Well, let me begin by first obviously thanking you for the opportunity to speak about some of the stuff that we do here at Washington University. And so as far as kind of some of the kind of innovations, what I'm going to focus on is using resting state fMRI for operative planning for tumors adjacent to eloquent cortex. And again, I think we've gotten some, you know, flavors of how resting state can influence it in some of the previous talks, but this is also very much a technology talk in that how do we bring this to the real world? How do we kind of transition some of these exciting scientific findings, these scientific capabilities into actual technologies that we use on a day-to-day basis to enhance our ability to map function? So disclosures, I own some stock in some companies, I'm a consultant, I have some intellectual property. None of this has a financial relevance to this talk. And so fundamentally, and I think we've been hearing in the last two days, that we're always balancing this tradeoff of cytoreduction against functional preservation. Remove more, remove less. And some of the classic sites that we're always making this decision on is around motor cortex or speech cortex. Now historically, some of the classic techniques that we've talked about is stimulation, evoked potentials, and they have their limitations. They certainly have their strengths, but they also have their limitations. With evoked potentials, we're limited to motor cortex. With stimulation, it's iterative, you can do one site at a time. There's after discharges, people have mentioned the risks of seizures. There's a limited repertoire, and some people may debate that, but still, you know, it can be challenging beyond the classic motor and speech areas that we map. So now one of the things I think that can really expand our capability of mapping function is using endogenous brain activity for brain mapping. And one of the techniques that I'm going to talk about today, there's others, but the one that I'm going to focus on today is using magnetic resonance imaging, or MRI, resting state networks, and how we can use that to identify function prior to surgery. Now classically, we've always used functional MRI in this task-based sense. And, you know, again, a classic example shown here kind of by Biswall in 1995 is, again, a finger-tapping task. And what you see here, let's see, is that basically you typically in task-based stuff is you do task versus rest in a block design. You average those two states, and you get something where you see kind of on average that there's a blood oxygen level dependent change, and in this case, in motor cortex. Now what Biswall discovered back in 1995 was that if you now look at that blood oxygen level or bold signal, that it has a natural fluctuation to it. And if you pick a seed or a voxel over motor cortex, that if you look at all the other voxels that are co-fluctuating with that, it will recapitulate the motor cortex on the unilateral, on the kind of ipsilateral side, and on the contralateral side, and that that matches up very closely with the task-based activation. And so now you've got a way to basically identify some functional cortex without the person doing anything. And Biswall compared early on with motor cortex that both the task-based stuff matched up very closely with the resting state correlation areas. Now I think what's interesting about intrinsic, meaning this kind of endogenous activity versus evoked activity, is that only a small fraction of energy is actually used in the evoked activity. The vast majority, 90 percent, 95 percent, is in this endogenous activity. And so also this intrinsic activity is quite difficult to suppress. What I mean by that is whether you're awake, you're asleep, you're sedated, or you're under deep anesthesia, these resting state networks are preserved, so that really gives you the opportunity to map people regardless of the state that they're in, and that has relevance certainly to pediatrics. Now how do we do this? So I'm going to just give you some kind of, you know, simple examples of some of the early ways that we do it. So first off, I'm just going to represent this rather symbolically. Imagine each of these dots is a voxel in the brain, and basically what you do is you pick a seed. You say, okay, I'm going to pick this voxel. I'm going to find the other areas of brain that correlate with it. And then basically that correlation structure is a network, and you can map that onto the brain, okay? Now people have been doing this now for the past decade to decade and a half, and there's about seven canonical networks that really define the functional architecture of the brain. And again, you've heard mention of the default mode network. Again, this is the network that is really one of the most dominant networks in the brain. It's associated with internal referencing, memory, and it's one of the networks that primarily degrades with Alzheimer's disease. So it's a very important network in our brain. Frontal parietal control, that's associated with executive function, inhibition of behavior, language which we're all familiar with, ventral and dorsal attention, how we deploy attention to the world either internally or externally, somatomotor, again, something that we're familiar with, and visual network. Those seven are what make it up. Now how do we apply this clinically? Now again, one example, kind of an early example, was we can use this for mapping. In this early, early patient, the patient was very good in the MRI, we put him in there for 12 to 15 minutes, you get a single scan, and now you can get multiple networks out of that single study. And in this case, it took a tech around one or two days, you've got to plant a seed in as I mentioned, and then you can map language, default mode network, attention, vision, and auditory function in this particular patient. But it took work. It took a technician, it's kind of done ad hoc out of a scientific lab. Again, another early example in this patient, and I think what's interesting when you compare it to a task evoked MRI is that, so if you look on, again, we've got our structural MRI with the tumor, we've got task evoked, and you see that, again, this isn't a finger tapping example, you see a bunch of areas lighting up, some of that looks like motor, some of it doesn't. And I think what's interesting is when you look at the resting state network, you see an area that pretty much lines up with motor, but when you look