Hi, everybody. Thank you very much for the organizers, Mitch, Hugh, and Richard, for inviting me. I'm going to talk today about Magnetoencephalographic imaging and how we use it at UCSF and some of the valuable information we get preoperatively that could be used in conjunction with interoperative mapping. I'm going to talk first about—this is not working, I guess—introduce Magnetoencephalographic imaging. For those of you who may not be that familiar with the technology, as it's not widely available, a few centers do it, and we at UCSF have been pioneering some of the efforts in integrating this with preoperative mapping. And then I'm going to talk examples about resting state connectivity that we do with Magnetoencephalographic imaging, and then talk about some of our ongoing work on speech and language mapping with MEG. So the first part, just what Magnetoencephalographic imaging is, is recording magnetic fields and reconstructing whole brain activity videos of brain activity. So earlier we heard about four-dimensional mapping that Eddie and Bob have eloquently pointed out. What they actually haven't told is that actually brain activation is five-dimensional at the minimum, if not eight-dimensional. The five dimensions are space, three dimensions, and time, which we've talked about, but also the different oscillatory frequency bands fluctuate at different time scales in different brain regions. So here's just, for instance, an example of a snapshot of the brain at about 200 milliseconds after you listen to a speech sound, and we see bilateral activation of auditory cortex. If you zoom in at any one of these regions, you can look at the color map where time is on the x-axis and frequency is on the y-axis, and power fluctuations are shown in color. You can see low frequency fluctuations, bilateral auditory cortex. If you just focus on one of these color plots, you can see the low frequency is the first band in the bottom, and then you can step up the alpha band, the beta band, and the high gamma band, and we can look at power fluctuations in each of these bands at any voxel in the brain. And we can do this non-invasively now with magnetophotographic imaging, which is combining MEG recordings, which is enabled by these superconducting quantum interference devices, which are just pickup coils made of superconductors and just some junctions, and we have about 300 of these covering the whole head. It looks like an old-fashioned hairdryer that you place your head in, and we record the magnetic fields, but what we are able to do is to use, harness the ability of machine learning and physics and signal processing to be able to reconstruct from these recordings outside the brain videos of brain activity. What's really enabled this technology in the past decade, I would say, are harnessing and leveraging powerful breakthroughs that have occurred in imaging algorithms that leverage progress in machine learning and probabilistic graphical models, and various techniques that are used in self-driving cars and Siri. All of those things can also be applied to imaging, and we and others have done it where we're able to use this information to reconstruct activation sequences with higher fidelity than has been done before, and also we can reconstruct functional connectivity, which I'll show you in just a bit. These advances open up powerful applications for MEG imaging, both in basic neuroscience and clinical neuroscience research that I'll talk about. Just to give you an example, somewhat similar to what Eddie had showed earlier, is if you're listening to a speed sound, and you can talk about listening in different phases of a task. Here's just looking at one second split into four quarter of a second time scale snapshots where we can reconstruct the brain activity at each of these snapshots. We think of listening as happening in the first 300 milliseconds, but actually it's quite dynamic within that time period. You can see also in this, this is reconstructions from MEG recordings during one second of listening to speed sound, and what you can see is I'm showing you average of multiple frequency bands and statistical thresholds that are reliable activations, and you can see things that Eddie had reported and Bob had reported in ECOG and others have reported in fMRI, which is within the first 200 milliseconds of listening to speed sound, you are actually activating the motor system as well. So you can see here in the first plot on the left, you can see the motor strip being activated while you're listening to speed sound. So this is something you could start to do looking at speech perception, looking at action perception connections with magnetophysiographic imaging. We can also image the resting brain. Most of you know that as you rest, your oscillation, neural oscillations are dominated by the alpha band. We can look at not just the power of where the alpha band is coming from, but we can also start to look at the connectivity of different voxels that are fluctuating in the alpha range. So if you, that's what's shown here is that you can actually plot the, this is actually what's shown here is a directed graph of each voxel. It's showing its alpha band connectivity with other voxels in the brain. So you're looking at a mean directed weighted connectivity of this graph, and what