---
title: 'Emerging Artificial Intelligence Technologies for Neurovascular Disease Diagnosis and Treatment Guide'
type: 'lesson'
url: 'https://doc.social/courses/neurosurgery-and-artificial-intelligence/emerging-artificial-intelligence-technologies-for-neurovascular-disease-diagnosis-and-trea'
course: 'Neurosurgery and Artificial Intelligence'
course_url: 'https://doc.social/courses/neurosurgery-and-artificial-intelligence'
position: '1 of 3'
author: 'John Bennett MD'
transcript: 'yes'
---

# Emerging Artificial Intelligence Technologies for Neurovascular Disease Diagnosis and Treatment Guide

## Summary

_This lesson is a recording without a written summary._

## Transcript

9,227 words · 69 min · 4 speakers · automatically transcribed (aws-transcribe) and not checked by a person; it may contain errors, particularly in names, drugs and dosages. The recording is the source.

[0:09] Speaker 1: One. Good morning. This is Doctor John Bennett broadcasting from Miami Beach, home of Neurosurgical TV. I'm armed with the uh presence of uh Cyprian Ionita. He's a PhD from University of Buffalo. I initially had an, I had a webcast with the Cyprian a few years back with um 3D printing and neurosurgery and learned that they're closely associated with the Jacobs Medical School and um
[0:40] Speaker 1: The University of Buffalo and Ciprian, I'm sure we'll talk about that. And uh we'd like to welcome Cyprian and thank you, it's all yours.
[0:48] Speaker 2: Thank you very much. Thank you for the introduction. Uh, I will, uh, present today a, um, uh, a first lecture of a series of three into Artificial intelligence. And, um, the, uh, uh, full disclosure, the entire work that I'm presenting today, it is actually the work of my, uh, PhD students. Um,
[1:13] Speaker 2: And um uh I'll uh uh present it um uh in the, in the, in this lectures. So, um, my idea is to do a kind of like a very uh high level uh kind of a presentation on the AI with direct, followed by direct applications to neurosurgery and uh um
[1:39] Speaker 2: The lectures are organized in such a way that we go from things that are easily acceptable in uh in the clinics and in the neurosurgery.
[1:49] Speaker 2: And then increasingly growing into complexity and impact. This lecture today on the AI will be mostly focused on the fact that, listen, this is the message is look, AI can replicate.
[2:10] Speaker 2: What humans do in the uh diagnosis of stroke, as well as a highly trained neurosurgeon.
[2:19] Speaker 2: And the only and the advantage that is through this application is that it does it.
[2:27] Speaker 2: In about 6 milliseconds to about 10 milliseconds, let's say per case, it can go through an entire CT scan and right away the results are are presented to you. So in that context, um.
[2:43] Speaker 2: They will focus on stroke and the uh kind of like uh uh um severity and identification of the patients.
[2:52] Speaker 2: And, uh, and disclosure, full disclosure, uh, this work has been funded by different um federal institutions like NIH and National Science Foundation.
[3:06] Speaker 2: And we also have, uh, we're working towards a small startup, CAS AI at the University of Buffalo. So really quickly, uh, we've seen the AI and the machine learning, all the buzzwords are everywhere nowadays. We heard it from self-driving car, facial and speech recognition and general and game playing, but I don't.
[3:34] Speaker 2: I think in the medical realm has been uh uh yet such, such a, a big uh um
[3:44] Speaker 2: A turnaround or such a big acceptance, mostly because of the really deep implications that will have on personal level or on patient health and also regulatory kind of organizations where we need to make sure that what we're doing it is right so
[4:09] Speaker 2: What can we do like in the medical applications and we've now we've seen it over and over again. Radiology Society of North America has an entire dedicated an entire journal where the AI is used for segmentation, really identify regions and segment them out, which means isolate them and extract features.
[4:34] Speaker 2: Uh, why is that? Well, because it is mostly because of, uh, time savings. Uh, segmentation for a radiologist can take minutes, dozens of minutes, up to hours if, uh, if they have like a, a very difficult uh uh um uh structure in their images.
[4:54] Speaker 2: So Other things is that it can look at different patterns into anything like a tumor or Other or hemorrhages and so forth and then based on existing data sets it can actually a
[5:14] Speaker 2: Identify disease severity and can come up even with device suggestions lesion significance, and this is what the next two presentations are going to be focusing on today. I will really just focus on mostly on the first part of it time savings and identifications of stroke lesions in order to streamline the work into the hospital.
[5:44] Speaker 2: So what is what is a machine learning in itself is actually a way.
[5:50] Speaker 2: Identify a mathematical model that
[5:54] Speaker 2: What does a transformation like A goes into B. For example, you show a bunch of fruits and you teach them, hey, this is an apple and this is a pear, or insofar and so on. So it is.
[6:08] Speaker 2: Really we teach the machine a mathematical function that does that kind of a transformation. So how is it done? Well, we have an input in this case a database. We have a database of images and we assume that there is some sort of transformation that takes it to an output. We have a whole bunch of CT scans that some have hemorrhage, some are not, and we have, and we have to say yes, there is a hemorrhage or not, for example.
[6:37] Speaker 2: And then we compare with a ground to label label. This is where the expert users are located. Without the expert users, without the radiologists, without the neurosurgeons, all this cannot be accomplished.
[6:53] Speaker 2: So the having medical staff, the neurosurgeons, having the cardiologists and all the, all the folks into the clinical setting that makes this decision, it is extremely important. Without them, actually none of this will be possible. And what do we do, we compare this, this what machine tells us with the groundroot, and if we don't like it, we go and we readjust.
[7:22] Speaker 2: This function and we can go this thousands, 10,000 times and so far and it can go for days in training until we find the proper transformations that takes this point into that into an output that agrees with the clinical.
[7:41] Speaker 2: Uh example was a was a clinical labeling, let's put it that way, really, here is an example, you take an image and you have all these outputs to know is that for these outputs you really have to, uh, they are reset. In Other words, we need to talk with the physicians.
[8:06] Speaker 2: And with the radiologists.
[8:09] Speaker 2: To actually identify these labels.
[8:13] Speaker 2: The machine, once it's trained, will not be able to go outside.
[8:19] Speaker 2: Uh, this labels in order to identify a disease that is not, has, has not been taught to recognize and what it does, it calculates probabilities.
