Today I'll just be talking a bit, a short history of some of the fiber tracking we've been doing for preoperative mapping and how that's evolved over the last 16 or so years. So this work started with Dr. Berger at UCSF and this was just towards the end of last century which is a funny way to say things. But this is sort of where we're heading now, looking at language pathways and visual pathways and then trying to assess what's been damaged and how they've changed pre and post surgery and then finally how these things might actually help to not only understand the impact of surgery preoperatively but also understanding how to predict long term deficits. This is a, I'll just start off by showing some of the pictures. Diffusion fiber tracking is always good for pretty pictures. Everyone likes that. So this is a gratuitous 3D spinning picture of different language pathways. So just to begin at the beginning very briefly and just reacquaint everyone in case you forgot, the way this algorithm works is really straightforward. We acquire the diffusion MRI data which everyone calls DTI although technically DTI is only a model that's performed on the data but it was the original model that got everything going and you model the data to try and infer the directions of the pathways within a voxel and then you come up with some kind of algorithm to start connecting those arrows in each voxel. So that's essentially how we come up with the fiber tracking and it's not really a talk about fiber tracking but just wanted to give the background in this context. Fiber tracking itself and diffusion modeling itself data and acquisitions have evolved a lot over the last two decades as well and concurrently. This is one of the first patients that we did at UCSF. This was done I think probably 2000, not sure, and we had a lot of cortical stimulation points that we could associate with the fiber tracking using DTI fiber tracking and it was really amazing and I think that first one was better than the following 50 but that's kind of how it goes. So this was followed up as we worked on a technique a bit. Even though the DTI fiber tracking has limitations, it's still the most widely implemented method that's out there, early on we could see things that were quite reasonable with it that made us believe that we were doing something sensible. So for example here you see the fiber tracks that one would get by mapping from a shoulder motor stimulation and wrist motor stimulation which were acquired intraoperatively and then looking at how those fiber tracks, you can see them spiral around the tumor in this case and spiral around each other in the known architecture for these cortis-pallate track pathways into the cerebral peduncle. So just to say what is it we're doing, well with fiber tracking and streamline methods there are a number of other methods but they all do more or less the same thing. You try to use the data to infer the direction of a white matter fiber bundle within a voxel and that could be a tensor model, that's the DTI, the so-called DTI and essentially what it does is it says within each voxel you fit the best ellipsoid that you can and then or you could look at higher order models of which there's one called Q-ball, there's spherical deconvolution, there's many different names for just having more realistic shapes of what goes on in white matter voxels. And then there's the way you propagate and connect all those dots essentially or arrows and using different algorithms. So just as a very simple example, if there's just a single bundle of parallel fibers then the diffusion tensor ellipsoid would give you something shaped like that and then if you do some higher order model it should give you the same shape more or less. But if you have two bundles crossing for example, the diffusion model will give you the best shape to this thing which is basically a pancake or something like that and in which case if you go to try and connect through something like this you can see you're going to obviously make errors while if you do the higher order modeling you can get a more realistic shape. So this is one of the first things that evolved in terms of how we did fiber tracking preoperatively and this was a paper that we published in 2008 and just showing how we could use this technique and so we've been doing this now for probably since before that using this model and this type of data for fiber tracking for intraoperative use. You see now you have two bundles going across each other and when you connect and make the pathways you can actually make two separate bundles and the algorithm won't get confused. So part of what you can do with fiber tracking data in the surgical suite is keeping just as the talk we just had with respect to functional MRI, we also know that there are obvious errors and there are pitfalls and problems with this technique. So we need to really understand how much we can trust the data we acquire and how the techniques might actually be improved and one way to do that is to use the stimulation as was just presented with the functional MRI as a kind of standard in order to tell us if we're doing something sensible and in this case looking at two algorithms, one is a probabilistic tracking which is just a stochastic method instead of a non-stochastic method that is seen has better properties. If you use a DTI model to do that, what you can do is you can make an experiment where essentially you get the cortical stimulation and you say well look, there was a hand motor stimulation in the cortex so we know there must be a subserving pathway. Now we do the fiber tracking and the one question we can answer is, is that fiber track there or not? If it's not there then we know that the technique has failed. So if you do that, what you'll find is that there is a limited sensitivity to that traditional diffusion tensor imaging and that's actually pretty well known from early on. For example, when people made the first beautiful maps with fiber tracking, all you would get, you would mostly just get the kind of lower extremity motor track and many times hand motor or face motor, all of these regions would not be connected and that's because of the complexity in the crossing fibers as you get this SLF crossing the cortal spinal track. And so you can see even with a probabilistic method, it's still, you get maybe a little bit but really you don't get. So with these higher order models like the one I just showed you, we actually started to get more realistic motor pathways and then we could actually build the different motor pathways and using the cordal stimulation, we can actually test the sensitivity of different approaches and I think this still remains a very important approach. This is not a perfect approach because there are many other technical issues when making that comparison but it's certainly the best we can do right now in terms of trying to validate