Advances in Treatment for Epilepsy-Visualization Techniques to Identify New Surgical Targets in the Very Young Population

Video Transcription

So my talk is a bit more technical, and I'm not going to spend a lot of time on the exact rationale.  I'm just going to give you an overview of some of the techniques that are currently under development that could potentially allow you to find targets in younger children and  infants in whom you're planning stereo EEG or other types of interventions. The brain undergoes a stage of intermediate myelination from roughly six months to one and a half years.  And by age two, the MRI is properly myelinated, or at least to the extent where the contrast between the gray and the white matter is returning.  But before that, you need to be able to do surgery, and it's not always obvious. So I'm going to talk about some evolving techniques based on networks, based on convolutional  neural networks, and based on source imaging, just to give you an idea. Here's my disclosures. They are not relevant to the talk today.  And at the end of the talk, I want you to have a rough idea of changing modalities that are evolving into the use of earlier localization.  There is a trend towards machine learning and deep learning, and that is being applied to neuroimaging, including in epilepsy.  And I will cite from other people's work and some of my own. And finally, recognizing that while we are here focused on resections tailored towards  the generator of the seizure, sometimes if the generator is unclear or the network is already very advanced and expanded, we may be able to address key nodes of the network.  So there's work to discover these networks that facilitate either the propagation of the network or to become dysfunctional networks on their own. First, again, the why.  We've discussed this. And again, this is about TSC. So you can see how we all became friends and how you have basically a lineup of four people in the TSC field lined up today.  So I apologize for that bias, and next year as a chair, I'll make sure to include a bit more of a broader thing as well. No offense, Rene.  But the why is important, and here's an example of TSC. And the TASRIN study monitored 160 children prospectively in five centers across the nation  for six years, five or six years, with serial EEG, serial neuropsych eval, serial MRI. And when we looked back, because it's not a part of the formal study, we found 20 children  to 19 children that were operated under age two at a median age of about one and a half years old and were followed for nearly two years, at a minimum follow-up of 12 months.  And they had similar favorable outcomes, as we were mentioning before from Dr. Sedo. The poor outcome was also present.  And I'd like to show you these slides here at the bottom that did not make it in that form into the paper.  But basically here in purple is the Mullen score of patients that did not have any seizures with their serial exams over time plotted at the bottom.  And in blue, patients with epilepsy over time that did develop seizures but were controlled, and the uncontrolled in green and red.  The red, refractory epilepsy that did not undergo surgery, and in green, those that underwent surgery at approximately a median age where this truncation exists.  And you can see the change of the direction of the curve. Of course, that could be in part a fitting effect.  But we see this trend in these raw scores in language, receptive and language scores as well. And it was very consistently present over the cohort.  So it's kind of rare that you prospectively can track developments and see the effects of an intervention like this.  Now, I think the most important message may actually be hidden here is that there were many, many patients with ongoing seizures that were not specifically targeted.  So when your surgery is not successful in the sense of an ENGEL score because patients have ongoing seizures, it doesn't mean that you don't need to operate.  You can operate on the most devastating seizure type or the most burdensome seizure type. And these patients still did remarkably better and are incorporated in this developmental  improvement curve. All right. So I'll talk about source imaging, neural networks, and lesion mapping. So at its very core, localization in EEG is just looking for phase reversals.  And you can see this juicy spike at F4, C4. And then maybe with a field to FZ and probably a little bit of frontotemporal field as well.  And that used to be the core of localization. And thankfully, now, thanks to Dexter, we can do retrospective research and figure out  based on the blood pattern on the wall, we can deduct and calculate backwards where the gun was fired and entered the victim's brain.  So similar to that, in electrical source imaging, you generate a lead field matrix, which is basically a hypothetical source in every voxel in the brain, come up with a big  grid of what those sources would look like on the scalp. And then just like Dexter, you calculate backwards once you have your EEG signal where the original  signal may have been generated. So it's a reconstruction.  Now, this solution is fundamentally ill-posed because normally you have only 18 or 19 or 21 electrodes on the scalp surface.  And you have an ill-posed problem, which means that there is an infinite number of solutions that are equally feasible, at least from a mathematical point of view and physical point  of view, to explain the surface signal. And here in the cartoon, you can see that a perfectly legitimate dinosaur was reconstructed  thousands of years or millions of years after the last dinosaur was roaming the earth. And it is a very feasible and even believable dinosaur, but it is not correct.  So when you have a large number of possible solutions and not enough data, you need to introduce constraints to the model or get more data.  And based on that first approach, we said, well, in situations where you're dealing with multi-lesional epilepsy, and we are assuming that one of the lesions or more is primarily  responsible, why not constrain these lesions not just to the cortex, but actually to the lesions?  So here is a situation where we built a Lietveld matrix for the whole brain, then had lesions, in this case tubers, individually compete to show us visually which of these lesions  is best able to explain the abnormal surface signal. And we translated these into color maps.  So if you add enough components to an independent component analysis, like ICA or something like that, you can mathematically eventually get a fairly good fit of every lesion.  So we included a penalty for mathematically complex solutions. And this is what the final product looked like.  