[8:31] Speaker 2: That it tells that, hey, based on my first iteration, the probability that this is a cat is very high.
[8:38] Speaker 2: Obviously we compare this with the normal things and with the normal labels. In this case, the prediction says that it is a cat, which is untrue. The ground truth is dog. Then we go back and readjust this function until we have a really good.
[8:59] Speaker 2: A prediction of this of this label and as it goes at it, you see like this predictions, these probabilities that this particular image is a dog is increasing while the Other ones are decreasing and we keep going at it until a through this through this iteration process until the error, it is a lot less between the prediction.
[9:25] Speaker 2: And on the actual labels. So we are looking for this function to get better and better. And finally, after many, many iterations, thousands and so far, the algorithm will do this kind of predictions. The note is that the algorithm is not never going to say this is a dog with a 100% accuracy. It's going to give you predictions.
[9:52] Speaker 2: Or probabilities based on the labels that you have in there. And if we go to a to a medical space and we say, hey, is this a intracranial hemorrhage is a ventricular hemorrhage or a subarachnoid hemorrhage, it's going to give you those probabilities that that particular presentation, it is one of those factors, so.
[10:16] Speaker 2: Finally, when we get the algorithm to say, hey, this is a dog, and the prediction and the ground truth agree, we kind of like minimize this error and we stop the training, and now we have a machine that every time we present something to it will be able to predict this kind of a label with a with a high.
[10:39] Speaker 2: With a high accuracy. So this procedure is done over all data. Usually for for this kind of trainings you need thousands, if not maybe 10s of thousands of images, which is extremely difficult in the medical world. It's not as easy as a dog that you can go online and download millions of pictures.
[11:00] Speaker 2: but only by going through this Approach, the methods, the model becomes generalizable, and you can use examples outside of the data set that you have. So another thing that about the AI.
[11:19] Speaker 2: And the classification of the objects. Let's suppose we don't have that. We have in this case, a patient information about hypertension or things like blood pressure and so forth. There are the the AI, what it does is actually separating.
[11:40] Speaker 2: These datasets and in this case you kind of see there is a clear separation or others are a little bit more complex. What the AI does it actually calculates this very fine line in here to separate.
[11:55] Speaker 2: Very like complex data structures or things that are not quite that obvious and one key factor is the following that if you have data sets where you have you label them versus let's say a disease or you're trying to say this patient has a has a tumor or not or so if
[12:20] Speaker 2: You do not see these trends with the naked eye. You look at the data and you don't see the trends. It's very unlikely that the AI will be able to do a good separation at it. So kind of advice is that every time we use parameters to characterize a disease, let's say a pressure drop, pressure gradients, or velocities across an intracranial stenosis, and so far it was a NOVAR, so.
[12:50] Speaker 2: If we plot that data and we do not see them separating, it is very unlikely that the AI is going to have like a magic kind of a point and immediately is going to separate the patients with the disease versus ones without. So in that.
[13:10] Speaker 2: Uh, um.
[13:13] Speaker 2: A kind of an idea. Let's, let's dive into the neurosurgery application. So how do we do this, uh, this, uh, training and so far? We actually go through thousands and thousands of patients. We look at the angiograms and we know precisely.
[13:30] Speaker 2: Who has what kind of disease we know precisely what is the outcome. These are patients with multiple imaging techniques and multiple very detailed outcomes. It's 3 months, 6 months, 1 year, and so far. Then we have.
[13:50] Speaker 2: Uh
[13:52] Speaker 2: Experts in this case neurosurgery, neurosurgery fellows or neurosurgery residents that can go through any of these data and identify every single aspect into the image they can say, hey, there is an aneurysm right there and an aneurysm here, and then what we do, we are kind of like picking these data sets in this case an angiogram, and we are compare it with what the fellows or the neuro neurosurgeons have done.
[14:21] Speaker 2: Neurosurgeon is more it's less we can compare notes with them when we are in doubt, but most of the things that you see over here are actually the work of of the fellows, residents and and trained people. So when we take this.
[14:39] Speaker 2: Data sets we train them and we create this function that does a transformation from an image like this to something like that and you say well what is the point of this? I mean I know an aneurysm how it looks like well the thing is if you will have someone let's say go and try to
[15:01] Speaker 2: Uh, get the data set and outline an aneurysm and then take some measurements from them. I don't know, like how much could the contrast took to flow through or what is the size of this aneurysm, what is the dome size, and so it might take a couple of minutes for them to bring up this, the, the data set and then and do all these measurements really.
[15:26] Speaker 2: The AI to go from something like this to something like that. It is fractions of the second once the model is trained, which brings
[15:35] Speaker 2: A lot of improvement in the workflow of the of the people who use it. So as of now for AI medical applications, at least the ones that we're working, what we're trying to do, we're in the phase where we're demonstrating equivalence to human performance. The organization like Food and Drug Administration or any kind of Other ones in Europe will not.
[16:03] Speaker 2: Accept any AI unless we kind of like show that they are they are equivalent, and I think the field it is in this phase in here and what we want to use it is to optimize the workflow. We want faster diagnosis. We want treatment planning. A lot of the folks now are asking us, well, give me the best place to drill a hole so I can place the clip over an aneurysm neck from the first shot.
[16:33] Speaker 2: Others are asking, well, tell me how to get this deep brain stimulation.
[16:40] Speaker 2: Uh, a pathway in such a way that I spare very difficult areas. Others are asking for patient triaging.
[16:50] Speaker 2: Some neurosurgeons are asking us, well, tell me.
[16:56] Speaker 2: The hopeless cases, so I don't spend time on those.
[17:00] Speaker 2: Um, uh, I need to, I need to kind of know if a person that comes with a, it's worth actually reserving the Surgery room and, and, and go, go for it or not. Others are asking, tell me what device to use.
[17:16] Speaker 2: Tell me the sizes. What, how many coils am I going to use? What type of coils? So I go from the first shot and be done. And finally, complication detections, things that are subtle that you cannot see, like placing a stent over a branch, and you don't see, you don't think that it's going to occlude later and like a pipeline, and then weeks later you figure out that, you know, the ophthalmic artery is gone, for example. So today.
