Regular MRI is a key element in the follow-up of pwMS, particularly those on disease modifying treatment (DMT). Why? Because MRI is significantly more sensitive in detecting MS disease activity than counting relapses and measuring disability using the EDSS (Kurtzke’s Expanded Disability Status Scale). I’m quite precise here referring to the EDSS since other clinical instruments may be better to monitor pwMS over time. Indeed, at Barts Health, we have been collecting a handful of test results other than EDSS for some time, and these feed into management decisions. Check here and here for more.
But let’s not get distracted and rather focus on MRI follow-up. Many of you may be aware that in the UK, treatment decisions for high-efficacy DMTs have to be discussed and agreed at multi-disciplinary team (MDT) meeting consisting of at least two consultant neurologists, an MS nurse specialist, and neuroradiology support to interpret MRI. Involvement of a neuropharmacist is optional – at our centre we are blessed with an outstanding, award-winning one.
In treatment-active MS centres, DMT decisions are commonly based on MRI changes alone. But there are problems: Firstly, MRI scans obtained at different MRI systems are often difficult to compare since the techniques used differ, more or less. However, if you are keen on detecting even small changes, this “more or less” difference between MRI acquisition techniques may be key for a decision to switch (or not!) to a different DMT. Neuroradiologists would then often use phrases like “given the difference in techniques, no changes were detected”.
Secondly, counting MS lesions, or assessing changes in size of such lesions, is one of the most boring activities for any neuroradiologist not involved in MS research, which is the vast majority.
Thirdly, the time allocated to assess MS follow-up scans varies depending on other pressures, and in the NHS there always are other pressures…
We have highlighted the issues related to the first point in two papers, one a perspective on what should be done, and an audit on what is actually happening on the ground – There is a clear need to “unite the Kingdom” in what is being used to obtain images.
The second & third point is increasingly being addressed using artificial intelligence (AI), i.e. assistive software that highlights areas of change thereby guiding the clinician in deciding whether a significant change has indeed taken place or not. As it turns out, AI may well impact on point one above as well !
Over and above MS lesions, these AI-technologies are able to measure whole brain volume, and even segmented white and grey matter volumes to estimate the degree of brain atrophy.
The paper below by Diana Sima and her colleagues is the latest one in a series making the case for using AI-assisted MRI assessment in clinical practice. Using model simulations they assess the potential health economic impact in detecting MRI changes in pwMS.
Click on the title and you can read the paper free of charge. It is a quite challenging read for the non-expert though, and one of the co-authors, Wim van Hecke, therefore provided me with a more accessible summary, which is here:
There-is-more-to-MS-disease-activity-than-meets-the-eyePlease note, the authors of this work are all employees of icometrix, including their CEO and CTO, so do read it with conflicts of interest in mind. Having said that, I’m pleased to report that BartsMS, together with colleagues from the University of Nottingham, have submitted a joint bid with icometrix to explore AI-assisted MRI in routine care of pwMS – a project called Assist-MS. More about this soon.
@KlausSchmierer
Disclaimer: The opinions expressed here are those of the author and do not necessarily reflect positions of Barts and The London School of Medicine & Dentistry, Queen Mary University of London, or Barts Health NHS Trust.
Could we use AI to determine how EBV causes ms ?
I think this might be happening already; there’s people with machine learning background involved in studies like this: https://www.thelancet.com/action/showPdf?pii=S2352-3964%2821%2900365-0
Awesome news! Without playing the blames game (as there are many fantastic practitioners out there). Can’t tell you the additional frustration/worry we have been exposed to by having lesions missed on MRI’s being denied more effective meds because of this (had to pay for additional specialists to look at MRI’s taken and they were picked up).
Making access to best reading technology for pwMS equitable, on the highest level possible, is one of the drivers of this initiative & project.
Hopefully, humans can’t be as accurate at detecting small but significant changes as well as computers / software.
Is the software readily available or will some hospitals have access while other won’t?