you can see that it reflects the alpha posterior, alpha network, and we've used this now. This is very easy to obtain. We just have subjects come in, they close their eyes, and they're just ruminating, and we see this idling network that we can reconstruct. What motivated this was a postdoc of mine, Adrian Gogesberg, who's a neurologist in Geneva, took data from one of our brain tumor patients, and we had this massive brain tumor, and he developed this technique where he was looking at the relative connectivity of resting state connectivity in alpha band of this tumor, and found that it was largely disconnected from the rest of the brain, and consistent with that disconnectivity hypothesis, resection of this massive tumor in this patient produced no deficits. And so we published the technique, and our first finding, an analysis of neurology in 2008. Since then, we have done a large series of these questions. The first question we want to ask is, is this sort of resting state cortical connectivity, is that helping predictive of surgical mapping results? So here's a paper led by Juan Martino, one of our visiting neurosurgeons, along with Feroz and several others, where we looked at a series of patients where we had both positive and negative connectivity, preoperative mapping, interoperative ECS mapping, and preoperative resting state MEG. And what we found, and this is just showing you three cases, the one on the top left is a patient where there was a mapping site for speech arrest, and around that site you had high functional connectivity shown in yellow, and low functional connectivity shown in blue in these maps. So the top right plot is another site where you found responses to sounds in intercortical mapping, and the connectivity around that was actually high and positive. Shown in the bottom is a site where there was negative mapping, there was no site that was deemed to disrupt language function, and these sites, there was really low connectivity in this mapping. So in this paper we showed that there was great high negative predictive value, that is, when you found regions of low connectivity preoperatively, it is highly likely that your interoperative mapping will be negative in the standard tasks that one does. We also wanted to find out if cortical connectivity is predictive of long-term outcome. We saw a hint of that in the first series that Adrian did, so we did a larger series led by Feroz during his research year in the lab. And here's just an example, I'm going to show you three cases. Here's a case where resection in this patient showed preoperatively very low connectivity shown in this sort of blue map, and this patient had very good recovery. Another patient had very high regions of high connectivity in eloquent cortical regions and had poor recovery. And here's a very interesting case where this is a patient who had high connectivity in the margins of the tumor, but low connectivity within the tumor, and this patient actually had very good outcome in recovery. So when we look at our overall cohort, we found we could categorize, we did a first step at this looking at preoperatively, we could categorize patients as with decreased connectivity or statistically neutral connectivity of the tumor compared to the contralateral side or increased connectivity. And we found was there long-term, which is post-six-month morbidity, and we find that in patients with decreased connectivity, none of our patients had poor outcome. So this suggested that there was very good negative predictive value. In terms of positive predictive value, our results are 60-30, and this is because our assessment of functional connectivity is unrest, and our assessment of outcomes are language tasks. So it suggests that these regions of high connectivity might be implicated in non-language functions. Recently, Dario Englert, another neurosurgery fellow who did his research here, published looked at patients with epilepsy, and in epilepsy patients, again, so far we've been looking at alpha-band connectivity. So one of the things we've known, and it's been known by many other studies, is that you have alpha slowing in epilepsy, and we were able to replicate that. And what Dario found is that there's widespread disconnectivity in epilepsy. So this is looking at the global connectivity, which is asking what are each voxel, how is its connectivity with the rest of the brain. And what we think is that this is a normalizing mechanism, so you have an irritable zone that is causing high excitation, and one way to compensate for this sort of irritable zone and its impact is to overall reduce the global connectivity in the brain. What Dario also found was that you can derive a single number for global connectivity of a particular subject, and we were really excited and surprised by this result, is that this degree of global connectivity was correlated in a multi-factor learning model with duration of epilepsies, as well as frequency of consciousness-inducing seizures. That is, subjects who had reduced connectivity had, so you can see here, what's plotted on the x-axis is the frequency of a seizure, and the y-axis is plotting the global connectivity. So subjects with reduced connectivity had fewer seizures. So we were also able to look at regional connectivity, and this is an example of where we