[17:45] Speaker 2: As far as concerning stroke, I'm going to focus on these 4 aspects. I'm going to demonstrate equivalence in human performance, optimize the workflow. I'm going to show you that it does faster diagnosis and how it can be used for patient triaging.
[18:00] Speaker 2: So it is now standard for our uh um neurosurgery department to have a protocol like this in here. As soon as the patient hits the um
[18:14] Speaker 2: Um,
[18:16] Speaker 2: The emergency room. Uh, the first stop, it is actually the, um, CT scanner, not the MR, not the engine room. They go to the, uh, CT scanner, and there is a four-step protocol.
[18:32] Speaker 2: First of all, there is a non-contrast CT that is looking for hemorrhage.
[18:37] Speaker 2: Followed by a large occlusion. Once this is ruled out, we go through a second part where is a large vessel occlusion.
[18:46] Speaker 2: And this is an angiogram. This is a CTA.
[18:49] Speaker 2: While this happens, the non-contra CT, it's used for for AEC scoring if a hemorrhage has not been detected, and this is an Alberta scoring system where we are looking for an APEC score below the 6 hours to determine whether a patient is eligible for thrombectomy or not, and
[19:11] Speaker 2: After the 3rd, after the scoring, it goes through the CT perfusion, and the CT perfusion, based on the diffuse Clinical Trial and a lot of people are using now Rapidd AI AI to do this. We're looking at the
[19:27] Speaker 2: How much of the, what is the ratio of the infarct core to the penumbra, which penumbra is for tissue a risk for, for infarct. So based on this ratio in here, we decide whether a patient is going to go for the thrombectomy. So in general for mechanical, now the mechanical thrombectomy in our, at least at our center has become one of the most used.
[19:55] Speaker 2: Uh, methods to, uh, for these patients. Uh, and, uh, during this procedure, uh, uh, there, a catheter is placed in a device such as a uh a clot retriever is deployed over the, uh, over the actual lesion.
[20:11] Speaker 2: And the entire system is pulled out removing the clot. Uh, there are quite, um, their American Heart Association has very thorough guidelines about people and on how to use the imaging in order to detect, to determine eligible patients. And right now, uh, what we're trying to do is to optimize this workflow.
[20:38] Speaker 2: To be from whatever it takes now, uh, when the patient hits the um
[20:45] Speaker 2: Uh, the emergency room, it can take up to half an hour to 1 hour for it to get to the thrombectomy. So what we're trying right now is to do to use the AI to optimize this kind of a workflow. And I will go with development on the AI for ICH algorithm.
[21:03] Speaker 2: One of the applications for the AI, it is to identify, to identify the hemorrhage. Why is that? Well, it's mostly convenience, convenience workflow, and it is kind of a trivial radiological task if you think about it.
[21:22] Speaker 2: Most of the residents, even the fellows uh uh that are in Other departments, not necessarily Neurology, are, are, should be able to recognize patterns like this. Why do we do it is because that's how we start with all these technologies. We start with things that
[21:42] Speaker 2: are kind of, uh, you know, safe. We demonstrate equivalence and then we increase in uh in complexity. So right now, uh, let's say, uh, what is the accuracy of an AI application? It's about 96% accuracy. It has 96.
[21:59] Speaker 2: 7 sensitivity. It can go through on a normal computer. It can go through an entire stack of 320 or 512 slices and identifies where, whether there is an IC age or not.
[22:18] Speaker 2: Um, not only that, the, um, we can train this, uh, if we have enough data, we can train the AIs not only to say, hey, there is a hemorrhage, but we can actually have it, uh, you know, classify them. Well, it's an intra intraparechimal is a subdural, it's an intraventricular, or is a subarachnoid hemorrhage. So, um.
[22:44] Speaker 2: A lot of the AI applications like right now and recently at the Radiology Society of North America, there was a challenge about 2 years ago in identifying this type of this kind of intracranial hemorrhages.
[23:01] Speaker 2: And the person that it was a young gentleman from Japan that developed the software was able to detect this this kind of a
[23:14] Speaker 2: Uh, hemorrhages with an accuracy of 96% or even above. So, um, now, this is a medical product. So we see, heard about Rapid AI. They have a application which is both an invis AI, and both of them are mobile applications, which means the patients get this scan.
[23:42] Speaker 2: It tries to identify whether there is a hemorrhage or not, and then it blasts it. It blasts it to the entire.
[23:50] Speaker 2: Uh, staff, uh, of neurosurgery and neurology.
[23:55] Speaker 2: Immediately the doctors, if they are in an elevator or walking in the hospital, they get a notification on their phone. They go open their phones and they see the image, and then it allows them farther than that to actually go and communicate with their college, with their colleagues to about the next step, steps into this procedure. And again this is just convenience and workflow within probably a couple of
[24:23] Speaker 2: Uh of minutes, the entire staff is notified. They can see the, uh, the attending can get the results, looks at it, and gets his team going to actually, with regard to the next step in into the treatment of this of this patient. Um, well, let's go through a kind of like one on one, how is uh we do this. This is how my team has done the same kind of thing for intracranial hemorrrrhages.
[24:51] Speaker 2: So in this case we just determined the purpose is clear, determined patients with hemorrhage with a noncontrast CT. Then what we did, we went and asked our fellows to identify 200 intracranial hemorrhage patients. We took every single slice and we went through every single one of them and labeled it.
[25:14] Speaker 2: We have about 6500 slices and say this is positive, this is negative. We did it for different combinations of slice thicknesses and 2.5, 3 millimeters, 5 millimeters, and so forth to accommodate different kinds of acquisitions because this kind of, what I, I apologize, what I did not say in here, these platforms.
[25:39] Speaker 2: Can be deployed not only at the central hospital. It can be deployed to Other hospitals in the regions that do not have a
[25:51] Speaker 2: Neurology or neurosurgery staff in the hospital and as soon as their CT scanners detect they perform the CT scanner and upload these images, the center, the main center in this case, the comprehensive stroke center, which is at Buffalo University, immediately gets notifications. There is like 12 to 14 hospitals which are connected into a system like this and can upload these cases. Going back to this example.
[26:21] Speaker 2: Then we kind of like uh the network.
[26:26] Speaker 2: Again, it is trained on this kind of slices, and these are typical slices that we have that we upload into the into the network or into the AI, and the output is very simple. We call this attributes.
[26:43] Speaker 2: Is there a hemorrhage or is there not?