at the resting state, you also see dorsal attention there. And again, that speaks to us about when you're doing a task, it's never a pure cognitive function. I mean, when you're doing a finger tapping test, you're still using attention. And so that sometimes these resting state networks can be very specific to the functionality that you're trying to map. But again, as I mentioned, it takes a lot of work to actually create these maps. So when we first started doing this, again, we would have a technician, we'd have various people working on it, it would be several days, you know, you'd have to do it under a kind of a research protocol, and we really wanted to transition to something that was more efficient, more effective to use. And so we started using more advanced classifier algorithms to automate this process of finding networks. And so what we would do is we would basically, each voxel that would come through would basically be run through, I'm going to talk about this in a little more detail, a kind of a multilayer perceptron. It's kind of a neural network model. It's one of the supervised classifiers that are being used, for instance, for speech recognition. So that it could basically create an automated output for each voxel, which voxel does that network most likely represent of our seven canonical networks. Now the way we did it, we had to train this multilayer perceptron using kind of a priori data. And what we did is we basically scoured the literature and found all the sites that are associated with cortical activations relative to these networks. And that involved, I think, you know, a multitude of publications, but it ended up generating 169 regions of interest. And then we then trained it on 21 normal subjects. So we basically got average cortical activations for these various tasks and these various locations. And then those became the seeds, if you will, for our multilayer perceptron to then create correlation matrices. So that basically we could find out what those structures looked like on average. And then when we run new patients through it, each voxel gets referenced against those correlation matrices. So that not only do you basically find, and you can now kind of allocate the entire structure of the brain to each of these seven canonical networks. And that maps to not just cortex, but also to deeper gray structures in the cerebellum. And you can do it for an individual subject, which is the most important thing. And it's automatic. So we can do it either in a kind of probabilistic model. We can do thresholds where we can say, okay, you know, at a 50%, 70%, 85%, 95% thresholding, what's the likelihood that this is default mode versus motor versus vision, et cetera. Or we can do a winner-take-all where we parcellate the entire brain. Now again, of our first foreign to say, okay, like, you know, let's ground this into some level of reality, was for the somatomotor network, we then said, okay, for the centroid of the somatomotor network, how well does that predict the central sulcus? And it predicted exceedingly well, the p-value of less than, you know, 10 to the minus 15. So it always matched up with the central sulcus across, you know, multiple individuals with a high level of variability. Then we wanted to take it one step further and start comparing it to things we know, such as cortical stimulation. And again, we had a number of patients, both tumor patients and epilepsy patients, and I'll show you some examples of each. And for our tumor patients, again, we had seven. And we could, the first thing we wanted to demonstrate in these clinical patients, can we find all these networks, and what happens with the tumor? Does it kind of ruin it? And as it turns out, it doesn't. And we could find that these networks were really hugging the edges of tumors. So we could really reliably see kind of that these resting state networks when we create, we call them, you know, kind of casually the rainbow brain, where all the networks, you know, superimposed, that we could find that. And sometimes that these networks were actually in the tumor, meaning that the tissue, there was some viable functional tissue still within it. And again, we had some early examples where, okay, well, let's compare this to our awake craniotomies. And we would find that stimulation often aligned with the resting state networks that we mapped, such as motor and speech. But we wanted to be more quantitative about this. And so we then looked in our epilepsy patients where we could have grids. And we could then start to kind of look at, you know, where we would co-register the electrodes to the anatomic anatomy and with their resting state networks to start doing some kind of analysis and see if we could predict stimulation positive sites. And again, the patients would both have resting state MRI before surgery. They'd have their electrodes placed. And then basically we would stimulate through those grids for standard mapping and compare to the clinical standard, both in terms of sensitivity and specificity. And again, just showing you some examples that for motor, for the somatomotor, there was a very tight correlation, both high sensitivity and specificity to kind of where motor cortex was located with stimulation. And also, I would say, a high sensitivity for language and probably a lower specificity. And when we boil this down to receiver-operator curves, that again, for the area under the curve for motor was 0.89, whereas for language it was 0.76. And that's really due to a higher sensitivity, meaning that very often that there was extra areas identified with the resting state network. And very rarely was there a stimulation electrode outside that resting state network. And again, false negatives count in this, meaning that if you're mapping these networks and you basically identify areas that outside of that is not in the network, but it is a stimulation positive electrode, that's a bad thing, versus if you've got some areas within the network where there's a negative electrode that is truly negative. And so, we then mapped out, okay, how far out do you need to go relative to one of these networks depending on various thresholds where you could have a potential false negative. And really, once you're around 15 millimeters outside of these resting state networks, you have a 3% chance of a false negative, which isn't bad. So, we took that and we said, okay, this is starting to look interesting. Now, how can we start to think about