looked at the increased, we looked at the connectivity of the epileptogenic zone, and we found that when you have increased connectivity in the epileptogenic zone, and it is resected, it is associated with favorable post-operative seizure outcome. We have several other disease populations where we've looked at resting state functional connectivity. In the interest of time, I don't have to show you that, but we have some very exciting results in dementia, and schizophrenia, and neurodevelopmental disorders, and stroke, and traumatic brain injury. Overall, we're finding that there's great value in using resting state functional connectivity. It's a very simple assessment. It's an alternative, and perhaps a complementary assessment to resting fMRI in stroke cases where we have vascular confounds. We find that MEG resting state connectivity actually provides more useful information than fMRI, which has these vascular confounds. We can also do speech and language mapping in an MEG scanner. It's basically, you have an awake behavioral setting. So we routinely now do auditory verb generation tasks in our preoperative mapping. Our first efforts at this were to try to see if MEG could be used to perhaps replace or substitute WADA assessments. And so when we first did this, we looked at a cohort of confirmed patients, confirmed by WADA. You can question whether it's an appropriate gold standard. Nevertheless, we took this as a first effort to look at patients who we knew were left dominant and patients who we knew were right dominant. And what you're looking at are snapshots of every 100 milliseconds after they were listening to a speech sound. In this case, this is an auditory verb generation task. So they're listening to a noun, and they have to produce a verb. You could look at time windows when you're listening, or when you're speaking, or when you're preparing. And we can see that patients who are left dominant show a clear pattern of left lateralization. And patients who are right dominant are largely bilateral, but right dominant as well. And we now use this protocol routinely to determine language dominance, and that could potentially make an impact in whether or not awake surgery is done or not. We've used this in a variety of research projects. I just want to highlight a recent paper that's coming out in General Neuroscience in the next month or so, where we looked at patients with colossal agenesis. And we find that the absence of a corpus callosum really shifts and changes language dominance. So here, we're looking at 300 to 700 milliseconds after listening to a speech sound, and you find in control patients, if you look, the top is the control. The second row is patients with colossal agenesis. The left hemispheric activation is very similar. But if you look at the right hemisphere, you can see that control patients really, on average, do not show right hemispheric activation, whereas these colossal agenesis patients do. And you can see that if you categorize laterality index plotted on the left is the healthy controls laterality index. If it's left-dominant, it's about the axis zero. And if it's right-dominant, it's below. You can see that patients with colossal agenesis, with partial colossal agenesis, are bilateral. And complete agenesis, many patients show right lateralization. So which is very exciting, it suggests that colossal connections, which might be impacted in surgery, you could impact language dominance and harness potentially contribute to brain plasticity, as we heard from Hugh's talk earlier. So the summary of this portion of the talk is we've shown that MEG imaging demonstrates the transient nature. I forgot to mention this, is that one of the realizations we did, when you're able to image the brain activity in, say, this tens of hundreds of milliseconds, you realize that the notion of language lateralization is actually a very transient phenomenon. Your brain is bilaterally active at the time of listening and at the time of speaking. But its dominance, language dominance, occurs in this intermittent window between listening and speaking. So we have to exploit that window almost in a subject-specific manner, also in a task-specific manner, depending on if you use a verb generation task or a picture naming task, the regions you extract to derive language dominance could be different. So we can know, hopefully I've shown you that this can be accurately now reconstructed with MEG. We can also do higher-order language functions like verbal working memory. I'm just going to show you a little bit, very similar to what Eddie showed earlier about repetition. We do a syllable reproduction. So this is nonsense word reproduction task. We don't do this clinically, but I wanted to show this to you. We can use subjects' hear sounds, which are like ba, da, pa, ba, and they have to reproduce them. And we can look at the windows of listening and speaking and look at accuracy in performance of the task. And we can ask which brain region's activity correlates with accuracy in performance in this task. And we can reconstruct a largely dorsal speech motor network that are involved in neurobehavioral correlations, which we can look at either in a stim lock manner or a load effect where we can change the memory load. So a syllable reproduction task is a potentially