- 284
Transcript
Also, Brainlab asked me to give a presentation of what we try to do to advance the workflows that we're currently using. And I'm a heavy user of the Elements system as well now that my disclosures are seen down there. And I'm talking about those advanced imaging technologies. Now we're coming out of a world where we were basically treating nuclei and nodes of networks. And there was a nice discussion on that in one of the symposia on Monday. And [inaudible 00:00:34.959] has actually laid out that actually we've been always been thinking about nodes, and networks, and all that. But now we're in the realm of where we are able to show the networks and to use connectivity determination with MRI, and that's basically the main part of what I'm going to talk about.
Just as an introduction, you see on the left side there's a spherical rendition of the likelihood of diffusion if you do not restrict diffusion into a certain direction. On the right side, so that would be like a free diffusion into all directions. Then if you measure that in a brain with an MRI scanner, you cannot really appreciate direction from that. However, if you're in a situation where you are working in larger fiber bundles, for example, then there will be a preferred diffusion of the protons along the external structures and a little bit outside also of the external structures. And that is a preferential diffusion and you're able actually to measure that with an MRI scanner and then to draw conclusions after that.
How that basically works, and just to elude on how complicated that might be, I'm just showing what you can do. You need a certain gradient which you can see in the upper right here. That means you need to scan the brain in a certain direction. You need to know the geometry and the arc settings of that. Of course, that results in a very, very complicated, what's called a gradient table in order to evaluate later where you have been scanning what you have been doing. And then you need those almost quasi-anatomical scans that you intercalate in all those scannings here in-between because the system understands, or basically, we know and make the system understand that, from the different scanning directions and from how you scan that, there is going to be a lot of inaccuracies introduced in that. However, that is called diffusion-weighted imaging. We typically do that with 61 directions, and then we have 8, so that ends up in 69 directions because we have 8 intercalated B0 sequences in there that looks quasi like a T2 sequence.
And then you have to make your decision, what you do. You can use local approaches to evaluate those images. You can use global approaches to evaluate that. And I'm not going to go too much into detail. There is a lot of MR physics and post-processing with that. What's typically used in today's planning system, that is actually what we use also is the deterministic approach. The deterministic approach is typically realized by FACT algorithm or HEFT algorithm. FACT is fiber assignment to continuous tracking, which basically is the brute-force algorithm that calculate and estimates in a given box or in a given volume, it tries to understand what the main direction of diffusion would be. And those deterministic approaches in most of the softwares that are used today are typically one-tensor models. And that is actually one of the restrictions because at a certain junction of a crossing, a sprouting fiber, and I show you there's a lot of crossing and sprouting fibers and kissing fibers, that algorithm has to make a decision where it's going and where it is going to track. And that is actually also the main limitation. While that is a very robust on the test-retest relation, it's a very robust algorithm, that potentially does not show you everything. That's a very important limitation that you need to know.
Now, what you can do then is that's basically another rendition of what you've seen before. You can do nice depictions of the pyramidal tract, for example, and other larger fiber tracts. But what you have to know and what you have to obey basically there is you're having a mathematical algorithm appreciate anatomy for you. And anatomy, that is probably not anatomy. So you need to really understand the anatomy beforehand before you interpret what you do. And I'm not showing here the paper by Kinoshita and co-workers who basically have in 2005, I guess it was probably the first paper on glioma surgery in the central region where they basically purely ignored electrophysiology telling them you're approaching the sample region and just relied on a diffusion term by imaging depiction, very thin depiction of the pyramidal tract and left one of those two patients, [inaudible 00:04:57.719], but it is due to the editor's credit that this was published actually and which was one of the great examples of how not to do things.
Now we're not the first ones to talk about connectivity. We're not the first ones certainly to talk about how the thalamus looks like. This is early work by Heidi Johansen-Berg from Great Britain, that's already 2005 or 2003 and they already did segmentation. And also, they already saw what we today use a lot for tremor, that is the dentato-rubro-thalamic tract or the cerebellothalamic tract that can be nicely targeted to additionally identify the ventral intermediate nucleus, or the subthalamic, posterior subthalamic region for tremor surgery.
