MRIs Miss Brain Lesions Causing Epilepsy - AI Can Find Them - Dr. Konrad Wagstyl, KCL & UCL Great Ormond Street Institute of Child Health, UK
Many epilepsy lesions that cause seizures are missed on routine MRI scans, delaying accurate diagnosis and potential surgery. Dr. Konrad Wagstyl discusses his research using AI and large MRI datasets to detect subtle abnormalities such as focal cortical dysplasia and hippocampal sclerosis. The conversation explores how these tools help identify surgical candidates earlier, enable precision medicine, and improve outcomes in people with an epilepsy!
📺 Watch the Interview
🎧 Listen to the Podcast
👉 Listen on Spotify | Apple Podcasts | YouTube
This episode is sponsored by EASEE® by Precisis who’s had no influence over the editorial content or discussion. Learn more about EASEE® here.
Episode Highlights
Why some epilepsy lesions are missed on MRI scans
Detecting lesions and focal cortical dysplasia using AI
Identifying hippocampal sclerosis with automated MRI tools
How lesion detection changes the epilepsy surgery pathway
The MELD project and global MRI datasets
Improving diagnosis in MRI-negative epilepsy
About Dr. Konrad Wagstyl
Konrad is a Senior Lecturer in Biomedical Computing at King’s College London (KCL) and an Honorary Associate Professor at University College London’s (UCL) Great Ormond Street Institute of Child Health . He holds a Wellcome Senior Research Fellowship at both institutions and leads the IMAGINE Lab, which focuses on developing advanced imaging and machine learning techniques to better understand and treat neurodevelopmental disorders, particularly epilepsy.
Full profile: Konrad-Wagstyl
Topics mentioned
mri-negative epilepsy
focal cortical dysplasia
hippocampal sclerosis
epilepsy surgery pathway
stereo eeg (seeg)
machine learning in neuroimaging
lesion detection in epilepsy
sudep risk
Related paper
“Automated and Interpretable Detection of Hippocampal Sclerosis in Temporal Lobe Epilepsy: AID-HS”, Annals of Neurology, https://onlinelibrary.wiley.com/doi/full/10.1002/ana.27089, Mathilde Ripart et al, 14 Nov 2024, DOI https://doi.org/10.1002/ana.27089
-
Trailer
00:00 Konrad Wagstyl
“That's the kind of golden, the golden bullet. If you can find it on an MRI scan, the chances of seizure freedom are really high. If you can't - so that's the group that we're really trying to help here - there's a range of options. So firstly, if they're first not seen, typically they'll be given anti-seizure medications, which are unlikely to work.”
Intro
00:16 Torie Robinson
Heaps of people with a focal epilepsy are told that their MRI scan looks normal, even though they are or have experienced epileptic seizures! To be blunt: there are subtle brain abnormalities that are being missed. In today’s episode! Dr. Konrad Wagstyl from King’s College London explains how biomedical computing and large international MRI datasets are helping train AI tools to detect otherwise missed lesions like in focal cortical dysplasia and hippocampal sclerosis - that may otherwise go unseen and delay access to effective treatment - including surgery.
If you’re new here, please subscribe so you don’t miss our chats with global leaders in the sphere of epilepsy - and let’s get into today’s episode - presented in partnership with EASEE®, by Precisis GmbH.
Initial problem solved - brain abnormality identification
01:02 Torie Robinson
So, you've been researching children and adults with focal epilepsy where the MRI appears normal. What problem are you trying to solve here with your work?
01:10 Konrad Wagstyl
So, we're working on children and adults with focal epilepsy and in around one in seven, the seizures are caused by a brain abnormality or lesion where if you can find this brain abnormality, you can operate to remove it and that potentially stops the seizures. However, it's very difficult to find these on an MRI scan, often they're missed, and that makes it very difficult to figure out where to operate, what the best treatment is. So, we've been working with people around the world to create datasets to train AI models to find these abnormalities to help radiologists, neurosurgeons and neurologists figure out how best to treat these patients.
Type of lesions: hippocampal sclerosis and focal cortical dysplasia
01:48 Torie Robinson
And what type of lesions are we talking about? That can be missed on the standard imaging and in what part of the brain do they reside?
