Okay, thank you very much and good afternoon, everyone. So AI and Endoscopy 2023 to be followed by 2033, I think by Scott, so I'll be just giving you a glimpse about where are we today and how can this help in your practice today. So at this year's meeting, I think the theme you could have easily seen has been about artificial intelligence. And so the questions always come up is, how does it work? And what can it actually do for you in practice? So let's just look at this. So how does it actually work? So you know, it's your polyps, how polyps detected, how does the bounding box come out? This is what the promises of AI and endoscopy, we of course, are focused on computer vision in the practice of endoscopy. We can look at quality assessment, can it help us improve the quality of both upper and lower GI endoscopy, documentation and automated reports that would make our life so much easier if that was done automatically. And of course, in education, and of course, we've got large language models now which are being used for education. So how does it work in this? Let's look at this. So everybody like James Bond? Yes. So this is a clip and you can see that this is how the models are trained. So it's telling you here, this is a vehicle, that's a motorbike, that's a person, James Bond right there. And so you train the model by telling them what's a person, what's a vehicle, who's James Bond, who's his assistant. And sometimes there can be false positives. So sometimes the algorithm does think that I was James Bond, but you know, that's really not true. So you can get some false positives with it. But at the end of it, it also gives you the accuracy of how accurate it is in recognizing the car, which is about 98%, and then in recognizing a person close to 100%. So that's how these models work. So now apply that same concept that I just showed you in that clip of James Bond to colonoscopy. On the left, you have a colonoscopy, which is being done without the AI. On the right, it's being done with the AI. And you can see that the similar box concept, that bounding box appears, which tells you, okay, here's a polyp, and it draws your attention to a subtle lesion, which is there. So this is how it can tell you about it. Taking it a step further, it can characterize it for you. This is the CADX application of computer vision. And you can see it's seeing the polyp, and it's calling it a hyperplastic polyp. On the right lower quadrant is a heat map, which is also telling you the location of the polyp. So it can recognize polyps, it can help characterize polyps for you. It can probably do a little bit more. If you look at the left hand side of the screen, with the speedometer, it's telling you it's a neoplastic polyp rather than a hyperplastic polyp. But right below it, it's also telling you what the size of the polyp is. So you can see here's a polyp, it is diagnosing it for you, and then it is telling you that this is a diminutive polyp. So very soon we'll help with this, in which it's recognizing the polyp, it's characterizing it for you, and it's also telling you the size of the polyp in this situation. How about in patients with Barrett's esophagus, the same concept, okay? So here's a software which is looking at the entire area of Barrett's esophagus, which is grading it for you. And then there are subtle areas of neoplasia within it. This is an area of intramucosal cancer. And now as you get close to it, it is telling you what that entire extent of that area is, which is clearly seen on narrowband imaging, but also telling you the extent. So for therapeutic endoscopists, you know exactly what you want to resect in this situation. So these are examples both for the upper GI and the lower GI tract for looking at these areas in the field of endoscopy. So if I can go to my next slide, please. Moving on with that same concept in upper GI to low-prevalence lesions like early gastric cancer. So I think colon polyps, many of you may think, well, will it really help me? My ADR is already great. But for low-prevalence lesions such as Barrett's cancer, early gastric cancer, I think in those cases, since we don't see enough of those, I think there may be a real help with computer-aided detection and computer-aided characterization in this situation for you. This is the same thing, at least on a retroflex view, another area of early gastric cancer, which is seen, and it's being able to recognize this thing for you. So moving on from computer vision to quality assessment, right? Because that's a key thing. We want to perform endoscopy, but we want to do a high-quality examination. So how can this help us in performing a high-quality assessment in endoscopy in this situation? This is assessing the BALPREP, the Boston BALPREP score that we all use. So you can see it goes from 0, 1, 2, and 3. 