Clinically, I'm talking about tremor and depression. And since there is another talk going to be on tremor surgery emissions, I'm just going to briefly go over that and leave that open and then go into what we nowadays do for depression surgery in order to...what we think makes it work better. So that's the dentato-rubro-thalamic tract that comes from the cerebellar nuclei then crosses over...most of the fibers cross over under the red nucleus, they have a little touch with the red nucleus. There is one loop that goes down to the spinal cord there. And then we have that other loop that goes into the ventral intermediate nucleus. Ventral intermediate nucleus, there's a synapse there. DTI cannot show the synapse, it only shows the fibers, it does not show the directions. But in that first approach that we published in 2011, we were able to show the coincidence of an electrode position. There is very, very limited current. In this case, it was 1.5 milliamp checking out to something like...sorry, 1.5 volts checking out to something like 1 to 2 milliamp. Could be very, very beneficial and that basically was the first presentation of the DRT as a potential target for deep brain stimulation. Now, DRT DBS looks like this. You want to be in that anterior half of that dentato-rubro-thalamic tract, not too much encroaching on fibers that are going to the back because you will intermingle if you're too far to the back, and do that without testing. Intraoperatively, you would intermingle with fibers that already belong to either the lemniscal system leading to persistant paresthesia or you have fibers from the prelemniscal radiation that then leaves the patient unstable in gait and may be falling. We do that on a regular basis. That's basically images from the Elements system. So you can see how with the one-tensor model of the Elements system you can nicely show that. And you can do more advanced stuff which we have published here. There was a patient that was operated outside actually in a very good ventral intermediate nucleus location, but he had a very proximal tremor of a "yes-yes" type. And there was a tremor that was basically of the trunk. And we decided to do something else and explanted the first system, did a DTI scan, and then implanted the second system. And implanted that actually significantly deeper with DRT guidance and actually got a fourth patient a much, much better result. More for tremor in the next talk though.
Talking about depression, which is actually the topic that I would like to show in order to show what we can do in advanced imaging. There are some targets that have been used to treat depression. There's no real consensus about what you should use to treat depression. We have, and Tom Schlaepfer is in the room, who's the psychiatrist that I'm happy to work with and collaborate with. We have...to 10 years ago decided to do medial forebrain bundle stimulation in order to treat those patients with the idea that most of the depressed patients suffer from anhedonia and hopelessness which is looking toward a very important deficiency of the reward system. Medial forebrain bundle system itself has to do with the reward system, that's why we're stimulating that. Interesting is that we have two sadly stopped trials on depression. You all know that, the BCBS trial, the Reclaim trial by Medtronic was stopped in 2010, the Broaden trial was stopped in 2012, published in 2017. And they were stopped basically based on futility analysis. There's the stimulated patients and the non-stimulated patients, that's 90 patients overall up to the futility analysis, that's something that you define before. Before you actually enter the trial yourself, you define what would be the criteria to then pursue the trial or to stop it. And what basically happened here is there was a sham condition and a stimulated condition and that's what the improvement looked like. There was some improvement there, but the improvement was not enough. It was not to be differentiated between responders and non-responders. And one, that's the response line down there. So that group in the time that we're looking at, which is 12 months to 1 year, there was no response to be seen in that whole group. Then the authors basically conclude and say, "Hey, that would probably have worked better if you would have used that tractographic imaging because we now know more about the Cg25 network and the region." So we were now talking about networks. And then there's the beautiful work by Patricio Riva-Posse and Cameron McIntyre from the Emory group. And basically, they're now looking at a conference, like a nexus of a couple of fiber tracts that come together to stimulate them. And we totally agree with them, that is the way to do it.
