Transcript
Thank you for the introduction, so, first of all, I want to thank Brianlab for inviting me to present our work about MRI distortion. So basically, my talk will mainly focus on trigeminal neuralgia. So we all know that it's a well-documented syndrome. The incidence increase with age. Medication remains the therapy of choice. But of course, when failing, you have other treatment options like we have seen already in the previous talk [inaudible 00:00:28.552], and one of them is stereotactic radiosurgery. So I will focus on that. So all technical characteristics, the way that we are performing our trigeminal neuralgia is the fact that we are using different MR sequences. So we are using a CISS, a T1 and T2 weighted images. And then in order to cope for stereotactic coordinates, we are using a stereotactic CT of 1 millimeter and we are applying registration of the images in order that the MR is also in stereotactic conditions. And then we are placing an isocenter of 90 gray at the root entry zone of the nerve. And then we are trying that 50% isodose line is just outside the brain stem. So that's our treatment strategy. And to do so we're using a Linac, so we are using 7 no-co and their arcs with a 4-millimeter collimator in order that we have a spherical dose distribution. So that's a little bit the way that we are treating our patients. But when we speak about accuracy, yeah, we have to see about what is the least accurate process within the whole chain.
And so I group them in three different parts. So we have the accurate targeting, so we have to have a precise imaging in order that you can perfectly delineate your target that you want to treat. Then, of course, you have to have the tight distribution, and then of course, once you have all this, that's on planning, of course, then you have to be sure that the patient is positioned correctly in order that you can perfectly treat the patient and that you can deliver the dose at the right spot. So my main focus will be now all around the precise imaging. So like the previous speaker was already mentioning, MRI, they are distorted. They can be distorted, and you have two different kind of distortions. So you have the machine-specific ones where the static field can be a problem. So that's hardware-induced inhomogeneities. And then you have the gradient fields. So the gradient field is on top of the magnetic field and due to imprecise calibration and inherent non-linearity, you can have some kind of spatial information problems.
And then, of course, we have also patient-specific ones, the patient-specific one. This is a chemical shifts, so that's basically the different resonance frequencies between fat and water. So then you can have a little distortion in your image. And you have the magnetic susceptibility, which is the tendency of a material to magnetize due to the field and that's more the between tissue and air. So those are the kind of distortion that can happen into your image.
And what they can lead to what? So the previous speaker already introduced that perfectly. So basically they can lead to geographical miss. So we have also publications. But I won't take too much time about that because it was nicely presented previously. So that basically we have misses, we can have the geographical miss. So what is the reason? Because the images that you get, it's just a snapshot, it doesn't give you any information about the real anatomy. So if there is a little distortion you don't see it or you are not aware in the images. So how can we cope for it? So we can cope for it at the beginning. So we can do some kind of measurements of our MRI scanners, so we can use phantoms. We have a lot of protocols that showing how often you have to do the recalibration of your MR scanner, or do testing to see if you have a linearity or not. So that's one way to solve it. You can also modify your scanning protocol in order that you can cope for the kind of distortions. And you can correct for then at the time of performing your MR. So you have also industry-based so the the machine itself, machine-specific distortion corrections. You have 2D or 3D that can be can be performed. Of course, I wouldn't stand here if that would be the ideal solution. So basically the problem of those kind of solutions is the fact that it's only measurements on phantom, so you can get rid of the machine-specific distortions, but you cannot get rid of the patient-specific distortions. So the presence of a phantom inside of the patient can alter also the accuracy of your machine-specific distortion. So that would be the idea.