powerful task. If you have high-resolution imaging like ECOG, you can use this to really unravel a large network involved in verbal working memory. So the dominant view about verbal working memory is that we hold things in this phonological loop, which is considered the cycle between Wernicke's and Broca's area. Using the techniques that we have developed, what we show is there's actually a very interesting orchestration or reverberation between Broca's and Wernicke's area that support accuracy in verbal working memory. What I'm showing you here, blue is the activity in putative Wernicke's-like regions, and red is activity in Broca's-like region. And I'm plotting correlations with performance over the course of the task. So zero is when you start listening to a speech sound, and then there's a break because of reaction time. And the zero on the right is where you're looking at the response when you're about to speak. And if you look from the right, which is just before you're about to speak, you can see that there's activity in Broca's area that seems to be correlated with the task. And that wanes, and you have activity in Wernicke's area about 200 milliseconds before you're about to speak. And if you go further back, Broca's comes back on. If you go even further back, Wernicke's comes on. So you have this sort of orchestration of a phonological loop reverberation, which is very interesting. Most people think of these networks of being active all the time or sustained activity, which might be wrong. We know that these regions are, in fact, impacted because of huge talk in plasticity. I wanted to highlight two things about brain plasticity. Brain plasticity in disease could be profound. I'm going to show you an example that's completely non-neurosurgical. There are patients with schizophrenia who are doing this task, verbal working memory. They're just repeating speech sound. And we see that these patients activate ventral visual word form areas. This visual word form area is actually an interesting area, which has been shown to be harnessed in multiple domains where you have plasticity, in braille reading subjects, in deaf subjects, and so on. In a purely auditory task, this area is activated in patients with schizophrenia. Not just activated, it actually helps with correlating with performance. There is Wernicke's, like superior temporal regions, which in healthy controls are correlated with performance. But in patients with schizophrenia, only those with low hallucinators use this standard Wernicke's region. High hallucinators, the correlation between performance and activation of this region breaks down. But if you look at visual word form area, in healthy controls, there's not much of a significant correlation. But in high hallucinators, this area is highly correlated with performance in this task. So a disease can impact, and we saw evidence surgery could also potentially impact, large scale impact, beyond sort of traditional networks that we see. One of the things I just wanted to highlight at the end of my talk is that intense computerized auditory training also is a means to impact cognitive performance and neural networks. So we've done this study with a collaborator, Sophia Venerado, who's a psychiatrist, where we looked at patients with schizophrenia and we make them do intense auditory-based computerized training, and we showed large scale cognitive changes, large scale improvements in performance, and improvements in performance are related to cortical activation. That's what is shown here, is that to train patients, cortical activity are directly correlated with performance improvements and changes in cognition as well. So it's something that you can think about as adjuvant potential, adjuvant in your patients as they live longer because of the improvements we've done in surgery. So the summary of this part of the talk is that these dorsal speech networks can be reconstructed during cerebral reproduction. We can show activations, we can look at brain behavior correlations, and these things can be altered in disease, can be altered with training. So just to conclude my talk, MEG imaging, I hope I've convinced you, provides a powerful platform to examine links between brain function. We can look at everything from the relationship between genes to behavior. The first part of my talk I talked about resting state functional connectivity with MEG as a useful tool for cognitive neuroscience studies in clinical populations, and the second part we focused on using MEG to enable, which enables inferences by brain dynamics supporting speech and language function, including hemispheric dominance assessments as well as higher order language function like verbal working memory and also examining its plasticity. I want to end by acknowledging several of my collaborators, Mitch has been a great supporter and a lead collaborator of ours, as well as several neurosurgeon. I've had more neurosurgery fellows and residents in the lab than I've had radiology residents, which speaks to the nature of our collaboration, as well as Bob and Eddie and Feroz who are in the audience here, as well as many others in the lab. Thank you.

Magnetoencephalographic imaging

Preoperative brain mapping

MEG recordings

Brain activation

Functional connectivity

Resting state connectivity analysis