a true clinical implementation of this? And, well, one of the first things we did, we had a number of these patients who we started, you know, these brain tumor patients. We ran through the algorithm, both those resting state and they did our standard clinical paradigm for task-based fMRI. And we rated and we did kind of some comparisons. And so, this is across our first 20 patients. We basically first did an imaging score where all of our neuroradiologists rate the subjective imaging. Does it give you excellent topography that would match with your expectations? Is it adequate? Is it borderline? You're not quite sure if that's giving you adequate information or is there a new diagnostic utility. And for imaging score, our resting state did better than for both motor and for language than task-based fMRI. Now, I think also one of the really important things at failure rate that there was a much higher failure rate for task-based fMRI versus resting state where our failure rate was only around, you know, 5 percent for these first 20. And so, that got us really excited about the possibility of then now starting to use this on a more regular basis. So, then it became kind of an effort, an IT effort. And what we did is we really created a kind of IT pipeline so that when a patient got imaged, they got their standard imaging, their anatomic brain tumor protocols in addition to their resting state fMRI. Again, this is just 12 extra minutes in the scanner. And then that data then basically came into kind of our PAC system. There's the anatomic portion that was segmented, both the DTI and the gyroanatomy. And then also we had kind of that analytic perceptron that we created embedded into the IT software. So, it automatically analyzed it. We did a number of quality metrics to make sure that the data was of good quality. Those resting state networks then got turned into DICOMs that then got downloaded into our stealth station. So, basically, again, just kind of a living example is a patient would come in, you know, with a brain tumor, you know, maybe through the emergency room. They would get what we call an advanced stealth. And then basically by the morning, all their kind of, you know, their motor, their speech, and their DTI was all automatically processed and they would be part of our stealth station. So, that you could very quickly get functional information to make some decisions on. Now, oh, I'm sorry. And so, thus far, between, you know, January of 2014 and June of 2015, we've done 235 cases. And now we're really kind of starting to look at, you know, what we're doing with it. And we've actually, you know, to current date, I think we've done close to about 500 of these cases. And, again, the majority have been, in terms of pathology, been for glioma cases. In terms of the majority of these are used for when we're making decisions on open tumor operations, but a significant minority is also for laser interstitial thermal therapy. It's great to have that to help give you, kind of to help guide your trajectory for where you're going to put the probe. And I think, again, just some interesting examples. This is a patient who, again, as you can see, has, and again, this is in stealth, so left is left here, but basically had a large low-grade glioma in his temporal lobe. And actually had a seizure, and he's completely aphasic, and then he got this imaging. We could still find his resting state functional networks, even though he's completely aphasic. And actually, we re-imaged him after he got his speech back, and the maps are all the same. So, again, these are very robust, and so that you can really do some mapping in difficult situations. And here's an example where we've got a thalamic GBM, and I've got to figure out where I'm going to put my laser probe. And again, we map, you know, speech, motor, corticospinal tract, and again, I have some information so I can really know where, what trajectory to do. And Dr. Rogerman already mentioned that this is useful for pediatrics, and again, this is becoming a standard protocol for pediatrics as well. So, you know, there are failures, and we, again, now across our series, we compared our failure rate between, again, resting state and task. Our resting state failure rate, when you take all comers, you know, was 12 percent versus a 38.5 percent with task-based fMRI in these patients with brain tumors. And the failure is usually due to motion, that the patients move in the MRI scanner. And the second most common reason is artifact. It is a little bit more sensitive to artifact than task-based fMRI. And again, in terms of how this is changing our practice, when we look at the number of awake craniotomies that we've had to do over the past couple years, the number of awakes has dropped by around 30 percent. So I think that there's a number of cases where you're like, hmm, you're kind of on the borderline. Should I do it awake? Should I do it asleep? Now, because we really have come to rely on our resting state fMRI. So when I think, you know, as we start to conclude, historically we're all, and again, I think that this is a group of enlightened people, and we're already talking about other things to map. But I think, you know, again, when we think about kind of our field writ large, you know, we talk about motor and speech. And I think that's kind of where the light is, because that's what's governed a lot of, you know, our technology has been really geared towards that. And I think as we think about, you know, kind of, you know, these other areas of the brain and how they can potentially impact function, again, whether it be the default mode, the dorsal attention network, the ventral attention networks, like how we take those into consideration for the patient's functional outcomes, I think we can start to expand what we consider eloquent. So with that, I'll finish, and thank you for hanging in there and listening to me, because I know I'm at the very, very end here. Resting State BOLD data really provides a unique information on the connectivity in the brain. It has advantages over task-based fMRI. It does not depend on patient cooperation. It can be done on anyone that can get an MRI, and multiple functional systems can be scanned in a single imaging sequence. So resting state MRI really can be a powerful tool for preoperative brain mapping. So with that, thank you very much. And I... Thank you.

resting state fMRI

operative planning

brain tumors

functional cortex

brain mapping