So me, as a physicist, we are doing a lot of patient-specific QA. So depending on the machine, depending on what we want to treat, we are doing specific QA. So we were also thinking why not performing patient-specific imaging QA? And so measuring the patient-specific distortion. So, how can we do for it? So we can start with the CT. We take the CT as a reference because CT, no interior [SP], there is no distortion. The only problem that can happen is in the Z-direction. And there, your accuracy is as accurate as its couch movement. So let's take into consideration that the CT is perfectly and we can use it as a reference. Then I'm diving a little bit into the Elements software. So we have one app that is called the Cranial Distortion Correction. And when we look into that software tool, basically what's happening is that you have an elastic deformation to correct the rigid fusion. So sometimes when you are doing CT-MR registration, you cannot correctly align your MR in your CT. Sometimes or some people are saying, "Oh, the software doesn't work well," but it also may be the reason is because you have a distortion in your MR scanner. The good thing about that tool is the fact that you can have multi modalities. So in that situation we are just looking at CT-MR, but you can also perfectly correct for DTI imaging like Anthony was showing previously. And it creates additional image sets. So we have new images that are perform. So we have an MR that is correct, that's still your baseline, the MR that was performed at the machine site and then you have a corrected one. Of course, the question is when you get two MR sequences, which one is the correct one to use? And so that's the big idea.
So how does it work in practice? So basically, you do first a rigid fusion between the CT and all the kind of images that you want to correct. Once that is performed, is subdividing the image into cubics of 3 by 3 by 3. Once we have that it's looking at the optimal alignment for each of these little patch, and once we find the optimal alignments, you will calculate the deformation fields within those spots, and then afterwards you will correct the images. And then you get a corrected image that you can perfectly look and align with whatever you want. So, like I was already mentioning, yeah, you get a new image. It's software-based. How can we rely on that? Or how can we be sure that it's the correct image and we'll not miss the target?
So we have already one clinical validation. It's a publication in medical dosimetry in 2018. So they did a prospective study in 9 vestibular schwannomas, and they were introducing distortion within the images. And then they wanted to see or to try to understand whether or not the software is working correctly. And the conclusion, basically, was that it is working. And that it can adequately correct brain MR distortion. So there is already one validation. So that's made us also a little bit more feeling like why not go on and move on also to those kind of distortion correction? And that's why we wanted to dive in into our patient population. So we looked about 12 last week, the trigeminal neuralgia patients. So we did a retrospective study. We looked into the MRI scanners. So basically 3 of them were specific machine corrected, and 11 of them not so, depending on our MR that we have. We need to correct machine specifically already or not. So that was the idea. Starting from that, we had the contour target on the CT, we projected it on our MR that was the baseline. So the MR that was coming from our machine-specific MR sequence. And then afterwards, we were running the software and we were trying to correct the MRs of the specific sequences that we were using to treat patient. And once we come out about the corrected images, we, again, projected the CT contour onto that MR. And we looked where we ended up with our contour.
So we did a contour-based analysis. And the contour-based analysis was basically looking at different indices. So we took two out of them. So I will just take then two of them. So we looked at the Dice index which is looking at the volume of your contour, whether or not you have a good agreement, so closer to 1 the better it is. And then the Hausdorff Distance, Hausdorff Distance is basically the distance between the two points of gravity. So how far are we away from those two contours. So we see for those 12 patients that we were in very good agreement except for one. So you see that the Dice index was very low. So, and it was also directly linked to our Hausdorff distance, so we were far away from each other. So when you look into the images, so basically the upper part was the way that we were treating our patients. So with CT-MR and the registration. And underneath first, the MR that was distorted. And so you can see here that there is a little shift. So out of our patients, it was 1 out of 12 that we saw those kind of shifts of 1 millimeter. So it's only 8%, but that shows that even though that you have a machine-specific distortion, even though that you can do machine-specific QA, you see that sometimes because there is also patient-related distortion that that those kind of things can happen. And the part about that kind of images, so we have on the left, you have the MR that was performed at the day of treatment, and on the right the MR that was performed six months after follow-up. We corrected the both of them. So we can read the software. The left image is the one that you saw with the 1-millimeter shift. Six-millimeter patient came in with the follow-up MR. And that was any distortion scene. So basically it's really depending on your sequence, depending on a lot of factors. And those kind of factors, you don't have them by hand, so you don't know what exactly you have to do to correct them.