[26:46] Speaker 2: And then what we do is, obviously, in order to improve the um uh the AI performance, uh we'd have to do sometimes a some imaging segment, some imaging processing. And the key is that things like bones and Other structures and calcifications and so far they can throw off your AI.
[27:11] Speaker 2: So at the end of the day, AI is just looking at the numbers and tries to figure out what's going on in, in there and where these numbers are distributed, where is their intensities, and the simpler you make the problem for the algorithm, the better it becomes. And in this case, we do segmentations to throw away the bones and Other things like that. And when you look at an AI.
[27:37] Speaker 2: You are going to see something like this.
[27:40] Speaker 2: They are going to say first of all, well, what, how, what am I going to do with this? Well, most of the AIs cannot deal with full images data. We, there is a downsampling in there. 256 by 256 is actually quite big. There are others that go to shrinking the images to 64 by 64, and then the data is tested, is split, and what we do, we take 60 to train 10 and tend to validate.
[28:10] Speaker 2: And 30 in here it is you don't show that data ever to the AI algorithm and this is key, no, no kind of like a a.
[28:23] Speaker 2: Journal or regulatory organization will accept you testing your data on the same, uh, on the same one that you use for training.
[28:34] Speaker 2: And these are kind of like details of what's going on in here, but what technically is going at this level, we're taking every single image and we do some mathematical Operations, and these mathematical Operations are optimized in order to give you this output prediction in here, and this is kind of like the general output of a
[28:56] Speaker 2: Of an AI and I'll go a little bit into detail, more detail now, but here MCCV stands for Monte Carlo cross validation, which means that we look at this data and we split it in different configurations because you cannot risk at some point that you train your AI on a biased kind of data. You randomly split, split this data in order to make sure that.
[29:22] Speaker 2: Your AI performs good regardless of the splits that you have. So in, we do in general 20. Other people do less, Other people do a lot more. But at the end of the day, it gives you this plus minuses that you have in here.
[29:39] Speaker 2: The question is how good it is the AI in actually identifying a hemorrhagic stroke, and when you look at the probability for accurate diagnosis, it's about 96, and we do all calculations F1 score, sensitivity, specificity, positive predictive value, non-negative predictive values, the accuracy of the actual.
[30:01] Speaker 2: Um, the accuracy of the actual algorithm, Matthew correlation coefficients and the area under the curve. So we kind of like go and this is kind of standard for every single one that we do, for every single evaluation that we do. The training time.
[30:19] Speaker 2: Pretty bad, uh, if you ask, uh, it's about 18 hours. But in the prediction time, not per case, per slice, it's 6 milliseconds.
[30:31] Speaker 2: So once you train your algorithm, it's just go through these batches of images in really enlightening mode. So here are some examples. It's not always great at predicting. Here is a.
[30:49] Speaker 2: A successful one. There is a negative one, and again failed in here. Here it's a, it's a, it was a stroke positive and here is a stroke negative where it, it, it fails for for both of them.
[31:07] Speaker 2: So the next thing that we want to do is break down the AI and going back to this example where we try to predict what's happening to the AI when we present an image like this, it is to look inside the AI and see how it does the decision, and this is what is referred a class activation map.
[31:28] Speaker 2: So in Other words, as I said, we break open the map and we create this, we create this heat map showing where in the image network is drawing the classification from. So if we break down the AI, we get things like this, and what it does, it tells us that into the decision.
[31:51] Speaker 2: To classify this object as a dog, it uses a lot of information from these areas. In Other words, this area, it is highly counted towards towards a decision of the of the network. And then what we did, we went back and broke down the AI and we looked at it and see where was the actual.
[32:16] Speaker 2: Pixels that counted into making this decision because what if let's say our data was biased our CT scanner was somehow broken and
[32:27] Speaker 2: And adds something into an area of the bone or an artifact or image artifact and the AI somehow is being trained on that artifact to determine whether there is a a hemorrhage or not. And what we do, we break down the AI and it puts out these probabilities.
[32:50] Speaker 2: And using these probabilities, it actually decides whether a hemorrhage is present or not. And then we take these probabilities, we overlay over the actual CT scanner and say, yeah, it looks like this AI is looking at the features that matter and makes the right decision.
[33:08] Speaker 2: And without this, it's really very, very difficult to um
[33:16] Speaker 2: Um, who accept.
[33:19] Speaker 2: The results of such algorithm. So it is now almost a requirement for different journals like AI and radiology or so forth for the users or the scientists that perform these kind of studies to provide interpretation of the AI decisions.
[33:41] Speaker 2: And here is another example where it's a different kind of of hemorrhages, and again it accounted fairly accurate where the, the, where the hemorrrrhages and it shows display of probabilities in order to make that decision. The next thing that we want to do, it is actually an AI hemorrhagic stroke segmentation. In Other words, find out what is the hematoma expansion.
[34:12] Speaker 2: And uh
[34:13] Speaker 2: We did the same kind of things. We went through the 20, 200 confirmed intracranial hemorrhage cases and noncontrast City, but this time instead of like saying, hey, there is or there is not a hemorrhage.
[34:28] Speaker 2: An expert user went and by hand in the semi-automatic, actually, first use the threshold, identifies the location of the IC agent, kind of like had to draw this kind of maps.
[34:43] Speaker 2: This is a different kind of a pro uh problem.
[34:46] Speaker 2: In this case we're taking an image like this and we are trying to do a kind of labeling. Most of the people in the medical field knows what segmentation is pretty much is find out what is in within that image, and that is done again and even before, even for the the AI before it's done using what is referred as a convolutional neural networks.
[35:12] Speaker 2: So what do we do, we take like this kind of a square, we slide it over the entire images, and this filter in here, it is what we calculate, what do we optimize.
[35:25] Speaker 2: Technically what it does, it takes this, multiplies it with that, and replaces it, it replaces one value with what was the result of this mathematical operation and the AI, what it does, it actually calculates every single one of these parameters.
[35:45] Speaker 2: Uh, in, in the process.
[35:48] Speaker 2: And then what it does also, it does something like cold pooling. What it means it
[35:53] Speaker 2: It shrinks the image. You just take, like, for example, these pixels you see in here, and you just simply um you can go and uh and you'd look at max pulling, you're looking at the maximum.