01:55 Konrad Wagstyl
So there are multiple different lesion types that can cause focal epilepsy. Hippocampal sclerosis is the most common one in adults and in children the most common and one that's very hard to spot is focal cortical dysplasia. But there are a whole range of other pathologies becoming more rare. There are some tumours that can cause seizures and all of them in theory, can be missed and in practice actually are missed. Some of them are much harder; so focal cortical dysplasia is very commonly missed. Hippocampus sclerosis is missed about 20% of the time and they vary. Some are more clear and some are harder to spot. But we have seen cases of all of them being missed.
Impacts on surgical pathway
02:35 Torie Robinson
And so, when these lesions are identified, how does that change the surgical pathway for people?
02:41 Konrad Wagstyl
That's the kind of golden, the golden bullet. If you can find it on an MRI scan, the chances of seizure freedom are really high. If you can't - so that's the group that we're really trying to help here - there's a range of options. So firstly, if they're first not seen, typically they'll be given anti-seizure medications, which are unlikely to work. So, this will carry on for many years until it's clear that these aren't working. They might have multiple MRI scans during that period. They'll have their inpatient hospital stay, often 10 days. There's also a particular invasive investigation called stereo EEG where they'll put electrodes into the brain to try and figure out where the seizures are coming from. If you have a clear MRI lesion, often many of those delays, many of those investigations aren't necessary and you can proceed for consideration for surgery much, much quicker.
03:27 Torie Robinson
This is so exciting. So, tell us about your paper, specifically on this; so, what was the question initially? What were your discoveries and were there any surprises?
Papers regarding cases project and cases that MRI didn’t identify
03:37 Konrad Wagstyl
So, I'd like to highlight two papers in fact. So last year we produced one tool that is good at detecting focal cortical dysplasias. This was called MELD Graph. And we produced a second tool in a separate paper called AID-HS (so Automated and Interpretable Detection of Hippocampal Sclerosis). And those tools were one for focal cortical dysplasia, one for hippocampus sclerosis. And this is working with many hospitals around the world who have shared their data with us. We give the computer many examples where we know where the abnormalities are (where the focal cortical dysplasia is, or the hippocampus sclerosis) and we train it to find them. And we're looking at the focal cortical dysplasias; one really interesting thing is that the model picks up around 70% of them, which is pretty good, especially as they're hard to find, but in the ones that are MRI negative, the ones that were missed, it's still picking up two thirds of them. And so, this is the group that really can make a big difference. It's clearly using something slightly different to what the humans are using to find these abnormalities, but it can pick up missed ones really, really regularly. And I think that's one surprising thing is that it seems a little to be doing things slightly different to the human.
It’s an open source tool
04:47 Konrad Wagstyl
And then the other thing that's been fantastic is we've made these tools open source, which means that any hospital in the world can download them and try them at their site as a research evaluation. And the uptake has been a really surprising thing. We know of at least 250 sites that have downloaded the code and it's really a fantastic bit of our work when you hear reports of, you know, “We've tried it in this patient and it found something that we hadn't seen, but we definitely can see it now.” and that's kind of makes our day.
Improved quality of life, inc. cognition and SUDEP risk
05:14 Torie Robinson
I can, well I know, can feel it here right now. The difference that would make to a person's, well, treatment, but also potential quality of life, and extends (in many cases, I imagine, should they be suitable for surgery or just, or potentially not) improves their life expectancy in some cases, correct?
05:40 Konrad Wagstyl
Unfortunately, if, if you're at the epilepsy is not being well treated by medication (but this is true in focal epilepsy as well), there is a risk of sudden unexpected death in epilepsy (SUDEP), and if you remove the seizures, obviously that risk goes down. Yeah, life expectancy, but even just life, quality of life. Some people go from 20, 30, we've heard of a child with 70 seizures a day; if you can bring that down, or, you know, as far as zero, that's obviously a completely different quality of life.