this technique for interoperative use. This is another study we had done where we looked at using the cue ball technique, we looked at the cortal spinal track and we looked at connectivity to the different motor pathways and this was done with white matter stimulation and just to back up for a second, so here for example, there was a stimulation point in the white matter and then what we did was we looked at the closest associated fiber track and why do we do that? The reason we do that is because there are a number of inaccuracies that occur when you try to correlate preoperatively acquired data with the interoperative data. First of all, with respect to the stimulation, there is a spread of that current so there is a distance from which the track can be stimulated that's remote from the actual position of the track. We don't know quite what that distance is, there have been estimates made in the cortex but in general we don't know how tissue geometry, how pathology and all of these other things might affect that penetration depth. Secondly there are of course the shifts that occur when you have the surgical opening and the surgical resection and so all of these things will add up to some kind of offset between where you might stimulate interoperatively and where the preoperative fiber tracks might lie. The new efforts going towards interoperative imaging may actually help to address that. The truth is though the techniques for doing fiber tracking require data that are still not really available in high fidelity on an interoperative machine but what at least the interoperative machine might be able to do is to help us to register better what we acquire offline in a sense with what's going on during surgery in real time. Going backwards then, what we could do is we could actually look at some of those offsets and the other thing we could do is we could also say look, if it was a hand motor pathway that we think was stimulated, does it actually go to the expected location on the cortex? So just at a gross level we could separate into the function that was stimulated and where from a somatopy point of view where it connected and in that case we actually found really good concordance. So this data keeps being really believable to at least to a certain level and so it seems to really work very well as an adjunct to interoperative mapping both subcortically and cortically. When we looked at the tissue shifts, we actually did twice studies looking at the offsets that we observed interoperatively and on two different groups of patients. The first study was published in 2007 and we found an offset of about nine millimeters with the DTI technique. What was really interesting back then was that the variability of that offset was very small which is something that's really more important than the offset itself to some degree because it's actually how well we could actually predict the distance from the preoperative mapping where one might find the stimulation interoperatively. So that actually really helps to zero in on where the surgeon can expect to find the stimulation during surgery. In the following data sets that we analyzed, we looked at both DTI and cue ball tracking and we did find that the cue ball did a better job overall than the DTI in terms of reducing that distance. What we did observe as well was that sometimes when the fiber tracking was not complete, that is to say, let's say there's a lot of edema and you're not getting the full bundle, then the track can appear to be further away from the stimulation than it actually is. But for the cases where we actually feel like we actually did stimulate and we were able to get a complete track, it was always about half a centimeter away with the cue ball technique. I want to just switch topics slightly and evolve into looking away from motor and visual. We've started to look at some language. In the last four or five years, we started looking at some language preoperative mapping. As many of the previous speakers explained, it's a very complex world, but we just took sort of a very pragmatic approach to going out of after this. We had some initial experience that was recently published in General Neurosurgery, but we've been doing this more or less since 2012 in terms of making the preoperative language pathways with high angular resolution diffusion imaging data sets. Our point at this was to try to develop something that was reproducible and feasible and would really give us something sensible over time in terms of pre-surgery planning, in terms of assessing the connectivity for language pathways, and also really importantly to see if we could actually correlate and predict long-term deficits in patients based on the language pathways. For this approach, we defined eight sort of putative language pathways. I know within that community there's a lot of debate about what you call which one and which one exists, but we don't really go into that so much. We sort of took what we believe is sort of a consensus to some degree and tried to map those pathways. Those are the dorsal and ventral pathways, RQET, SLF23, SLFTP, MDLF, ILF, IFOF, and SINET for the ventrals. So in terms of this clinical throughput, we've probably done over 400 patients now in the last four years or so, and we create the maps preoperatively, and then they're exhibited intraoperatively for mapping. So we wanted to look in some initial look at some subjects and see what we could understand about the impact on resection of these tracks, so we took 35 patients that were special because we had gotten also postoperative fiber tracking on them, and so that gave us something that we could look at in a different way. So these patients had fiber tracking just before surgery and within three days post-surgery, and then in terms of the language evaluation, they also had a long-term follow-up, and so this made this data set a very nice one for us to explore, to see initially what's going on with the fiber tracking for language pathways. In order, as a first pass at the technique, we used two operators, and so all of the fiber tracking techniques, most of them, pretty much all of them that are done are really operator-dependent. There have been attempts at automatic methods, but the registration is really a challenge because of all of the distortion that occurs due to the tumor, and the kind of images that are acquired for this type of data is also prone to distortion on top of that. So all of this work is highly expert-trained dependent, and that's an important thing to remember. So you have people at local institutions that learn how to do this and learn very well how it works, and then they really gain a lot of experience, and then they're the ones that can do it, but it's really operator-dependent. You launch these techniques, you use prior anatomical knowledge to really clean up the results that you get and make these fiber tracks. So