For the medial forebrain bundle that is work that stem from 10 years ago. That was the time when I was working in Vancouver with Chris Honey as a fellow at that time. And what basically happened is that we had a patient. So we were looking for something completely different. We were looking for the accuracy of DTI. And we had a patient that reacted with a hypomanic response, STN patient. And the hypomanic response, we attributed that to a co-stimulation of fibers that go into the medial forebrain bundle, something that was at that moment in time not really described in the human. So we had to first describe the medial forebrain bundle, at the same time describe that side effect. That worked, however, and was published. And the idea at that moment in time basically came up. If you push somebody out of euthymia, so like a normal balanced state of mood, if you push him over into mania or hypomania and then mania, what that mean that the depressed patient had a deficiency and is reacting to that system and you could push him actually from being depressed over back to his euphoric state, maybe that would be detrimental in the case of Parkinson's patients but probably very good for depressed patients. We did some analysis on that. We did analysis of where are the other targets? Actually look at the other targeted are, in our opinion, located just around the fiber tract as we have described it. So if you see that goes up to Cg25 up here, and then also the anterior thalamic internal capsule. And also and [inaudible 00:12:00.272] goes here, the anterior thalamic [inaudible 00:12:03.440] posterior terminals also have to do with the medial forebrain bundle system. Not going to bore you with that. But the most important thing is that if you stimulate in that region you want to be right in front of the red nucleus. And it's large in circumference. That is a tractographic targeting. You want to have in there the mesolimbic fibers that go to subcortical reward-associated regions like nucleus accumbens and septal region. But you always also want to have the mesocortical part that goes up to Brodmann area 9, 10, 11, 11m, and a little bit of 47.
We did some heavy reconstruction of that region, which is, this is a little bit of workflow. We don't have to understand that in detail what we're using for that is something that you cannot today use in your navigation systems. This is called global tracking. The global tracking have the advantage of better resolving the crossing/kissing branching problem because it is a heuristic that basically looks from the outside of the brain to the inside. It's not a local thing, it's a global thing. And what it does basically, it interprets the MR signal that you scan also in a diffusion where the scan that I just showed right at the beginning, and it interprets it as a polymer. So the brain is interpreted as a polymer and that has basically, one, a lot of context and the global tracking appears to be, at least from the tracking algorithm if you want to do tractography, one of the best tracking algorithms that you can use.
Now the deep medial forebrain bundle is a complex structure. Just showing that. And I've talked about where the medial forebrain bundle is actually projecting. There's a large projection through the Brodmann area 10 here. And you can dissect that out. And in those three postulations of the Desikan/Killiany atlas that we use to parcellate, just a parcellation default that we use, 97 of those fibers go in the region that I've just described. And you can, with a deterministic approach, you can do that. You can nicely show the target region. We've shown that in a paper last year. And that's a comparison with stuff that comes from Balint Varkuti who was at that time still working with Brainlab and that on the boot [SP] lab you can also do that with deterministic tractography, reliably show the fiber tract.
However, the fiber tract itself is a complicated thing to track. For the sake of time. I'm going over the surgical procedure just showing the results. So that seems to work. That's 50 weeks, 49 weeks follow-up of the first FORESEE cohort. So one who is a responder stays a responder. So just trust me that that appears to work in that small group of patients that we're looking at.
Now the first thing that I wanted to show as something new, a couple of slides, is what we do with machine learning. We realize that the medial forebrain bundle is a very complex structure and even if you talk to colleagues who know a lot about the anatomy, it is very hard to talk to each other about where am I going, which part is important, which part is not important. And the complexity is really a problem also for the tracking procedure. Medial forebrain bundle itself as we know it now, the system itself comes from the ventral tegmental area. It has something that is the inferomedial branch that is more the rendition of what the medial forebrain bundle in classical rodent and cat anatomy looks like. And then you have those extensions to the orbitofrontal prefrontal cortex and dorsolateral prefrontal cortex. That's what we classically call the superior lateral branch. But you have some motor extensions also there. So there's a lot of dopamine innovation of the motor cortex also going through the cerebellum's fastigial nucleus and stuff. And you don't want to have those fibers because if you have those fibers in your tracking procedure, that will alter what you do with the tractography planning of that.