[36:06] Speaker 2: Value in here and you replace and you create a smaller image and even a smaller there is different ways it's a max pulling average pulling that's that's less less important.
[36:19] Speaker 2: And when you do this, you go through, you've heard probably of the unit. This is what the unit does. You take the image, you do those mathematical Operations, then you make a smaller image and a smaller image and a smaller image. You get at this point where you do like kind of they encode or detect features into the image, and then you go back up onto this and you reconstruct the image with the information where is the labeling that you have at this point, and you recreate the image.
[36:48] Speaker 2: In this case it is going to come out and it's going to be fully segmented, so we took the data, we pretty much did the same kind of thing thresholding we augmented because we didn't have enough data into this.
[37:05] Speaker 2: And created something like a network like this, the same kind of splitting 240 patients by 40 by 120. We used somewhere into the vicinity of 2700 images.
[37:21] Speaker 2: And then we test it and see what are our ability to to actually segment this and find out what's going on. So when we look at now, instead of like saying whether this now we have quantities to measure and we are looking at the median volumes and how do they compare with the actual AI.
[37:48] Speaker 2: Well, the average user said it's about 13.5. Here we look at the hemorrhage prediction, it's 13.4. The difference between expert users and the actual AI is 0.1 mL.
[38:05] Speaker 2: I'm sorry.
[38:09] Speaker 2: So when we look at now Other things, this is a dice coefficient, and it kind of like compares structure differences.
[38:17] Speaker 2: For example, if I have a cube and I compare with another cube, what is the over predicted by the AI?
[38:26] Speaker 2: Are they the same? This is kind of like tells you how much, what is the similarity between the two, between the two volumes. And when we look again sensitivity, specificity, negative predictive value, positive predictive value, and the accuracy, this is pretty much perfect.
[38:45] Speaker 2: The, the key is in here is that
[38:49] Speaker 2: It takes about 24 milliseconds to calculate the volume.
[38:54] Speaker 2: Well, for a human, it can be anywhere to a couple of minutes in order to do it.
[39:01] Speaker 2: And here are examples where between successful and good dice coefficients and here in the the
[39:10] Speaker 2: You see the comparison between the ground truth in here in red and the prediction and the actual red curve on the outside is the edge of the ground truth, what whatever the expert user has done.
[39:25] Speaker 2: We're not always good. Here is a case of a very faint hemorrhage. This is a ground root.
[39:33] Speaker 2: And yeah, I totally failed to uh to get it.
[39:38] Speaker 2: The next thing is large vessel occlusion. Again, this is another thing that is.
[39:44] Speaker 2: Just convenience.
[39:46] Speaker 2: And all platforms Rapid AI, AI, Canon, and probably the Other vendors as well, now they provide this because it's a fairly easy way to detect it. And in this case, the large vessel occlusion applications that we are showing you is being trained on 3000 cases, validated on 450, and what it does, it automatically goes and detects where the occlusion is.
[40:14] Speaker 2: And based on this,
[40:16] Speaker 2: Whether the patient is a good candidate for mechanical thrombectomy or not.
[40:23] Speaker 2: Here is another example. Um,
[40:26] Speaker 2: This takes a little bit longer. It takes about uh 20 seconds.
[40:30] Speaker 2: And detects occlusions in M1 and M2.
[40:34] Speaker 2: And uh I'm showing in here the Canon version which has been clear for uh in July last year. But these AI and um Rapid AI have already platforms uh both of them uh on a mobile units uh as uh for, for this application. Um, most of the time, the algorithm will not show you.
[41:00] Speaker 2: Um, what is going on where, where the AI is, is just going to bring you to the slice where it believes that the, uh, large vessel occlusion is occurred and allows the physician to uh confirm it or not.
[41:15] Speaker 2: Here are different examples, right MCA, uh, right uh ICA occlusion and MCA occlusion, uh, for, for this, uh, for these patients where the uh LVO has been detected.
[41:29] Speaker 2: Others, uh Rapidd AI actually, instead of having that kind of a box, it gives you an overlay like this while the VS AI again, does the same thing. It brings you to the slice with a maximum intensity projection showing you most probably the uh the presence of the um
[41:53] Speaker 2: Um, of the large vessel occlusion. The notification, let's say if this CTA has been done at the satellite, uh, hospital, to the point where all the team receives a notification that there is a large vessel occlusion coming, it is about 7.32 minutes. It's slightly longer than the ICH detection.
[42:17] Speaker 2: Now, we're moving into a little bit even more complex part and this is the uh uh Alberta Stroke Program, the Early CT score uh score. And again, all these um uh
[42:32] Speaker 2: Um, uh, platforms provide this. The one from Canon is not submitted yet. And what it does, it's calculating the stroke to determine eligibility for thrombectomy within the six hour frame uh uh window. And, uh, again, it just goes, plays automatically these ROIs without any user interface, um.
[42:56] Speaker 2: And uh uh calculates, uh, shows the uh Hansel units if it detects hypodensities, automatically is coloring it into that way and presents the um in really fraction of the second prevents uh uh presents the uh aspect core to the Neurology and Neurosurgery department.
[43:19] Speaker 2: So, um, Rapidd aspects is shown in here. They have kind of like the same kind of Approach in uh into measurement. This is their interface, um.
[43:31] Speaker 2: And, um, and, um, uh, with the AI uh the same. I'm going to move to now into the actual CT perfusion and the AI infarct uh core segmentation, and this is based on didilution curve theory. We inject the contrast and we're going to have this kind of a distribution as a contrast uh propagates through the uh neuros through the um
[43:56] Speaker 2: Brain tissue and uh what um CT perfusion does, it calculates various parameters based on the shape of this curve.
[44:06] Speaker 2: So, um, maps like this are generated at every single, um, uh, voxel into the image. Uh, our center is really.
[44:18] Speaker 2: Um, very thorough about it. We are doing pretty much on every single patient minus some of the one that have intracranial hemorrhage.
[44:29] Speaker 2: So, um, uh, the idea is to find out into these maps, how much tissue is actually dead and it cannot be savage and how much tissue it is actually at risk of being uh uh uh uh uh of, of, of being infarcted.
[44:49] Speaker 2: So there are a lot of advantages in this case for the segmentation. When we look at these images, it is in general really difficult to figure out if we put, let's say, a user.