So I personally think that as we are using, whether it's a Gamma Knife, a Linac-based, or another device we are using the most accurate devices to treat the patients. So I think that the acccuracy need to be maintained, so we have also to look into all the kind of accuracy that we need to maintain. And one of them that is sometimes overseen is the fact or MR sequences. Back when we are speaking about 7-Tesla, you know that those kind of images will be more distorted so we have to have some kind of tools that can correct for them. If you're using the DTI sequences, you have to be sure that what you see is also at the right spot. So I think we have to try to use the Cranial Distortion Elements, so to be sure that we correct all the kind of image sequences in a correct way that we are sure about what's happening anatomically in our patient, and that we can hit the target for follow-up and every kind of other thing, so that we can be sure that what we see is also what is happening within the brain of our patient. And I'm just presenting for trigeminal neuralgia patients. But it's also the case for all the kind of MR sequences that we are using. I thank you for your attention.
And so I group them in three different parts. So we have the accurate targeting, so we have to have a precise imaging in order that you can perfectly delineate your target that you want to treat. Then, of course, you have to have the tight distribution, and then of course, once you have all this, that's on planning, of course, then you have to be sure that the patient is positioned correctly in order that you can perfectly treat the patient and that you can deliver the dose at the right spot. So my main focus will be now all around the precise imaging. So like the previous speaker was already mentioning, MRI, they are distorted. They can be distorted, and you have two different kind of distortions. So you have the machine-specific ones where the static field can be a problem. So that's hardware-induced inhomogeneities. And then you have the gradient fields. So the gradient field is on top of the magnetic field and due to imprecise calibration and inherent non-linearity, you can have some kind of spatial information problems.
And then, of course, we have also patient-specific ones, the patient-specific one. This is a chemical shifts, so that's basically the different resonance frequencies between fat and water. So then you can have a little distortion in your image. And you have the magnetic susceptibility, which is the tendency of a material to magnetize due to the field and that's more the between tissue and air. So those are the kind of distortion that can happen into your image.
And what they can lead to what? So the previous speaker already introduced that perfectly. So basically they can lead to geographical miss. So we have also publications. But I won't take too much time about that because it was nicely presented previously. So that basically we have misses, we can have the geographical miss. So what is the reason? Because the images that you get, it's just a snapshot, it doesn't give you any information about the real anatomy. So if there is a little distortion you don't see it or you are not aware in the images. So how can we cope for it? So we can cope for it at the beginning. So we can do some kind of measurements of our MRI scanners, so we can use phantoms. We have a lot of protocols that showing how often you have to do the recalibration of your MR scanner, or do testing to see if you have a linearity or not. So that's one way to solve it. You can also modify your scanning protocol in order that you can cope for the kind of distortions. And you can correct for then at the time of performing your MR. So you have also industry-based so the the machine itself, machine-specific distortion corrections. You have 2D or 3D that can be can be performed. Of course, I wouldn't stand here if that would be the ideal solution. So basically the problem of those kind of solutions is the fact that it's only measurements on phantom, so you can get rid of the machine-specific distortions, but you cannot get rid of the patient-specific distortions. So the presence of a phantom inside of the patient can alter also the accuracy of your machine-specific distortion. So that would be the idea.
So me, as a physicist, we are doing a lot of patient-specific QA. So depending on the machine, depending on what we want to treat, we are doing specific QA. So we were also thinking why not performing patient-specific imaging QA? And so measuring the patient-specific distortion. So, how can we do for it? So we can start with the CT. We take the CT as a reference because CT, no interior [SP], there is no distortion. The only problem that can happen is in the Z-direction. And there, your accuracy is as accurate as its couch movement. So let's take into consideration that the CT is perfectly and we can use it as a reference. Then I'm diving a little bit into the Elements software. So we have one app that is called the Cranial Distortion Correction. And when we look into that software tool, basically what's happening is that you have an elastic deformation to correct the rigid fusion. So sometimes when you are doing CT-MR registration, you cannot correctly align your MR in your CT. Sometimes or some people are saying, "Oh, the software doesn't work well," but it also may be the reason is because you have a distortion in your MR scanner. The good thing about that tool is the fact that you can have multi modalities. So in that situation we are just looking at CT-MR, but you can also perfectly correct for DTI imaging like Anthony was showing previously. And it creates additional image sets. So we have new images that are perform. So we have an MR that is correct, that's still your baseline, the MR that was performed at the machine site and then you have a corrected one. Of course, the question is when you get two MR sequences, which one is the correct one to use? And so that's the big idea.
