Will the NIHR give us the chance to test AI-assisted MRI in clinical practice?


I recently posted on the promises of AI-assisted MRI in routine follow-up of pwMS and mentioned that we’re applying, together with colleagues based in Nottingham and a commercial leader in the field (icometrix), for an NIHR award to see whether such an approach can indeed result in better “quality, efficiency and equity in the NHS care of multiple sclerosis” (from the title of our proposal).

Well, time is moving fast, and we’re up for interview in less than two weeks, so I thought you’d like to wish us well and perhaps give some feedback on the presentation. Here it is:


Disclaimer: The opinions expressed here are those of the author and nobody else.


About the author


  • What can AI pick up that humans can’t?

    Or are there something AI could miss?

    I think it’s a great idea btw, save neurologists time, however I hope it doesn’t make for complacancy

    • When it comes to lesions, evidence suggests radiologists are better when combined with/assisted by AI. When looking at volumetric measures (brain atrophy, etc) AI is superior hands down.

  • Great idea. It makes me wonder if the human radiologists will soon be having to retrain – like cardio-thoracic surgeons when percutaneous intervention was developed.
    I can’t see where the n=646 comes from – is this on the basis of a power calculation, or is it a pragmatic number based on the frequency of MRIs routinely undertaken at each site. How many sites are involved?
    Could there be a process whereby if the AI system spots something that the radiologist missed, or vice versa, they have a group of radiologists to review the difference and collate the information to look for correlations in analysis.

    • Thanks Mark. You could’ve written the protocol (or are a member of the AAC panel…). Some training will be needed, but not much. There will be two sites, Nottingham and the Royal London; the 648/site comes from a power calculation of the minimum detectable difference based on 48 clusters (=MDTs) per arm with 15 pwMS per cluster, and factoring in a 10% drop-out rate. The resulting sample size, at a power of 80% and an alpha of 5%, was 1296 patients with a minimum detectable difference of 8-9% which is just below the range (10-20%) of the difference estimated from pilot studies. There are secondary objectives covering the issue of expertise, a non-MS expert radiology team and a gold-standard reference team for which separate sample calculations were undertaken by our fabulous statistician Amy MacDougall (https://www.lshtm.ac.uk/aboutus/people/macdougall.amy).

  • Championing of this new technology is great and hope that this research trial gets funded. What is of concern though is ongoing lack of skilled workforce and the implication that this AI will fill the gap. This AI should be in addition to a well resourced and skilled workforce not a solution to a resource gap

  • Most importantly, I like the idea of the study, would have no hesitation in taking part, and would be grossly disappointed if allocated to the standard-of-care arm. Further thoughts from my perspective:

    I think having an annual mri is as likely to happen as spotting a unicorn on the M1. I have had 1 scan since diagnosis 5 years ago. The nice guidelines sets this as the baseline expectation for those treated within the uk, and it is not met. My care is at one of the lead tertiary centres in the uk, so I can only assume that the dgh patients get far worse. There is a clear need for studies which might improve throughput, and allow the nhs to meet its own standards.

    Potential for efficiency aside, enhanced monitoring in the current climate holds great appeal. One of the most disconcerting things about having ms is not being able to predict the future. The narrative around “smouldering ms” is particularly concerning, and therefore I was previously an advocate for very aggressive therapies and was frustrated that I was not eligible for these. However, since covid I have become a lot more conservative about immunosuppression, worried about the issues with vaccine efficacy. The software is potentially a more sensitive tool for patient monitoring; if I was monitored with this I’d feel more reassured about being on a first-line dmt if the software still indicated NEDA.

    Thirdly, I like the idea of doing this in the setting of an academic led, NIHR funded study, as whilst the commercial data looks promising it is somewhat rose-tinted.

    Good luck with it.

      • I hope the bid is successful.

        Am I right in thinking you are only looking at the brain and not the spinal cord, which I understand is harder to image?
        If your current bid and trial goes well, are you planning
        follow ups that are harder to justify in terms of quick cash returns but might lead to longer term rewards? I’m thinking smouldering, of course.

        • Thanks. Currently only brain indeed, although at the Royal London we would more often than not do the spinal cord in follow up as well. Smouldering pathology needs a bit more work, however certainly something we will look at once MRI acquisition is ready for a routine job at 1.5T or 3T.

  • If Icobrain proves a useful tool for radiologists, improving accuracy and efficiency then it sounds a no brainer (pardon the pun). And if it speeds things up (which it will, bet Icobrain is pretty much instant) then that’s brilliant too.
    Good luck with your grant application 🙂

  • Can the software look at brain volume or for changes between scans due to activity?

    I know you have to estimate a number of patients lost to the study, but do you think it will be as high as 10% when it is routine to attend every year for a scan anyway? I’d be more than happy to join – you’d probably have an easy time recruiting!

    • We will primarily assess disease activity (lesions) since this is what current underpins treatment decisions in line with NHS guidelines. However, brain volume indices will come as a compliment, and I expect will develop a more important role as the project evolves.

  • If I understand correctly, AI is already routinely used in reading mammograms and the like very successfully. Hopefully this will be the case here. I believe there is a dire shortage of radiologists – so one hopes this will ease the situation.

  • Prof. K – Thank you for sharing the presentation slide deck and soliciting our feedback.

    Might I suggest adding a few thoughts on how this technology would provide, “early detection and intervention” when diagnosing a pwms.

    After reviewing the presentation, you could possibly add the concept of early detection and intervention to the “Background” slide. It would fit nice at the beginning of 3rd bullet….Early and effective…

    Also would fit on the “Impact” slide, as a benefit.

    Just a few thoughts. Good luck with the presentation!

By ProfK



Recent Posts

Recent Comments