06:00 Torie Robinson
And the positive impact on cognition as well, or prevention of cognitive regression in some cases. And we often talk about SUDEP, which is really, really important, but also there are other ways, like I don't want to make people feel bad, but it's important to acknowledge other ways that people can die. Like, as a result, I fell on a railway line during rush hour during a seizure years ago. So, things like that, you know, if we can minimise that likelihood - ideally through seizure control or prevention, that is kind of ideal, I think.
Improvements in cognition
06:29 Konrad Wagstyl
Absolutely. I’d like to highlight a different paper that came out of the team at Great Ormond Street where we work. We were involved with it but led by Maria Erickson where she was looking at cognition over time, prior to surgery and then after surgery, and saw, you know, progressively over time, if the seizures aren't treated, cognition goes down, as you were describing. But if you are seizure free after surgery, it rebounds quite, quite noticeably. And particularly if you can stop the anti-seizure medication, then there's a big jump. And so, if you can offer that, that's really powerful.
07:01 Torie Robinson
And can often also prevent or halt neurodegeneration as well, at least in some cases, right.
07:07 Konrad Wagstyl
There's this very interesting area of research of epilepsy as this kind of progressive disorder that you're describing where there are long-term effects. We're not actively in that area, but it's definitely what motivates us to try and get these abnormalities diagnosed sooner and treated sooner.
Statistics from the papers
07:23 Torie Robinson
Can you give us some statistics from these papers please so we can put things into a bit of perspective?
07:26 Konrad Wagstyl
We were training our models on the focal cortical dysplasia side, a thousand individuals we were training them with. And that's sensitivity of around, I think… if we knew that the patient was seizure free (so we absolutely knew where the lesion was), we were picking up 85% of these lesions. It's sometimes hard to know exactly where the ground truth is. So, then we get 70% in the larger cohort, 65% in the seizure free. And I think there are some interesting features that we've really tried to build into these models so that they're interpretable for radiologists, which means that when you then get a new prediction, it produces a confidence score. If it says, I'm 90% confident, that means 9 times out of 10, it's correct. If it's 10% confident and the radiologist told us they want to hear about these, you then go and say “Well, we need some other information to make that… to be confident that we might then go on to operate like SEEG or this is consistent with their seizures.”.
The second model we've been training is this hippocampus sclerosis (AID-HS we call it), and this has been shared and also downloaded a lot, and there've been a number of other papers who said we've tried this tool. It works exactly as they said. So, we pick up 95% of the hippocampus sclerosis cases with this. In those that were MRI negative, it's about 80%. And what's been really cool to see again here is that because we've trained it with such a large data set, when a new hospital tries it out, it basically works for them as advertised, the same that we found it to work.
08:59 Sponsor mention
Before we move on - with thanks to EASEE®, by Precisis GmbH.
Downloading the research tool - for free
09:04 Torie Robinson
So researchers and clinicians who may be interested in this, who haven't yet heard of it, do they need anything fancy in the office or in their lab to be using this software that you've put together?
09:18 Konrad Wagstyl
We've tried to make it as usable for clinicians and researchers, with the big caveat that it's still just a research tool, but that they can download it. It works on a Windows, it works on a Mac laptop, it works on my laptop, it works on a Linux machine. So, the three main operating systems. We've tried to make YouTube videos to help people install it. And if you try it and get stuck, we help people to get over that hurdle. Because once it's in place, it's very easy to run.
Next steps in the research
09:42 Torie Robinson
Amazing. And so, what are the next steps in this research of yours?
09:44 Konrad Wagstyl
It is just a research tool at the moment. And when we spoke about this with our patient advisory group, they feel very passionately that it shouldn't be just available at the university hospitals. They want it in their local hospital. Many have this story of the first scan being read as MRI negative and only later something being found, and they want it to be there on that first scan.
So, our next step is to truly make these tools extremely robust so that they could work at any hospital, to show that they do work on these first epilepsy scans so that you can pick out the lesions and get them referred quicker. And part of that is then also working to make it very robust that it could integrate into a hospital system, that it's safe. There's all sorts of… to make it a medical device, we have to think very carefully about making it extremely safe and proving that it's extremely safe.
10:33 Torie Robinson
And what are the next steps in order to do that then?