a lot of time is made just figuring out how to make the fiber tracks in a believable way. So in this case, we used two operators who independently reconstructed the tracks pre- and post-surgically using an agreed-upon algorithm in terms of placement of where you put the constraints on the pathways in order to define them. We also came up with a visual rating scale where we could say there was low impact where the track either seemed unchanged or it was displaced but otherwise normally appearing, and then we also talked about a high impact where it was infiltrated, partially interrupted, or just disconnected completely. The concordance for the operator doing this approach was actually pretty high. So even though it's somewhat subjective, it's an algorithm that can be agreed upon and can be quite reliable. The differences in this particular dataset were resolved by consensus. In terms of the correlation, we had these 35 patients pre- and post-surgery of which 14 were low-grade with clinical deficits, but none before or after surgery. So 21 were high-grades. In terms of the clinical deficits, there were none before or after for the low-grades. That's correct. For the high-grades, we did have three that were pre- and post-surgery early, that is right after, and then eight that were new early post-surgery deficits, but four of them resolved at the later follow-up. Ten had neither pre- nor post-surgical deficits. In terms of the analysis, the way we analyze this data, there are very few that actually had long-term deficits, right? Only four out of the entire group. And in fact, the low-grades had none. And so the way we approached it is we said, look, if there are tracks that are clearly interrupted or affected, but there are no deficits in a patient, then in some ways it says that these tracks are somehow less important for that function to be preserved. At least they're not necessary. And so from that point of view, if you look at the low-grade data, more than 80 percent of the patients had something that were not even though they had no clinical deficits, the tracks were very affected. And so we suspect that it's harder to understand what's going on in low-grade patients because there's probably a lot of opportunity for plasticity. In high-grade patients, the effects are more acute, and there we actually did see something quite interesting, which is the patients who were the tracks that were affected, more likely in the affected patients, clinically affected patients, were the arcuate and the SLFTP out of those eight tracks. So essentially what we found that were these sort of dorsal involvement post-surgery, whether they had short-term or not, they were more likely to show long-term clinical deficits if these tracks were affected either by the tumor or the surgical procedure. And even for the ones that had within the patient subgroup that had early post-surgery language deficit, then it also predicted whether or not they would have long-term deficit. So from our initial experience, it seems like these are the two tracks. This might not be very revolutionary for language mapping, but at least it's something that shows us how these pathways could be used to start looking at this problem. Going forward, there are lots of areas that need to be further developed with respect to diffusion fiber tracking. The models of fiber tracking algorithms, ways of using different interoperative methods as standards for evaluating and validating the methods, and then putting techniques together like fMRI, MEG, that will be talked about later by Srinagarajan, and TMS, and so on. One of the things, though, that we've taken on as a challenge for us is to move away from this subjective analysis of fiber pathways and to try and come up with automatic methods to create these fiber bundles. And this will be important both as an extensible algorithm that can be applied to hundreds of patients, where you don't have to have a single expert that only one person in the world knows how to do it this way, and then it's something that can be also extensible in the sense of being applied across different institutions in an objectively similar way. This is an example of a patient for whom there was a short-term impact on full recovery, and it's a patient like this that we'd like to understand. Here is what the manual subjective approach did. There was a presurgical tract, and then there's the post-surgical IFOF, and you can see visually how the tract was changed pre- and post-surgery. So that's a pretty clear case how that was done. So here is a case where it looks like the post-surgery tract is actually better than the presurgical tract. And here you could erroneously think that perhaps the tract is actually somehow affected or destroyed, but it was just once you had the decompression from the tumor, it's possible then you were better able to actually create this fiber tract. Here's what a trained expert operator reliability looks like. This is a voxel-wise map, and it shows you that even though some core parts of pathways can be reliably made across operators, there's still many parts of the tracts that are really not reliably made between operators. There are many other things that contribute, but operator dependence is definitely an important aspect. In terms of one of the approaches that we are exploring right now for doing automatic methods, what we have here is using some advanced methods for registration of fiber bundles, we're able to identify, for example, which regions were resected from pre- to post-surgery. And once you do that, you can actually say which fiber pathways were actually intersecting the resected voxels. And that's what we're showing here. The red is actually those resected voxels, and these are showing the pathways, which in this case were some sensory pathways and SLFTP pathways that were affected. So this is a sort of a method that can be used. The other thing you want to know is how things change, and I'll just quickly end with that. Here is a change that might occur pre- to post-surgery, and here you see the pre-surgical pathways, and this is a post-surgical. And when you look at it visually, it's sometimes hard to appreciate what might have been different, but if you do this automatic algorithm method, you actually pull out the arcuate very nicely. So even though this is the part where the resection was in this case, this was connected to an entire part of the arcuate that was actually affected by the resection. So this is the methods that we want to use going forward. Here I just left in this little piece of thing here, which is a remote registration error by the tremendous distortion that happens, just to remind us that we still have a little ways to go to make these techniques perfect for perioperative and intraoperative mapping. Thank you.

fiber tracking

preoperative mapping

diffusion MRI data

algorithms

motor pathways

language pathways