[45:06] Speaker 2: And ask him, segment to me based on this image, what is the penumbra and what is the infarc core. It's going to be kind of difficult. And we can use thresholds. A lot of big companies like Rapid AI, Vital, and Vis AI, they are using different thresholds, different parameters, but if you use those parameters on the same patients, you're going to come up with different numbers.
[45:32] Speaker 2: And there is a need to unify this Approach.
[45:36] Speaker 2: We think that the machine learning techniques might be able to quantify this location better because not only they are going to be trained.
[45:45] Speaker 2: Who, you know, uh,
[45:48] Speaker 2: Based on the thresholds, it will take into account where it's located. And this is where the AI.
[45:56] Speaker 2: Uh, I
[45:58] Speaker 2: Advantage come into place. So, and how do we do this? Well, we go to the ground truth. We take patients that not only have CT perfusion, but they also have 48 hours follow up DWI.
[46:15] Speaker 2: Diffusion weighted images and we create the labels to train the AI based on this data that we know that is ground truth.
[46:26] Speaker 2: We take the DWI, we take the infarct levels, we use a threshold on the infarct labels, and now we go back to the CT data and say, I'm going to use this to train for you to segment on the CT perfusion results. Well, uh, we have developed our own software in order to align the MR data with the CT data, which is not quite trivial.
[46:52] Speaker 2: In this case we are using the ventricles. We are using the ventricles into patients to actually in the bone obviously, but to actually align the data from the DWI.
[47:04] Speaker 2: With um with the actual CT data. And this again, it is kind of the, uh you see uh the unit if you think, if you look at it, first is going down and then we go up.
[47:19] Speaker 2: U convolutions, uh, and we go from something like this to something like that. So again, we take the CT perfusion data. In this case, we can take just, let's say time to peak or mean transit time and train the AI in order to do this kind of a prediction. And here are the results.
[47:39] Speaker 2: We can do just mean transit time and then not only that, we can take something like this on data and do some mathematical, really standard mathematical, mathematical procedures or Operations to remove things like this, which we know that they cannot be actual strokes.
[48:00] Speaker 2: And finally we're getting the the volumes, and here he sees a comparison between the between the
[48:10] Speaker 2: Actual CT perfusion, AI predicted volume, and the DWI volume.
[48:17] Speaker 2: So, um,
[48:19] Speaker 2: It works.
[48:21] Speaker 2: And the most of the times, uh, we get the um dice coefficient of 0.87, but not always. In this case, we, we're showing here where we failed. And um it's mostly it's because when we trained the data set, we did not have enough cases for this kind of location of the uh of the stroke.
[48:44] Speaker 2: The AI will not be able to recognize lesions or Other things outside the dataset that has been trained for. So if you remember, we were kind of actually um so I apologize for this. We uh we're kind of
[49:05] Speaker 2: Um, limited in, um, in the, in the number of patients. If we want to be, we use 63 patients and we use the labels in every single slice, so it's 63x 320, but
[49:23] Speaker 2: In order for this to be really, really accurate and so far, probably at least one order of magnitude should be used in this case.
[49:33] Speaker 2: So, um
[49:35] Speaker 2: This is showing in here the failure, and then we looked at all parameters. We can train it on every single kind of a data set. We can use CB CBF, cerebral blood flow, cerebral blood volume, time to peak, delay, mean transit time in order to train these data sets. And here you see before mathematical, uh, standard mathematical Operations and after, when they come out to be pretty, pretty accurate.
[50:03] Speaker 2: When we're looking at Other parameters, dice coefficient, sensitivities, specificities, accuracies, and so forth.
[50:12] Speaker 2: Pre Positive predictive values and negative predictive values, they're almost perfect. The only thing, the overlay is not quite good all the time. Well, but we shouldn't be kind of like limited about this, uh, these values in here. So the best dyo coefficient we got is cerebral blood flow, uh, flow in this.
[50:33] Speaker 2: Um,
[50:35] Speaker 2: What if
[50:37] Speaker 2: Instead of just using one map.
[50:39] Speaker 2: Now, we're moving.
[50:41] Speaker 2: And
[50:43] Speaker 2: The sky is the limit. We can tell. We can feed into the AI whatever parameters we want, and we're recently, we really published the paper came out like 2 days ago. We look at different combinations. We took every single combination that the CT perfusion gives us.
[51:04] Speaker 2: And we do not use only one map. We are using CBV CBF, CBV, TTP, MTT, and delay.
[51:11] Speaker 2: Or we can combine 4 by 5 by 4 or 3 by 4, and so forth and so on. So,
[51:19] Speaker 2: Um, we do all the combinations. We train an AI for every single one and see how it performs, and we see that some of them are not that accurate, but something like, for example, CBF and CBV or CBF and TTP, I apologize, gives you something better than anything that is on the market out there, and we can take this, and I'm pretty sure that if we will do this with.
[51:46] Speaker 2: As I said, maybe one word of magnitude even of patients, we're going to get these numbers way, way high.
[51:55] Speaker 2: So here is an example, and it still and we look at all possible values when we do these measurements. We look at the median infarct difference. We look at the absolute error. We look at the dice coefficients, positive predictive values, and then we rank with the performance. And in this case it looks for us that CBF, CBV, and MTT will perform the best if we train an AI with this kind of data.
[52:21] Speaker 2: So, um, in that, uh, with that in mind, I, uh, this is the faculty team.
[52:29] Speaker 2: And um that uh uh that has been behind all this uh all this uh all this work. Doctor Siddiqui is our director and Doctor Levivi is our chairman.
[52:40] Speaker 2: Um, Jason Davis is a neurosurgeon who is also a PhD into data analytics and machine learning, and Doctor Kenneth Snyder is also a a neurosurgeon who has a PhD in bio bio uh Medical applications.
[52:56] Speaker 2: We're really, really, nothing would be possible without this gentleman. All the work that the engineers and the students that are doing, it starts in the Surgery room. They are the ones that come up with a problem, and then we're trying to solve that problem. It's not the Other way around. We sit up there and try to figure out some sort of, and let's see what we can apply this AI.
[53:21] Speaker 2: It is this is the proper way to kind of like do this and I'm kind of and the people that have developed all that code that you have seen are right here on this page.