And I think the easiest example to see is at the bottom left, where in black there is a resection cavity superimposed on the preoperative brain, as well as the source solution with  Loreda superimposed on the preoperative brain. And you can see that the core of the distributed source solution is somewhat scattered and  focuses along that rim of the future resection cavity. Now if you force the model to think in boxes, basically, you can see that the hottest lesion  is that red blob on the image right next to it. See if I can point it on the big screen here.  And the same goes for solutions that are on the border of the frontal lobe and the temporal lobe, where a distributed source solution cannot decide.  If you tell the computer, well, there is a tuber sitting there, then that will become the hottest tuber based on our methods.  Now, there are many drawbacks of these methods, and it's at best complementary to conventional source localization. But it is a good visualization.  And we're expanding on this tool, for example, by using high-density EEG for better fitting. And here's an example of high-density EEG.  So here is a low-frequency sampling of just 19 electrodes. And here is a high-spatial sampling. You can see that this map is already much more diverse.  Here is a dipole found with the low resolution. And then with the lesion-constrained solution, you can see how nice the tuber comes out in  both the distributed source solution and in our lesion-constrained analysis. So again, this is just one example of how creative solutions can improve visualization  techniques in young children. Another one is the use of convolutional neural networks in either MRI-negative epilepsy or  in MRI-positive epilepsy, where there are, in hindsight, lesions actually detectable. And a quick word about how these networks work.  So with machine learning, one basically applies statistical modeling led by humans, supervised  by humans, and you apply increasingly sophisticated models on data that has been broken down into specifically predetermined features.  So for example, an age range, or a head circumference, or a calcification, for example, in MRI.  And you can determine these features and then do modeling, including things like linear regression and then ever more sophisticated, and come up with a good fit, and therefore  a predictive value of your model. In deep learning, you take a step back, and the humans are less involved, and the features  are not predefined, and you let the computer figure out some of these imaging features in particular themselves.  And convolutional neural networks is a form of deep learning where convolutional filters are used, and that's what I briefly wanted to talk about here.  So the reason why, in machine learning, that the imaging world seems to be steps ahead sometimes of other applications of machine learning and deep learning is because of the  use of images that have not only a hierarchy, meaning that an image is built up of complex constellation of lower level features, which I'll explain, but also locality.  So we have two or even three-dimensional data where we make use not just of the fact that a complex picture, like in this case a cat, is built up of multiple little cat building  blocks, but also that a cat has ears at a certain location and eyes at a certain location. So there is two levels that we can capitalize on.  So for example, here is a circle filter, and there is an arch filter, and in the next level these simple features and shapes are combined into something that maybe resembles an eye.  And if you feed the computer thousands of images, at some point it will say, wait a second, we see this shape every time that somebody told me that there's a cat visualized.  and then in fact in the next level of machine learning where the next layer you can say oh wait a second there's an there's the set of eyes if we  combine these lower level features and you can have an air filter and so on and ultimately these building blocks build a cat. So we see a cat but the computer  sees numbers and colors and walks with these filters over these images and for example to detect edges and edges are important because that means that the  voxel or pixel next door is not similar to most of the other ones and therefore signifies a more important point in the slide. So you can imagine for  example here an ear and eye and a nose filter that will help detect and form the basis for recognition of these images. So rather than a cat  filters if you apply tuber filters that are self-learned by the computer we can identify lesions in TSC so here work from Ivan Sanchez Fernandez at Boston  Children's where with a high reliability but this was compared to normal controls then you can do transfer learning which is basically that you in a rare disease  all these machine learning algorithms have thousands and thousands or sometimes millions of images that they need for the machine learning to sorry  for the deep learning to train these networks but if you have a rare disorder how do you go about so one of the tricks is to replace the final layers in the  network here indicated in purple that are very specifically trained to a task so there's an image database that has millions of images the convolutional  neural network gets very smart but then in the last steps of the neural network the last layers it learns to recognize let's say different types  of sailboats and you're like well that's not what I'm interested in so you say well this network is performing so well but I'm going to take a few steps back  and replace the last layers with and train them on my tubers relatively rare disorder and then you can retrain and do the same thing so we trained a TSC  algorithm on multiple sclerosis patients another trick that you can do in rare disorders is actually to overrule the human annotation so when you feed the  convolutional neural network images that are imperfectly annotated you're basically training it with wrong data noisy data and if a computer finds a  high degree of certainty it can overrule what the human had originally marked in the training data set and retrain itself with better data and we did that with a  bunch of cases in TSC and a smaller data set that was it was correcting these noisy labels performed as well as a really large data set that that did not  allow to do that and then here is where this talk about the convolutional networks was actually heading is that we can detect subtle focal cortical  dysplasias in non-lesional MRIs in this example for example from the Canadian group in Montreal there is this was done on 148 cases from nine centers with  either MRI negative workup in about half of them and others had a subtle lesion that was visually detectable