Eye eye! DeepMind teams up with doctors to ogle eyeballs for illness

Not an ophthalmologist, but... high street opticians typically take a picture of your retina (called a fundus image), not an OCT, which can image the different layers, often taken as a cross-section, see https://www.opsweb.org/page/RetinalOCT

Unsure whether OCT requires iris dilation or not (have had it done a few times and you lose track of all the drops they're putting in).

OCT

@AC & @ibmalone

Staff shortages at my local hospital have caused my OCT scan to be cancelled twice.

Check your local opticians/ask - some do have OCT machines. Mine offers OCT scan as part of their "enhanced" eye test - cost about £40, takes only a few minutes. No need for dilation.

Re: So it's wrong 6% of the time ...

.. whereas us real-I humans have a record of being right that is /far/ superior... :-D

Re: So it's wrong 6% of the time

"Why is AI in medicine, in particular, not being held to the five nines standard expected of technology in critical systems?"

Because five nines has never been the standard expected for diagnosis of diseases. No human doctor has ever come close to that kind of accuracy, so it would be utterly insane to demand a computer must be that good before it can be used. If a machine learning system can assess scan as accurately as a human expert, what exactly is the problem? It frees up the humans to spend more time doing jobs the computers can't do, and brings essentially zero risk since no-one is going to start removing eyeballs based on a single scan without further tests regardless of who or what made the initial diagnosis.

Re: So it's wrong 6% of the time

"Whey are we expected to accept crap reliability rates when something is described as "AI"?"

Because we accept that kind of rate when someone is described as a consultant ophthalmologist as well.

The overwhelming majority of the paper deals with the comparative performance of the model, which is markedly more complex than the 6% figure quoted in the headline. For example in urgent care decisions the rate falls to well below 1%, while when presented with scans from a machine the model had never seen before the performance initially collapsed.

The other aspect glossed over is that this isn't a binary output. The model produces a probability matrix over both its suspected diagnoses and recommended next action. Organisations will be able to set decision boundaries at the appropriate confidence levels, in line with their risk appetites, to account for the inaccuracies in the model.

There are millions of these scans performed worldwide every year, each one needing to be reviewed in detail by a skilled clinician. This is a technique that promises to perform just as well as that skilled clinician, allowing us to triage and prioritise our responses. The savings and the potential improvements in care, training and diagnosis are huge.