So how does it work in practice? So basically, you do first a rigid fusion between the CT and all the kind of images that you want to correct. Once that is performed, is subdividing the image into cubics of 3 by 3 by 3. Once we have that it's looking at the optimal alignment for each of these little patch, and once we find the optimal alignments, you will calculate the deformation fields within those spots, and then afterwards you will correct the images. And then you get a corrected image that you can perfectly look and align with whatever you want. So, like I was already mentioning, yeah, you get a new image. It's software-based. How can we rely on that? Or how can we be sure that it's the correct image and we'll not miss the target?
So we have already one clinical validation. It's a publication in medical dosimetry in 2018. So they did a prospective study in 9 vestibular schwannomas, and they were introducing distortion within the images. And then they wanted to see or to try to understand whether or not the software is working correctly. And the conclusion, basically, was that it is working. And that it can adequately correct brain MR distortion. So there is already one validation. So that's made us also a little bit more feeling like why not go on and move on also to those kind of distortion correction? And that's why we wanted to dive in into our patient population. So we looked about 12 last week, the trigeminal neuralgia patients. So we did a retrospective study. We looked into the MRI scanners. So basically 3 of them were specific machine corrected, and 11 of them not so, depending on our MR that we have. We need to correct machine specifically already or not. So that was the idea. Starting from that, we had the contour target on the CT, we projected it on our MR that was the baseline. So the MR that was coming from our machine-specific MR sequence. And then afterwards, we were running the software and we were trying to correct the MRs of the specific sequences that we were using to treat patient. And once we come out about the corrected images, we, again, projected the CT contour onto that MR. And we looked where we ended up with our contour.
So we did a contour-based analysis. And the contour-based analysis was basically looking at different indices. So we took two out of them. So I will just take then two of them. So we looked at the Dice index which is looking at the volume of your contour, whether or not you have a good agreement, so closer to 1 the better it is. And then the Hausdorff Distance, Hausdorff Distance is basically the distance between the two points of gravity. So how far are we away from those two contours. So we see for those 12 patients that we were in very good agreement except for one. So you see that the Dice index was very low. So, and it was also directly linked to our Hausdorff distance, so we were far away from each other. So when you look into the images, so basically the upper part was the way that we were treating our patients. So with CT-MR and the registration. And underneath first, the MR that was distorted. And so you can see here that there is a little shift. So out of our patients, it was 1 out of 12 that we saw those kind of shifts of 1 millimeter. So it's only 8%, but that shows that even though that you have a machine-specific distortion, even though that you can do machine-specific QA, you see that sometimes because there is also patient-related distortion that that those kind of things can happen. And the part about that kind of images, so we have on the left, you have the MR that was performed at the day of treatment, and on the right the MR that was performed six months after follow-up. We corrected the both of them. So we can read the software. The left image is the one that you saw with the 1-millimeter shift. Six-millimeter patient came in with the follow-up MR. And that was any distortion scene. So basically it's really depending on your sequence, depending on a lot of factors. And those kind of factors, you don't have them by hand, so you don't know what exactly you have to do to correct them.
So I personally think that as we are using, whether it's a Gamma Knife, a Linac-based, or another device we are using the most accurate devices to treat the patients. So I think that the acccuracy need to be maintained, so we have also to look into all the kind of accuracy that we need to maintain. And one of them that is sometimes overseen is the fact or MR sequences. Back when we are speaking about 7-Tesla, you know that those kind of images will be more distorted so we have to have some kind of tools that can correct for them. If you're using the DTI sequences, you have to be sure that what you see is also at the right spot. So I think we have to try to use the Cranial Distortion Elements, so to be sure that we correct all the kind of image sequences in a correct way that we are sure about what's happening anatomically in our patient, and that we can hit the target for follow-up and every kind of other thing, so that we can be sure that what we see is also what is happening within the brain of our patient. And I'm just presenting for trigeminal neuralgia patients. But it's also the case for all the kind of MR sequences that we are using. I thank you for your attention.