That is the main reason why we started to work with machine learning algorithm. And Marco Reisart from my department program, what is called the hierarchical HArMonic filters for LEarning Tract from diffusion MRI. That is again basically something that works on the raw diffusion MRI data. And then it takes into account the appreciation of what we think is the medial forebrain. I have to make you aware, this is not something that... So the machine learning algorithm, this is something for retest reliability, right? So if you feed in something wrong here, right, it always shows you something wrong down here. But we pretty much think after those 10 years that we now know how the structure looks like. We have simulated a couple of times so now we think this is what typically the medial forebrain bundle looks like. And what the algorithm then does is it gives you a comparable tracking corridor. So the colored images that align below the streamlines here that I track are basically the...they are the color corridor that help you to, while you're doing the planning, appreciate actually how the medial forebrain bundle probably should look like.
And there is an example in here that is from a publication on that topic. We draw a rather large blackened eyebrow on top of the red nucleus here, like a region of interest there. We do first tracking that's unrefined. And then based on where the algorithm tells us where fibers should be, we set our exclusion...oh, sorry, that's another inclusion region of interest. And that basically shapes a little bit thicker streamline down here to something that's a little bit thinner. And then this case, you basically can see that makes the differentiation between still reachable and not reachable anymore. So that electrode position would probably be acceptable in that planning but certainly not if you refine that for what you think medial forebrain bundle is. You can do that, you can then exercise that through the whole targeting, use that, and then do your targeting there. We have done basically some preliminary evaluation of that. And this is one of the cases that we did. There's a couple of cases just mentioned in our paper. And in that case, here we had an inferior outcome. It's still an okay outcome, but an inferior outcome than the others. And we could see when you then track from the HAMLET, you see that this electrode is actually located outside the bundle, which probably shows that HAMLET helps us to do something.
Interesting part about that is that's not at all interfering with your deterministic tracking procedure which is shown in gray here. And that is basically where you start to overlay, that's the HAMLET's training and automatic detection. And Dr. Ottoman has given me signs to get ready. And I'm just showing this as my last slide, basically. You could use basically Quentry data baseline, or sorry, Quentry data transfer between institutions. So everything that is blue could basically be located outside like an expert system. So if we would, at a certain moment in time, all do medial forebrain bundle stimulation, which I hope we'll do in the not too far future, there's going to be certain institutions who can do like HAMLET and machine learning renditions and you could send us images. And we can do the machine learning algorithm manipulation of that image, and then send them basically...like so you can use them for your tracking as kind of an expert system.
I have one other topic but for the sake of time, I'm going to stop here and would like to thank you for your attention. Thank you.
Just as an introduction, you see on the left side there's a spherical rendition of the likelihood of diffusion if you do not restrict diffusion into a certain direction. On the right side, so that would be like a free diffusion into all directions. Then if you measure that in a brain with an MRI scanner, you cannot really appreciate direction from that. However, if you're in a situation where you are working in larger fiber bundles, for example, then there will be a preferred diffusion of the protons along the external structures and a little bit outside also of the external structures. And that is a preferential diffusion and you're able actually to measure that with an MRI scanner and then to draw conclusions after that.
How that basically works, and just to elude on how complicated that might be, I'm just showing what you can do. You need a certain gradient which you can see in the upper right here. That means you need to scan the brain in a certain direction. You need to know the geometry and the arc settings of that. Of course, that results in a very, very complicated, what's called a gradient table in order to evaluate later where you have been scanning what you have been doing. And then you need those almost quasi-anatomical scans that you intercalate in all those scannings here in-between because the system understands, or basically, we know and make the system understand that, from the different scanning directions and from how you scan that, there is going to be a lot of inaccuracies introduced in that. However, that is called diffusion-weighted imaging. We typically do that with 61 directions, and then we have 8, so that ends up in 69 directions because we have 8 intercalated B0 sequences in there that looks quasi like a T2 sequence.