We’re planning to test implementation at two sites first, so this will not be available across the NHS for now. AssistMS may provide the evidence required for broad NHS adoption. Having said that, there is prior experience with the technology at several UK centres. Elsewhere, the software is already being used for clinical management.
One of the challenges of commonly used AI technologies such as neural networks is the difficultly in building assurance in the performance of the algorithm.
What level of confidence in the algorithm will be required to trust it in clinical decision making?
A concern would be that if sufficient confidence cannot be achieved, an algorithm could fail to spot an indicator of progression that a neuroradiologist would have identified. Automation complacency could lead to a lack of scrutiny during any annual review process introduced to mitigate this…
Thanks Chris, and yes there are evident limits to what AI can do. This is why the technology will always be combined with the expertise/experience of neuroradiology. Everybody involved in this type of work has had to realise that assisting, much rather than replacing, the eyeball of the radiologist/neurologist is – at this stage anyway – the optimum way forward. I will talk about AssistMS, which is very much about implementation in clinical practice, in a separate post.
Definitely an interesting idea, AI Doc is currently doing something similar at a commercial level. They focus on triage with things such as stroke. hemorrhages and fractures.
Mind you AI Doc and icometrix recently teamed up on stroke: https://www.prnewswire.com/il/news-releases/aidoc-and-icometrix-team-up-to-transform-stroke-care-coordination-301317252.html
Nice work
Happy new year
And to you – always appreciate your comments (even if not commending😉)
I red somewhere that AI is used to examine mammograms for evidence of breast cancer. The human eye can only detect 7 shades of grey but AI is limitless
Happy New Year, Patrick. Indeed, and it is interesting that even in mammography there are still issues with implementation https://www.massgeneral.org/news/press-release/ai-assisted-mammography-prospective-clinical-evaluation
New LINAC-MRI technology also incorporates AI technology. As the data and algorithms grow, it is expected that the machine will be able to identify the tumour location, size, and surrounding tissue and determine the required radiation dose and number of fractions (sparing sensitive surrounding tissue). Pretty cool technology. If only MS receiving the same level of funding as oncology…..
I have been waiting for this. It has bothered me from the first how badly our MRIs are handled. I’m in Canada but don’t think anywhere has avoided the pitfalls. MRIs are regression tests and in other industries there are systematic test methodologies for similiar problems. Here leisions are not itemized nor are older MRIs used in the comparison. This makes it way to easy to miss things. Only big changes are being picked up from one MRI to the next.
Shouldn’t AI assistance also be used in the imaging process itself, rather than just in interpreting data? Is this something that is also being considered?
I ask this because, I had an attack some time ago. So we did an MRI. The radiologist not only didn’t find anything active but also missed a prior known lession on my spine which was around 9mm!!! Which begs the question should humans be trusted in the imaging itself? On the images that he took, there was indeed no lession visible (I assume he just didn’t scan the area enough)…
I think the issue here is less about AI applied to acquisition and/or reading processes, but rather the choice of the area scanned. Particularly with respect to the spinal cord, there is no agreement whether only the neck proportion or the entire cord should be scanned, indeed whether it should be scanned at all!
Could you clarify why? What about people like me whose main lesions are on the spine. I have lesions at C2 and T9.
Is there a reason why the spinal mri is not accurate?
At diagnosis, head an spinal cord MRI would usually be obtained as a rule. Follow-up may depend on preferential location of lesions, so in your case it makes obvious sense to include the cord regularly. However, there is no unequivocal routine. Personally, I would prefer head and spinal cord at every follow-up. This is based on clinical experience and studies like the one be Zecca, et al. 2016, available here free access: https://journals.sagepub.com/doi/pdf/10.1177/1352458515599246
Best practice is annual MRI for MS, but when I swapped clinics in 2014 I was never offered one. When I offered to pay in 2020 I was told I couldn’t even hope to get an MRI privately. Is this down to post codes? I slunk back to my original clinic and got one but it’s much further to get there. That = two per century!
Kit, you describe part of the (deeply concerning) inequity for pwMS across the NHS. Not sure why you couldn’t even get a self-funded MRI(?)