in hindsight but all of them were  histopathologically confirmed after surgery and they trained the networks on recognizing these areas and expressed the lesions in clusters that you can see  here with the probability and then ranked these groups of detected focal cortical dysplasias based on how certain they were so it wasn't binarized  it wasn't saying here is one and here is none it was a probability map which is basically the way you also think about the likelihood of finding a lesion in  practice and inviting you to think critically before you add for example a stereo EEG electrode and so here is is one example it's really small on the  slide forgive me but see where my mouse went here's an example where we can see the small lesion here in deep purple and compared to a less likely lesion in blue  down here there and you can also get an idea of when you put all of these in the common space of how focal cortical dysplasias are typically mapped this top  image is actually based on the actual resections and not necessarily on the detections from the algorithm and then the false positive rate also has a  certain distribution and it was interesting to see that this happened mostly in the insula and in the perihippocampal gyrus which may be areas  where you know the gray-white border is a bit more fuzzy and you can see that these features are picked up and falsely detected more so the false positive rate  in this cohort was about six per patient so that's not an insignificant amount of work for a neuroradiologist it also means that you cannot just add six  electrodes per se so it needs to undergo some critical thinking which is nice why it's working with probabilities in this case the cluster that was indicated as  being the most suspect or the highest degree of certainty was part of the ultimate resection in 75% of cases and finally I wanted to take a minute to  talk about how to localize these networks that may pertain to generating seizures and we went about to localize in vital spasms with a technique called  lesion network mapping and we again examined 123 children in the TSC a Tassern study that were prospectively monitored for for several years and we  had 71 cases sorry 74 cases of infantile spasms in this case so why do some patients turn out to have spasms and others do not we came up with the lesion  segmentation based on these networks that I just showed you earlier and we projected these lesions into a common space here the lobes are color coded and  ultimately can come up with a distribution of the tubers over the brain and you can see that it is quite stochastic although I deep down I  wonder if there is a sparing of the motor strip somehow and perhaps also primary visual areas which is just interesting and then when you separate  these distributions by having spasms or not at the naked eye there doesn't seem to be a lot of difference between these two distributions so it doesn't seem to  matter where tubers are necessarily located but then what is the answer why are some patients developing spasms and others don't and when we did the math  and tried to find how which voxels or which tubers would place the patients at risk if you just look at the heat map the best product predictive value was  about 24% so we could not predict it from the lesion but then there's a technique called lesion network mapping where basically if you have a tightly  defined phenotype rather than putting all the locations in anatomical space into a common space like the MNI brains you can actually figure out what these  lesions would theoretically affect which networks they would theoretically affect in a normative connectome based on 1,000 fMRI resting state data from adults in  this case we also built a pediatric cohort but so for example these patients have something called peduncular hallucinosis a rare stroke syndrome in  adults but one patient has a lesion here and in the basal ganglia the other patient in the midbrain and here is a lesion that seems to be in the in the  pons and the question is if you put these anatomical lesions on top of each other say well what's the common denominator in these really specific  phenotype rare stroke syndrome then there is either no overlap or very little overlap but when you look at what networks are inversely connected in blue  or correlated in red on F based on fMRI data you can see that when you place all these networks on top of each other and then threshold it it turns out that all  of these lesions have all of them talk to the basal ganglia right here so what's common is not the location of these lesions but what they are connected to  and when we did that for the our cases we found that in 71 out of 74 patients the tuber distribution pattern was functionally connected to the internal  globus pallidus and the vermis of the cerebellum and then we turned that question around and we said which foxhole is functionally more often  connected to tubers in patients with infantile spasms as compared to no infantile spasms to see how specific this finding was and in the regions  where the that are both blue from the previous page here and and yellow from the convergence analysis the findings are both sensitive and specific and I'll  stop here because it gets a little too detailed but basically what we are able to do is to figure out in a multi lesional epilepsy where you cannot  necessarily find which lesion is causing the problems we can identify key nodes of the network and maybe make them a minimal to future and neurostimulation  and this is you know data in from infants and we're not using EEG necessarily so you don't always depend on invasive monitoring to end a quick  case this patient had seizures on the left it and very subtle right here a little blurb and it illustrates very well why EEG monitoring in my opinion is  a dynamic study it's sometimes hard to do and appreciate things afterwards which is why we need to capture five or seven or nine seizures before you  sometimes are convinced of a pattern being truly ictal and corresponding consistently but this patient also had seizures from the right and he was  really quite sick and presented to the ER and to clinic remarkably often so we applied all of our tricks including high-resolution MRI the PET scan  high-density EEG source localization including lesion constraint approach and he was operated both on the left and with a laser and on the right side with  a with an open resection within one month from each other has been seizure free since with with very rare breakthrough seizures in case of fever  or missed meds and things like that and that's him and he's doing very very well he's wearing a Texas Children's shirt as a salute to Rohini and team thank you very much