And then you have to make your decision, what you do. You can use local approaches to evaluate those images. You can use global approaches to evaluate that. And I'm not going to go too much into detail. There is a lot of MR physics and post-processing with that. What's typically used in today's planning system, that is actually what we use also is the deterministic approach. The deterministic approach is typically realized by FACT algorithm or HEFT algorithm. FACT is fiber assignment to continuous tracking, which basically is the brute-force algorithm that calculate and estimates in a given box or in a given volume, it tries to understand what the main direction of diffusion would be. And those deterministic approaches in most of the softwares that are used today are typically one-tensor models. And that is actually one of the restrictions because at a certain junction of a crossing, a sprouting fiber, and I show you there's a lot of crossing and sprouting fibers and kissing fibers, that algorithm has to make a decision where it's going and where it is going to track. And that is actually also the main limitation. While that is a very robust on the test-retest relation, it's a very robust algorithm, that potentially does not show you everything. That's a very important limitation that you need to know.
Now, what you can do then is that's basically another rendition of what you've seen before. You can do nice depictions of the pyramidal tract, for example, and other larger fiber tracts. But what you have to know and what you have to obey basically there is you're having a mathematical algorithm appreciate anatomy for you. And anatomy, that is probably not anatomy. So you need to really understand the anatomy beforehand before you interpret what you do. And I'm not showing here the paper by Kinoshita and co-workers who basically have in 2005, I guess it was probably the first paper on glioma surgery in the central region where they basically purely ignored electrophysiology telling them you're approaching the sample region and just relied on a diffusion term by imaging depiction, very thin depiction of the pyramidal tract and left one of those two patients, [inaudible 00:04:57.719], but it is due to the editor's credit that this was published actually and which was one of the great examples of how not to do things.
Now we're not the first ones to talk about connectivity. We're not the first ones certainly to talk about how the thalamus looks like. This is early work by Heidi Johansen-Berg from Great Britain, that's already 2005 or 2003 and they already did segmentation. And also, they already saw what we today use a lot for tremor, that is the dentato-rubro-thalamic tract or the cerebellothalamic tract that can be nicely targeted to additionally identify the ventral intermediate nucleus, or the subthalamic, posterior subthalamic region for tremor surgery.
Clinically, I'm talking about tremor and depression. And since there is another talk going to be on tremor surgery emissions, I'm just going to briefly go over that and leave that open and then go into what we nowadays do for depression surgery in order to...what we think makes it work better. So that's the dentato-rubro-thalamic tract that comes from the cerebellar nuclei then crosses over...most of the fibers cross over under the red nucleus, they have a little touch with the red nucleus. There is one loop that goes down to the spinal cord there. And then we have that other loop that goes into the ventral intermediate nucleus. Ventral intermediate nucleus, there's a synapse there. DTI cannot show the synapse, it only shows the fibers, it does not show the directions. But in that first approach that we published in 2011, we were able to show the coincidence of an electrode position. There is very, very limited current. In this case, it was 1.5 milliamp checking out to something like...sorry, 1.5 volts checking out to something like 1 to 2 milliamp. Could be very, very beneficial and that basically was the first presentation of the DRT as a potential target for deep brain stimulation. Now, DRT DBS looks like this. You want to be in that anterior half of that dentato-rubro-thalamic tract, not too much encroaching on fibers that are going to the back because you will intermingle if you're too far to the back, and do that without testing. Intraoperatively, you would intermingle with fibers that already belong to either the lemniscal system leading to persistant paresthesia or you have fibers from the prelemniscal radiation that then leaves the patient unstable in gait and may be falling. We do that on a regular basis. That's basically images from the Elements system. So you can see how with the one-tensor model of the Elements system you can nicely show that. And you can do more advanced stuff which we have published here. There was a patient that was operated outside actually in a very good ventral intermediate nucleus location, but he had a very proximal tremor of a "yes-yes" type. And there was a tremor that was basically of the trunk. And we decided to do something else and explanted the first system, did a DTI scan, and then implanted the second system. And implanted that actually significantly deeper with DRT guidance and actually got a fourth patient a much, much better result. More for tremor in the next talk though.