3 is, of course, the ideal that we want. You can see with the help of this software, it's telling you what the Boston BALPREP score is. So at the end of your report, you don't have to think at the end of the procedure, well, was it a 2? Was it a 3? Was it a 1? Because you've probably removed several polyps along the way as well. So how does this perform? You can see that the overall AI accuracy for this was significantly better in grading the Boston BALPREP score as compared to at least novice endoscopists, but also better than expert endoscopists in this situation. So this is an example of how AI can help in endoscopy assessment of the polyps, or sorry, of the Boston BALPREP score. This is looking at secal intubation rate. At the end of a quarter, we always get asked, well, what was your secal intubation rate? You go back into the charts and you do it. It can be automated for you, and you can see it on the left-hand side of the screen at the bottom. When you reach the cecum, that will get colored in pink for you. So you've reached the cecum, so you can start withdrawal. You don't have to worry about, is this the AO? Is this the ileocecal valve? And then you start withdrawing. The withdrawal time also now on the top left with the speedometer can gauge you the speed of your withdrawal. Are you trying to do a two-minute colonoscopy to get to the second next colonoscopy waiting for you, right? You shouldn't be doing that. So it tells you the speed of withdrawal. If you're withdrawing too fast, it will go in the red. If you're withdrawing adequately, it will stay somewhere in that area. So after two minutes, it's recognized the landmarks. It's telling you now about the withdrawal time, and you can start recording that. And this will be recorded for you. So you don't have to go and ask the nurse, well, how much time did we spend? What was the withdrawal time for me during this colonoscopy? It is all automated for you with the help of this AI software. The next thing in quality is documenting the adenoma detection rate. How does this happen? Manually, right? All of us assign either a nurse practitioner or a nurse or a technician in your unit to find out what our ADR is because the endoscopy report and the pathology report don't talk to each other. And so it's very difficult to do it. This was a study comparing two different types of AI methodologies and comparing it to manual retrieval of your ADR and what they were able to show that your polyp detection rate, your adenoma detection rate, as well as the bowel prep rate was very similar between manual extraction by a human versus extraction by a machine. So what this shows is that the NLP was able to extract the entire data for the entire institution under 30 minutes, whereas manual extraction required 160 man hours for 600 patients. So if you have 6,000 patients, that's 1,600 man hours or person hours that we are all wasting when it could be done manually for all of us. So helps there for quality in endoscopy as well in this situation. Next let's go to documentation and automation of reports and how much of the mucosa are you actually seeing in documentation and quality. Here's looking at an AI system, which tells you what percentage of the esophagus, the stomach and the duodenum you have evaluated. So have you looked at the G-junction? Did you look at the incisura? Did you go into the second portion of the duodenum? So you can see that this is guiding you and telling you that you missed 20% of the stomach during your upper endoscopy. And I think for EGD, this will be critical because all of us want to do a two minute upper endoscopy so we can go to the next colonoscopy, right? It pays more, more RVUs, correct? I mean, that's reality. So this will help improve the quality of the documentation during upper endoscopy. So documentation can also be done. Clinical care coordination, but more importantly, education, right? And it's very difficult to talk about education and AI without getting into, if this mouse were to work, next slide please, about looking at large language models or, you know, chat GPT kind of functions. So here's a model which is called glassai.com. And those of you who haven't been to it should probably look at it. So you can give a use case. So I'm asking this 55 year old patient with worsening heartburn, and this is being done real time. It is telling you the differential diagnosis for the patient, that this patient has either GERD, erosive esophagitis, could have esophageal adenocarcinoma. But at the bottom, it's also telling you what the treatment plan for this patient should be. Now this is available. You can go onto this website and check it. This is AI working there for you in this situation. So look at all of this. It tells you the differential diagnosis, and it's telling you what the treatment plan and it's pretty accurate. Again, requires probably a manual check, but it's pretty accurate in this situation. The same thing can happen with chat GPT as well, right, to do it. So what's coming soon? Not Mission Impossible, but something similar to this. And so this is what I was talking about earlier. It is telling you what the nice classification of this lesion is, whether it's a granular, non-granular lesion. But then it will also predict for you whether this is suitable or unsuitable for endoscopic resection. So this is coming soon for all of us in this situation. The same thing for EGD. If you're doing ESD, you can see the knife and that's yellow is the submucosa. The knife is getting into the submucosa. There's a blood vessel which will be highlighted in red. So you will try to avoid that blood vessel while you're trying to do that. So you can see that we're moving into all these diagnostic and therapeutic EGDs into this situation. So I'll stop right there, and I'll tell you how it's going to impact GI practice for us is, endoscopy, CADx and CADe will help us with navigation, optical biopsy, and other, reducing our miss rates. And if you look at text-based data, it will help us in looking at endoscopy and pathology reports and also allow us to do a high-quality endoscopic examination. Thank you very much for your attention.

artificial intelligence

endoscopy

polyp detection

lesion characterization

documentation