[53:35] Speaker 2: In particular, Ryan Rava, the 3rd in line there, has done all the CT perfusion and CT infarct segmentation, uh, and, um, um, is, uh,
[53:48] Speaker 2: Uh, involved in, in all this kind of Research, um, the rest of the team has done a lot of work on the work that I will present in the next two presentations where we become a little bit more ambitious or riskier and move from just trivial kind of things like not trivial but workflow optimization into things like um.
[54:14] Speaker 2: Detecting complications or predicting in real time.
[54:20] Speaker 2: How your neurosurgical procedure goes.
[54:23] Speaker 2: And um my students over here are also very, and myself are very interested in talking to any of the um of the neurosurgeons to find out what is your take on actual AI. My email is right here, um, cnanit@buffalo.edu, and we would really, really like to talk to you and see.
[54:50] Speaker 2: Um, what is your take on AI in general and where you kind of like see, uh,
[54:59] Speaker 2: This kind of feature is helping you uh without and um without kind of picking anything. I mean, the AI will never take anything away from neurosurgeons or uh neurologists. It's just enhancing their activity. And as I said before, people ask us all sort of things like what is the likelihood of return to normal.
[55:23] Speaker 2: Uh, should I go for these patients to open up or not? Uh, should I go open skull Surgery, endovascular Surgery? If I go endovascular, what kind of should I do? Should I do a web device, a pipeline to diversion, or coiling? These are kind of like, oh, and uh if I go, for example, with a pipeline, what size should I use?
[55:46] Speaker 2: Could I go with a 4.5 millimeter, 4.25, 4.7 4.75. Oversize, the same size of the vessel or undersized?
[55:57] Speaker 2: So, um, these are things that I think the AI can bring, uh, can solve by
[56:05] Speaker 2: But it requires really.
[56:08] Speaker 2: A multi-center collaboration, Other neurosurgeons, uh, uh, one center cannot solve any AI kind of a problem.
[56:18] Speaker 2: It requires multi centers.
[56:21] Speaker 2: Multi-expertise and uh uh uh in large groups. Thank you guys. And uh I'm open for questions.
[56:39] Speaker 2: Hello?
[56:48] Speaker 3: Uh, Mr. Yolia, can I ask a question?
[56:51] Speaker 2: Yes, please go ahead. Uh,
[56:54] Speaker 3: so let me introduce myself first. I'm a neurosurgeon. I'm a resident neurosurgeon from Turkey. Excellent. And currently we are, uh, we are also working, uh, we have a small Research group that we are trying to set up. We're just at the beginning right now, and we have a very small, uh, AI prediction algorithm that we have, uh, developed on our own. It's just
[57:20] Speaker 3: Uh, right now, it's just, uh,
[57:24] Speaker 3: detect if there is any intracranial hemorrhage. Uh, nothing more. But we want to advance this. And, I am kind of, I'm a neurosurgeon, but I have some computer background. And I made that, uh,
[57:42] Speaker 3: that program, and we want to advance this, uh, in a way, uh, that we can use in our Clinical trials. And I wanted to learn if we can participate as a, uh, as a group, as a center to your, uh, projects, and, uh, if not, uh, if there are any open source
[58:05] Speaker 3: Uh, tools that we can utilize, uh, for that, uh, for, for our aim.
[58:15] Speaker 1: Yeah, this is exactly the thing I wanted to see happen, uh, Cyprian.
[58:19] Speaker 1: Mhm.
[58:21] Speaker 2: I, um, we are a, uh, Research center and that
[58:27] Speaker 2: Is craving kind of like collaborations with everyone.
[58:32] Speaker 2: Uh, with a lot, uh, uh, we, we cannot, uh, this kind of AI things, we, we cannot, even if we have ideas.
[58:42] Speaker 2: We are going to end with those kind of false predictions that I've shown you. If we're using only Canon data and then somehow Z Siemens data comes in, it's going to be, it's not going to work properly. We know, we know that that is dependent a lot on.
[59:03] Speaker 2: The data that you're training for. So, yes, we are very, very much interested in collaborating and working on, on this project. As far as concerns, the um the co-development, I think um
[59:18] Speaker 2: Uh, code development now in the AI is, has become quite trivial. Um,
[59:26] Speaker 2: There's YouTube tutorials. There is code available everywhere. You just, uh, um, you just an examples, you can, you can write that kind of a code, but the key, I think for medical imaging is not necessarily this code. It is what you do to the data before it goes in.
[59:49] Speaker 2: And understanding how you can reduce those variables that will throw off your AI.
[59:57] Speaker 2: So I think there is where the problem is or how to fix or.
[1:00:03] Speaker 2: Different patient variability for artifacts or how to account for the physiology or the physics of the actual imaging process. So I think as the, the medical imaging world, it is a little bit different than
[1:00:23] Speaker 2: Figure out cats and dogs and uh pedestrians versus uh versus dogs uh and uh and so forth and so on. So I think uh there is where we can actually collaborate and where also it is the uniqueness of the problem that we are dealing with.
[1:00:44] Speaker 2: So I
[1:00:45] Speaker 4: have I asked you that.
[1:00:47] Speaker 1: Yeah. Go ahead, Marc, go ahead, Marcos.
[1:00:50] Speaker 4: Yes. OK, great talk, really interesting topic. But do you think for brain tumor, we'll be
[1:00:58] Speaker 4: Difficult than stroke.
[1:01:01] Speaker 4: What is your opinion about that, which is the difficulty for brain tumors prediction?
[1:01:08] Speaker 2: What I've seen so far in terms of uh uh tumor, brain tumor is it is to figure out if it is, what type of a tumor it is.
[1:01:18] Speaker 2: Uh
[1:01:21] Speaker 2: Things like uh differentiated between a glioblastoma versus uh astrocytoma or something along those lines.
[1:01:30] Speaker 2: Um, also segmentation to take the volume out, uh, and give, uh, measurements, but I have not seen yet.
[1:01:43] Speaker 2: Things like, for example, uh, to um
[1:01:47] Speaker 2: guide a Surgery in real time yet.
[1:01:51] Speaker 2: Um, I have not seen, uh,
[1:01:55] Speaker 2: Yet, um.
[1:01:58] Speaker 2: Things like, uh, you know, to, to
[1:02:02] Speaker 2: Indicate kind of like how to uh the Approach, how to Approach this for, for evacuation and so forth uh those kind of things I have not, or how to Approach the tumor to spare like uh uh sensitive areas around it. So people uh neurosurgeons have asked us for this.