Talking about depression, which is actually the topic that I would like to show in order to show what we can do in advanced imaging. There are some targets that have been used to treat depression. There's no real consensus about what you should use to treat depression. We have, and Tom Schlaepfer is in the room, who's the psychiatrist that I'm happy to work with and collaborate with. We have...to 10 years ago decided to do medial forebrain bundle stimulation in order to treat those patients with the idea that most of the depressed patients suffer from anhedonia and hopelessness which is looking toward a very important deficiency of the reward system. Medial forebrain bundle system itself has to do with the reward system, that's why we're stimulating that. Interesting is that we have two sadly stopped trials on depression. You all know that, the BCBS trial, the Reclaim trial by Medtronic was stopped in 2010, the Broaden trial was stopped in 2012, published in 2017. And they were stopped basically based on futility analysis. There's the stimulated patients and the non-stimulated patients, that's 90 patients overall up to the futility analysis, that's something that you define before. Before you actually enter the trial yourself, you define what would be the criteria to then pursue the trial or to stop it. And what basically happened here is there was a sham condition and a stimulated condition and that's what the improvement looked like. There was some improvement there, but the improvement was not enough. It was not to be differentiated between responders and non-responders. And one, that's the response line down there. So that group in the time that we're looking at, which is 12 months to 1 year, there was no response to be seen in that whole group. Then the authors basically conclude and say, "Hey, that would probably have worked better if you would have used that tractographic imaging because we now know more about the Cg25 network and the region." So we were now talking about networks. And then there's the beautiful work by Patricio Riva-Posse and Cameron McIntyre from the Emory group. And basically, they're now looking at a conference, like a nexus of a couple of fiber tracts that come together to stimulate them. And we totally agree with them, that is the way to do it.
For the medial forebrain bundle that is work that stem from 10 years ago. That was the time when I was working in Vancouver with Chris Honey as a fellow at that time. And what basically happened is that we had a patient. So we were looking for something completely different. We were looking for the accuracy of DTI. And we had a patient that reacted with a hypomanic response, STN patient. And the hypomanic response, we attributed that to a co-stimulation of fibers that go into the medial forebrain bundle, something that was at that moment in time not really described in the human. So we had to first describe the medial forebrain bundle, at the same time describe that side effect. That worked, however, and was published. And the idea at that moment in time basically came up. If you push somebody out of euthymia, so like a normal balanced state of mood, if you push him over into mania or hypomania and then mania, what that mean that the depressed patient had a deficiency and is reacting to that system and you could push him actually from being depressed over back to his euphoric state, maybe that would be detrimental in the case of Parkinson's patients but probably very good for depressed patients. We did some analysis on that. We did analysis of where are the other targets? Actually look at the other targeted are, in our opinion, located just around the fiber tract as we have described it. So if you see that goes up to Cg25 up here, and then also the anterior thalamic internal capsule. And also and [inaudible 00:12:00.272] goes here, the anterior thalamic [inaudible 00:12:03.440] posterior terminals also have to do with the medial forebrain bundle system. Not going to bore you with that. But the most important thing is that if you stimulate in that region you want to be right in front of the red nucleus. And it's large in circumference. That is a tractographic targeting. You want to have in there the mesolimbic fibers that go to subcortical reward-associated regions like nucleus accumbens and septal region. But you always also want to have the mesocortical part that goes up to Brodmann area 9, 10, 11, 11m, and a little bit of 47.
We did some heavy reconstruction of that region, which is, this is a little bit of workflow. We don't have to understand that in detail what we're using for that is something that you cannot today use in your navigation systems. This is called global tracking. The global tracking have the advantage of better resolving the crossing/kissing branching problem because it is a heuristic that basically looks from the outside of the brain to the inside. It's not a local thing, it's a global thing. And what it does basically, it interprets the MR signal that you scan also in a diffusion where the scan that I just showed right at the beginning, and it interprets it as a polymer. So the brain is interpreted as a polymer and that has basically, one, a lot of context and the global tracking appears to be, at least from the tracking algorithm if you want to do tractography, one of the best tracking algorithms that you can use.
Now the deep medial forebrain bundle is a complex structure. Just showing that. And I've talked about where the medial forebrain bundle is actually projecting. There's a large projection through the Brodmann area 10 here. And you can dissect that out. And in those three postulations of the Desikan/Killiany atlas that we use to parcellate, just a parcellation default that we use, 97 of those fibers go in the region that I've just described. And you can, with a deterministic approach, you can do that. You can nicely show the target region. We've shown that in a paper last year. And that's a comparison with stuff that comes from Balint Varkuti who was at that time still working with Brainlab and that on the boot [SP] lab you can also do that with deterministic tractography, reliably show the fiber tract.