Video Summary

The presentation focuses on emerging techniques for localizing brain targets in young children for treatments like stereo EEG. Traditional MRI assessments are limited in children under two due to incomplete myelination. Advanced methods such as convolutional neural networks (CNNs) and source imaging are being explored. CNNs and machine learning offer probabilistic mapping for identifying potential brain lesions and networks involved in epilepsy, especially when traditional imaging fails. Case studies highlight how these technologies improve diagnosis and treatment, enabling more targeted interventions. Lesion network mapping is used to identify key nodes connected to symptoms in multi-lesion epilepsy cases. These tools allow for early intervention, potentially improving developmental outcomes by focusing on the most severe seizure types, even when ongoing seizures persist post-surgery. The talk underscores the shift towards machine learning in neuroimaging and its application in pediatric epilepsy treatment.

Asset Subtitle

Presenter:

Jurriaan Peters, MD, PhD, is the Program Director for the Clinical Neurophysiology Fellowship and Epilepsy Fellowship Programs. His research in the Laboratory of Translational Neuroimaging and in the Computational Radiology Laboratory focuses on novel MRI and EEG modeling techniques in the localization of the seizure onset zone in children with medically refractory epilepsy, on how focal lesions give rise to a generalized epileptic encephalopathy, and on early medical and surgical treatment of epilepsy to mitigate detrimental effects on neurodevelopment.
Dr. Peters is an Associate Professor of Neurology at Harvard Medical School, and is a pediatric neurologist with additional expertise in epilepsy, clinical neurophysiology and neuroimaging. He earned his MD at the Catholic University of Leuven, Belgium, and his PhD at the University of Utrecht, the Netherlands. He did his Child Neurology training and Clinical Neurophysiology fellowship training at Boston Children’s Hospital.

Keywords

pediatric epilepsy

stereo EEG

convolutional neural networks

lesion network mapping

machine learning

neuroimaging