Five cancer breakthroughs that needed AI
Results published with numbers attached, in randomised trials and multi-centre validations, including the one where it did not work.
Two of these change what a screening programme finds. One changes how well the doctor works, and comes with a second trial where the same idea did nothing. One changes which drug a pancreatic cancer patient is handed. The last reads an expensive test off a slide that has already been cut.
Breast screening found more cancer without more false alarms
In the Swedish MASAI trial, 105,934 women had their mammograms read either by AI plus a radiologist or by the usual two radiologists. The AI side caught 80.5% of the cancers that were there, against 73.8% — and called back no more healthy women for extra tests.
It has not been shown to save lives. The trial was not built to count deaths.
A stomach cancer screen that reads scans taken for something else
Stomach cancer has no practical screening test outside a handful of countries. GRAPE finds it on ordinary CT scans already being taken for unrelated reasons. For every hundred cancers sitting in those scans, it caught about twenty-two more than the radiologists reading the same pictures, and it has since been run across 78,593 scans in real use.
All of it was done in China. Whether it holds up where the disease is rarer is untested.
Watching the doctor instead of the polyp
A Danish trial watched how the doctor performed the colonoscopy — how fast they pulled the camera back, how much of the bowel they actually looked at — and told them afterwards, every time. They went from finding growths that can turn into cancer in 43.4% of patients to 48.6%, and actual cancers in 8.9% rather than 6.5%.
The same year, CADLY2 gave the spotting kind of AI to expert doctors examining patients with an inherited condition that makes bowel cancer very likely, and found nothing: 33.8% against 30.9%, close enough that the trial could not tell it from chance.
Half of pancreatic cancer patients got the other drug
There are two drug combinations doctors start with for advanced pancreatic cancer, and no agreed way to pick between them. A model trained on 2,202 tumours picks one, and the patients who happened to get the drug it would have chosen lived longer. Roughly half had been given the other.
It is a look back through old records rather than a trial, and the people who built it work for the company selling it.
Reading an expensive test off a slide already cut
There is a test that maps which proteins sit where inside a tumour, and it costs enough that most patients never get one. HEX predicts forty of those readings from the ordinary stained slide the lab has already made. Across 2,298 lung cancer patients it did better than the tests in use now at saying how someone was likely to do, and whether a particular treatment would work on them.
Nobody has yet given a patient a different drug because of it.
The Force shall free me.