Somewhere in the basement of your nearest hospital, there’s a room full of glass where every biopsy goes under a microscope and decisions are made whether it’s cancer, how aggressive it is, and how far it’s spread.
Since roughly 2011, as you’d expect, machines have been able to read some of those slides more consistently than humans. That’s been cleared by the regulators, built and sold by the largest diagnostics companies on earth, and yet, most laboratories, in most countries, still haven’t bought any of it.
Why? Is this not a perfect, impactful problem to solve with AI?
Yes it is. But healthcare has forever grappled with a big problem. Perfect products can do absolutely nothing within an imperfect system.
Understand this and you understand a great deal about how (and how not) to innovate in healthcare…
Three different purchases
One bit of orientation first. The difference between pathology, digital pathology and AI pathology, so we’re all clear…
Pathology here means histopathology, i.e. tissue. Someone takes a biopsy or removes a tumour, the tissue is fixed, set in wax, sliced at a few microns, stained and mounted on a glass slide. A pathologist puts the glass under a microscope and decides what it is, how aggressive it is, and how far it’s spread. That is where a cancer diagnosis is actually made.
Digital pathology is a whole-slide scanner turning each slide into a gigapixel image, the pathologist reports from a screen rather than a microscope, and no more do you have a physical thing to store, break, transport etc.. That’s all digital pathology is. No AI anywhere. What it requires (again, note this…) is scanners, a network, an image management system, integration with the lab’s information system, and a lot of storage (not cheap). One site quoted 500,000 slides a year generating a whopping 800TB.
AI pathology is algorithms reading those images. Flagging which cores look malignant, counting stained cells, grading, quantifying the protein expression that determines whether someone gets a targeted therapy.
Now, those are separate purchases by the way and you need the previous one to get the next one. There’s no AI without digital. An algorithm has nothing to read until the glass has been scanned, so every AI pathology company is selling to a site that’s already forked out seven figures for digital infrastructure. And by the way, on the best evidence available, the financial return is… wait for it… pretty much nothing.
Facepalm. I’ll go into that in a second.
One industry count said 31% of primary diagnoses in the UK are now reported through an image management system, against roughly 23% in Europe and Asia and just 10% in the US. So any debate you see about model performance is happening on a fraction of what you’d expect.
The sceptic built it
Steve Burnell is worth listening to on this because he no longer has a horse in this particular race. I did a podcast with him recently (soon to be released). He went to Roche through an acquisition, sat inside the Foundation Medicine and Flatiron deals, ran digital pathology globally, and now runs Tenmile, a $250m evergreen fund backed by Andrew and Nicola Forrest’s Tattarang in Australia.
He gives three reasons why the thing he built didn’t sell. And these should be a warning to all innovators across healthtech too.
A regulatory trap that still hasn’t been fixed. To get an algorithm approved you have to show that it agrees with the gold standard, and the gold standard is a human pathologist. And when researchers went looking for the public evidence behind 26 commercial AI pathology products, they found that 24 had reached the market by self-certifying as general IVDs, and 15 of the 26 had no associated scientific publication at all.
“How often do you think you can get two human pathologists to agree on the diagnosis or the count of a cell membrane? The algorithm does it instantly and you know it’s doing it very very well, but then it’s got to get two pathologists out of three to agree with it in a blinded way.”
Workflow. Pathologists are rapid with a microscope, pathology skews senior, and keeping everything as it is means zero transition costs and zero risk.
Pathology labs are cost centres in the basement. They don’t sell anything and they’re invisible in the accounts.
“They are not visible in the healthcare economics system. So spending money in the pathology lab is a difficult thing to do.”
I’ll explain what I think that means and this is useful context for entrepreneurs.
In the UK, most NHS acute income now arrives through the aligned payment and incentive approach, i.e. the trust and its integrated care board agree on one largely fixed annual sum covering nearly all secondary care, topped up by unit-price payments for a defined list of activities. Elective work is on that list. Unbundled diagnostic imaging is on that list. Genomic testing is on that list. Do those things and get paid a bit extra, per thing.
Histopathology isn’t on that list, so that means reading a slide is never a payable event of its own.
Why is that important?
Well, take radiology as a comparison. Because it’s on the list, doing a scan has a price, so the department has products with a price and can sell those products. You can then write a business case to buy some tech that helps you sell more products to end up with more income.
Pathology can’t do that. It gets an allocation handed down out of the fixed envelope, which makes it a spender and never an earner. Worse, an efficiency gain in the lab doesn’t come back as money at all. It comes back as capacity. Capacity only becomes money if the trust converts it into something on that list. Or sacking humans of course. Classic.
So even if you sped everything up in a lab by 10x, it doesn’t improve the finances by a penny on its own. It actually worsens, because you had to buy something to get that speed increase.
Can pathology labs charge external sites for their service?
Yes, e.g. through direct-access work for GPs commissioned locally, referral work sold to other trusts under the hub-and-spoke networks, private patients, trials work for pharma etc.. Worth knowing, but it doesn’t change the problem because the core hospital cancer workload that digital pathology exists to accelerate still doesn’t actually get you any income.
Fork out for digital infrastructure and of course there are benefits when you take your finance hat off... patients starting treatment sooner, fewer repeat biopsies, fewer wasted MDT slots, less work outsourced, less locum cover, shorter stays, better performance against the 28-day faster diagnosis standard. But what’s depressing is that whoever signs the PO for that massive £2m infrastructure change gets no financial benefit themselves.
It’s another budget boundary issue and they plague healthcare innovation.
I asked Steve, “is this just money saved here, spent there, and nobody zooming out?”