Clinical trial
10:36 Konrad Wagstyl
So, we're going to start clinical trial next year on this to really show what happens when you put it in. And I think part of that is, you know, do you… would you be referring more people for surgery? How does that impact health pathways? All these sorts of considerations that if you want to put something like this into the NHS, you really need to think about. So clinical trial, the kind of safety software system, and then impacts on NHS are the bits that we're working on and other… we're not just a UK based thing; we really want to make it work worldwide.
MELD project and countries involved
11:05 Torie Robinson
I was actually going to ask you about that. So, for instance, you know, people in remote areas or low to middle income countries, especially where often we have more people affected by the epilepsies anyway, this could be a tool that potentially is of benefit to them as well, I imagine.
11:18 Konrad Wagstyl
MELD, which is our project, is the Multi-Centre Epilepsy Lesion Detection project. And we have 40 hospitals around the world who are sending us data, sharing their expertise, and I think we now have data from every continent. So, we are really trying to make it as diverse as possible. There's no hospital that's too small to take part. So, we're trying to bring everyone along with this. And we've seen that there's, you know, it can have really big impacts. So, at Great Ormond Street where we work, we have world leading neuroradiologists and we can help them find these lesions. But we've also worked with sites where they don't have that expertise. They don't have a specialist epilepsy radiologist and there the benefits are potentially much greater, as you say. And we're working with teams across the world as well to try and figure out how best to solve these problems.
12:06 Torie Robinson
And I mentioned the more data collected or used -which would be completely confidential, correct?
12:12 Konrad Wagstyl
Yep.
12:13 Torie Robinson
he better the system becomes at identifying these lesions. Is that right?
12:18 Konrad Wagstyl
Yes, exactly. So, we now in the first project, I said we had a thousand, we've now collected 2,800, we're still collecting more. And every time we get, increase the number of patients, particularly for some of those rarer types of epilepsy, then you start to see even better performance.
12:33 Torie Robinson
Can you just tell us a little bit more about that then, the countries that are involved? Because this is like not overly common to involve all these different countries.
12:40 Konrad Wagstyl
Yeah, we have a world map and I wouldn't want to miss any. And as I said, it's every continent. we have fantastic collaborators in Australia, Japan, China, India, trying to go work my way across the world.
12:52 Torie Robinson
Hehe.
12:52 Konrad Wagstyl
South Africa, lots of European countries; Spain, Germany, Denmark, Portugal... I don't want to miss any, but there's many, many…
13:16 Torie Robinson
Yeah, yeah.
13:16 Konrad Wagstyl
…and obviously lots in the UK. And then if we go over to the Americas, we have North America, Canada, Chile, Argentina, Brazil, possibly Argentina as well, but I'll have to check that.
Become involved in the project
13:17 Torie Robinson
So, if anybody would like to get involved in your research, whether they be a clinician, a researcher, or whether they be somebody with an epilepsy or a family member, what should they do?
13:25 Konrad Wagstyl
We are very open to all of those three groups of people. So if a clinician wants to be involved, they could either do a research study using our tools, get in touch with us, we'll help you download them, or if you'd like to contribute data, also we'd love to have you as part of the MELD project. The same is true for researchers. We often have people come visit us and work with us on the data set. We are a really multidisciplinary group. We have computer scientists, engineers, clinicians. We need everyone to make these things work. And then, patients, we have this patient advisory group and we really value the expertise that guide us. As I said, they're the ones that have been pushing us to make it work on first epilepsy scans. All of these types of decisions are made with our patient group and so we're always open to new members joining.
14:11 Torie Robinson
Well, this gives realistic hope to so many people around the world. And I would say as somebody who's also a patient, I would be up for donating any sort of data that would prove useful. And that's what we need to improve diagnosis and care in the future, correct?
14:27 Konrad Wagstyl
And also, just your perspectives are always valuable. Every patient's valuable in guiding our research too.
Closing thoughts & thanks
14:33 Torie Robinson
Thank you so much to Konrad for sharing his fascinating research into how AI and neuroimaging can help detect subtle epilepsy lesions that are often missed on MRI scans - which can support earlier and more effective treatment decisions and improved quality of life.
Again, huge thanks to EASEE®, by Precisis GmbH, for partnering with us at Epilepsy Sparks.