OnCo
ideasIdea

AI that spots pancreatic cancer on scans taken a year before diagnosis

Pancreatic cancer is often visible in hindsight on earlier scans. Software trained on those pre-diagnostic scans could flag subtle changes while surgery is still possible.

Retrospective studies (Mayo Clinic, Johns Hopkins FELIX project) show deep learning detects pancreatic cancer on pre-diagnostic CTs 3-36 months before clinical diagnosis with high AUC. Propose training on multi-institution pre-diagnostic scan cohorts and prospective deployment on abdominal CTs of patients over 50 with new-onset diabetes or weight loss.

Hypothesis
Prospective deployment in high-risk CT populations detects resectable (stage I-II) pancreatic cancer in at least half of flagged true positives, versus roughly 15% resectable at usual diagnosis.
Rationale
Pancreatic cancer survival is stage-dependent and the pre-diagnostic window is documented; incidence in new-onset diabetes over 50 (about 1%) is high enough to justify targeted use.
What would test it
Prospective cohort of 20,000 abdominal CTs in the target population, with two-year registry follow-up for sensitivity and PPV.
Maturity
preclinical evidence
Who has to act
research
Cost to try
Medium ($1M to $50M)
Years to first evidence
4
Bottlenecks it attacks
  • 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.

Connected

8top

Pages like this

not linked directly; found by shared links

cancers

1
Pancreatic ductal adenocarcinoma

fronts

2
Early Detection & ScreeningImaging

technologies

2
AI in radiologyCT (computed tomography)

institutions

2
Johns Hopkins Hospital / Sidney Kimmel Comprehensive Cancer CenterMayo Clinic

bottlenecks

1
Most lethal cancers are found late