However, the fiber tract itself is a complicated thing to track. For the sake of time. I'm going over the surgical procedure just showing the results. So that seems to work. That's 50 weeks, 49 weeks follow-up of the first FORESEE cohort. So one who is a responder stays a responder. So just trust me that that appears to work in that small group of patients that we're looking at.
Now the first thing that I wanted to show as something new, a couple of slides, is what we do with machine learning. We realize that the medial forebrain bundle is a very complex structure and even if you talk to colleagues who know a lot about the anatomy, it is very hard to talk to each other about where am I going, which part is important, which part is not important. And the complexity is really a problem also for the tracking procedure. Medial forebrain bundle itself as we know it now, the system itself comes from the ventral tegmental area. It has something that is the inferomedial branch that is more the rendition of what the medial forebrain bundle in classical rodent and cat anatomy looks like. And then you have those extensions to the orbitofrontal prefrontal cortex and dorsolateral prefrontal cortex. That's what we classically call the superior lateral branch. But you have some motor extensions also there. So there's a lot of dopamine innovation of the motor cortex also going through the cerebellum's fastigial nucleus and stuff. And you don't want to have those fibers because if you have those fibers in your tracking procedure, that will alter what you do with the tractography planning of that.
That is the main reason why we started to work with machine learning algorithm. And Marco Reisart from my department program, what is called the hierarchical HArMonic filters for LEarning Tract from diffusion MRI. That is again basically something that works on the raw diffusion MRI data. And then it takes into account the appreciation of what we think is the medial forebrain. I have to make you aware, this is not something that... So the machine learning algorithm, this is something for retest reliability, right? So if you feed in something wrong here, right, it always shows you something wrong down here. But we pretty much think after those 10 years that we now know how the structure looks like. We have simulated a couple of times so now we think this is what typically the medial forebrain bundle looks like. And what the algorithm then does is it gives you a comparable tracking corridor. So the colored images that align below the streamlines here that I track are basically the...they are the color corridor that help you to, while you're doing the planning, appreciate actually how the medial forebrain bundle probably should look like.
And there is an example in here that is from a publication on that topic. We draw a rather large blackened eyebrow on top of the red nucleus here, like a region of interest there. We do first tracking that's unrefined. And then based on where the algorithm tells us where fibers should be, we set our exclusion...oh, sorry, that's another inclusion region of interest. And that basically shapes a little bit thicker streamline down here to something that's a little bit thinner. And then this case, you basically can see that makes the differentiation between still reachable and not reachable anymore. So that electrode position would probably be acceptable in that planning but certainly not if you refine that for what you think medial forebrain bundle is. You can do that, you can then exercise that through the whole targeting, use that, and then do your targeting there. We have done basically some preliminary evaluation of that. And this is one of the cases that we did. There's a couple of cases just mentioned in our paper. And in that case, here we had an inferior outcome. It's still an okay outcome, but an inferior outcome than the others. And we could see when you then track from the HAMLET, you see that this electrode is actually located outside the bundle, which probably shows that HAMLET helps us to do something.
Interesting part about that is that's not at all interfering with your deterministic tracking procedure which is shown in gray here. And that is basically where you start to overlay, that's the HAMLET's training and automatic detection. And Dr. Ottoman has given me signs to get ready. And I'm just showing this as my last slide, basically. You could use basically Quentry data baseline, or sorry, Quentry data transfer between institutions. So everything that is blue could basically be located outside like an expert system. So if we would, at a certain moment in time, all do medial forebrain bundle stimulation, which I hope we'll do in the not too far future, there's going to be certain institutions who can do like HAMLET and machine learning renditions and you could send us images. And we can do the machine learning algorithm manipulation of that image, and then send them basically...like so you can use them for your tracking as kind of an expert system.
I have one other topic but for the sake of time, I'm going to stop here and would like to thank you for your attention. Thank you.