Off-the-shelf algorithms are not really fit for purpose when it comes to analysing scans from cancer patients, suggests Professor Dow-Mu Koh, of London’s world-renowned Royal Marsden hospital. Now, under a unique new grant-funded scheme his researchers are now working with NTT DATA to create, test, and manage new algorithms that the hospital hopes may be able to accurately identify cancers and their metastasis using, yes, AI.
Professor Koh, a consultant radiologist, is a global authority on functional imaging techniques for tumour assessments. Sitting down to chat with The Stack about the programme, he says that “I think there is a gap now between what is out there, available commercially, and what can usefully be brought into the health service," but admits that "the potential is vast.”
“This landscape is evolving very rapidly, and I think everybody is sort of playing catch up, as to thinking about how best to plot these largely well intended, meaningful tools and translate them into meaningful use within a health system like the NHS” he adds, saying the vision is to get a well-governed, compliant, and effective end-to-end machine learning pipeline in place “with real patient data going through it in real time.”
The hope, in short, is to further develop and evaluate AI algorithms to improve the accuracy of cancer evaluation, including for sarcoma, lung, breast, brain and prostate cancers, with the aim of getting faster response times, more accurate diagnoses and better-targeted treatments.
The support of technology here is critical.
"There is a workforce shortage" says Professor Koh simply. "The number of diagnostic tests, when it comes to imaging, is rising at an incredible rate, especially in the last 10 years, and the UK has not trained sufficient number of radiologists, or we are losing people to retirement, or emigration; so the workforce has shrunken, but the work volume has increased."
Build it, not buy it...
NTT DATA’s UK and Ireland CTO Tom Winstanley, speaking separately to The Stack, says: “The challenge that we found in AI in in the medical domain is the practicality of comparing, contrasting all the different [evolving] technical models that are out there [and] actually being able to to deploy all of those in a secure, suitable for regulated industry space,” he adds, “because this area is kind of a cottage industry still…”
“Lots of solutions today are SaaS-type solutions; especially in that R&D space that just wasn't [what Royal Marsden] was looking for.”
Sign up for The Stack
Interviews, insight, intelligence, and exclusive events for digital leaders.
No spam. Unsubscribe anytime.
NTT DATA has spun up a bespoke on-premises environment for Royal Marsden to do the work in, tapping AI-ready Dell servers with GPU capabilities and deploying the CARPL.ai platform, which helps test AI algorithms and includes a large collection of radiology AI models. NTT is also providing specialist imaging consultancy services, the firm says.
(CTO Winstanley tells The Stack: “We're seeing increasing demand [not just in health] but in our manufacturing group, also in our banking group for these kind of controlled environments – not to the exclusion of hyperscale AI models leveraging all of those as well – but for domain-specific LLMs and SLMs where you want to have total control of the data.”)
The project is being grant funded by the National Institute of Health and Social Care Research to the tune of low single-digit millions, The Stack understands; one hope is that if Royal Marsden manages to create any uniquely effective algorithms for identifying cancer it could commercialise them and make them available more broadly across the NHS.
Data annotation
Asked about accessibility of adequate training data, the team are largely positive, but admit there is a lot of cleaning up to do.
As Professor Koh notes: “We're fortunate that we've been blessed with an in-house imaging repository that we have developed over many years with the Institute of Cancer Research, which is our academic partner.
"We have a system that has allowed us to bank quite a lot of datasets from clinical trials and also from real world into this repository… we've been able to work across the geographical divide and work with other partners, for example, in Europe [to aggregate], prostate cancer MRI datasets now numbering about 18,000 [images] as part of that partnership.”
Yet with Royal Marsden working with this project on some quite rare cancer forms, there’s still a lot of manual image annotation and metadata tagging of some imagery that needs doing before any AI pipelines can be deployed.
That’s not just for data hygiene reasons, but policy ones too. Of image annotation “it is still, unfortunately, much needed in even deep learning with generative AI; yes, you can do a first run of some of these tasks but… you need dedicated people with dedicated time to look at [the images].
"[Anonymisation is part of that because] when it comes to imaging, it takes on a slightly different layer of complexity. For example, if the images involve a face… [we have to deploy a] very specific way of actually dealing with with that, to get everybody into a position of confidence that what we are doing meets regulatory standards and governance standards,” Professor Koh adds.
Workforce burnout...
Stepping back from the specific details of the project, Professor Koh is keen to emphasise the urgency of technological progress in this space.
"Clearly AI has to be part of the solution; people are looking at to see whether some tasks can be reasonably done by a machine with some sort of human interaction at some point – either auditing what the machine has done or signing off what the machine has produced – because at the moment, we really do have the risk of a workforce burnout if we don't do anything about it. So that operates in the background," he tells The Stack.
"[But] if you look at [existing] algorithms, unfortunately, many of the tools that have been developed commercially were based on the enthusiasm of that generation when deep learning just came out. So they tended to be developed using modest datasets; the context tends to be very narrow.
"Then there is the practical issue of how well it's actually going to do in your environment. Also, these tools are very costly. For example, they will typically cost – depending on the model – anything between £20,000 to and upwards per year to actually purchase. In a cash-strapped NHS people are looking at value for money, and one therefore has to really verify that it does what it says on the can, and [proove how it can] meaningfully improve the workflow or the outcome of the patient, for example, for it to be adopted."
"What we're doing here is trying to operate against the canvas of all that forces to find the one that can truly help us do something.
"That's not helped by the fact that now with generative AI, new tools are coming in all the time that promise to to do something better than the last generation of the two. But I think you will recognize one of the problem of trying to test technology that is constantly evolving; before you can truly validate that technology, it has moved on! I think that regulators also find great difficulty [in knowing how to] regulate these tools. Because regulation really involves the burden of consistency, reliability, etc."
But the Royal Marsden team is keen to build on its groundbreaking 2023 research that found a new AI algorithm it develo[ed could help tailor the treatment of some sarcoma patients more accurately and effectively than a biopsy. Professor Koh says: "this is one of the projects we [will] run through the [new] MLOps platform so that we can then put prospective cases through and see how it actually affects the decision making."
The link has been copied!