[1:02:23] Speaker 2: Um, another thing is, actually, I've seen the AI which is out of area, my area of expertise in uh actually looking into the genomics of it and, uh, uh, determine uh whether it's going to be responsive to a certain, uh, drug and so.
[1:02:42] Speaker 2: So I've seen those applications as far as concerned as.
[1:02:47] Speaker 2: I, uh,
[1:02:50] Speaker 2: Kind of like focus on endovascular applications and stroke. And eventually, we're going to move towards, uh, towards also, um, some, uh, some brain tumors. Um, but the order on which this team is going to go about is endovascular applications, uh, a lot of them.
[1:03:10] Speaker 2: Um, and, um,
[1:03:12] Speaker 2: Our team is really, really keen on Spine, uh, and cervical, uh, intervention in particular. Um, Other than that, uh, it's, uh, as I said, as I, as I showed you before, we're a team of, um,
[1:03:26] Speaker 2: Uh,
[1:03:28] Speaker 2: Couple of people, you know, a dozen people. Uh, we cannot solve all the problems, but um
[1:03:36] Speaker 2: I think uh on the uh even on the tumors, there are a lot of applications and we'll be willing to collaborate with others um uh if uh if there is a need.
[1:03:58] Speaker 1: Is there, uh, is there a place where all the you can Research, look at all the Research that Other centers have done besides Buffalo? Like, where are the big Research centers, uh, would you say?
[1:04:11] Speaker 1: In this area of Artificial intelligence and uh
[1:04:14] Speaker 2: neurosurgeon,
[1:04:14] Speaker 1: John,
[1:04:15] Speaker 2: John Hopkins, uh, I'm talking here like neuro neurosurgery applications, and so, um, Johns Hopkins, uh, uh, it is huge. There is a Jeff Stewartson in there with his team.
[1:04:30] Speaker 2: Um, there is, uh, the Duke folks are also very, very much involved into this, uh, Mayo Clinic, uh, they have, uh, a lot of, uh, applications and so, uh, UCSF, uh, also a big center and a lot of expertise right there, um.
[1:04:51] Speaker 2: The, uh, one of the, uh, the issues that stops us from collaborating, uh, all these teams, uh, is, uh, is data sharing and the RRB's problems. Uh, so, uh, we're looking into ways to go to actually do machine learning without actually
[1:05:12] Speaker 2: Having the data downloaded onto our one center, so, uh, what is called, uh, as federative learning, uh, we're very interested into that, that uh the training.
[1:05:24] Speaker 2: Actually takes kind of place into the cloud and we are using the data and the labels at that institution center without ever needing to take it out. Uh, IRBs as of now, and, uh, uh, patient, uh patient information sharing, it is, uh, I would say the, one of the, uh the second.
[1:05:46] Speaker 2: Most, uh, difficult, um, or the second blockage into this Research.
[1:05:54] Speaker 2: First being
[1:05:56] Speaker 2: Um, you need expert level, uh, uh, expert, uh, users to label the data, and there's, you know, um, not too many neurosurgeons in the world or people with a neurosurgery expertise to do that and their time is very valuable.
[1:06:16] Speaker 1: Of course. OK, um, more questions or comments from the panelists.
[1:06:22] Speaker 1: The floor is open.
[1:06:31] Speaker 1: Um,
[1:06:32] Speaker 2: if
[1:06:32] Speaker 1: you
[1:06:33] Speaker 2: have questions and if you'd like, uh, to talk to, to my team and, uh,
[1:06:38] Speaker 2: Um, in a different context and so far we're all willing and open to, to have Zoom meetings and, uh, discuss with you. We're stuck, uh, in our labs in remote, uh, working and so, so we're more than willing to, to meet with you.
[1:06:55] Speaker 1: OK. You, you want me to put contact info in the chat?
[1:06:58] Speaker 2: Yes, so I'll, uh, I'll put my email in
[1:07:02] Speaker 1: here. Yeah, and I'll just copy and paste it.
[1:07:10] Speaker 1: You just emailed it to me?
[1:07:13] Speaker 2: Not yet. Uh unfortunately, I did it to only one person. Let's see.
[1:07:23] Speaker 2: There it is.
[1:07:25] Speaker 1: Oh
[1:07:29] Speaker 1: So,
[1:07:32] Speaker 1: Um, or did you show it on the screen or
[1:07:36] Speaker 2: you in the chat.
[1:07:38] Speaker 1: Oh, OK. Is it OK, OK.
[1:07:42] Speaker 1: Oh, there you go. Does everyone get that? You see, uh, see his email?
[1:07:50] Speaker 1: OK, it'll also be posted on the neurosurgical TV uh page.
[1:07:57] Speaker 1: I'll make sure I save that. Let me save that uh email.
[1:08:04] Speaker 2: So next, uh, next one, we're, I'm going to present, it's going to be a little bit shorter, but we're going to go, I'm going to show you how we can use some of these AIs in order to, uh, detect complications in real time during the procedure, to endovascular procedures next week with the shorter version of
[1:08:25] Speaker 1: this. OK, very good. OK, any closing comments or questions or just someone that wants to say hi.
[1:08:36] Speaker 1: But you know, recipient, this Technology is, is kind of new as far as interacting over something like this, and uh, so, and, and the audience uh is probably like me, they don't know much about it yet, you know, but hopefully we'll learn. Hello, Daniel, any comments, or Abu Bakar, any comments or questions, or just to say hello?
[1:09:03] Speaker 1: That's OK. OK, I guess everyone's shy today. OK, Cyprian, thank you very much. We'll have this, we'll have this edited, sent to you. Thanks for your time. Thanks for all the Research, and uh we look forward to next week.
[1:09:20] Speaker 2: OK, have a good one. Thank you.
[1:09:23] Speaker 1: Bye-bye.

## In this course

- [Next: Artificial Intelligence: Complications Detection During Endovascular Procedures](https://doc.social/llms/courses/neurosurgery-and-artificial-intelligence/artificial-intelligence-complications-detection-during-endovascular-procedures.md)
- [All lessons: Neurosurgery and Artificial Intelligence](https://doc.social/llms/courses/neurosurgery-and-artificial-intelligence.md)