Very good, so we'll go ahead and get started now. This is our fourth year, but it's our first year where we actually have AI systems in our unit. These are approved now throughout Europe, America, they're approved here in the Middle East. We're using these systems in our unit. Patients are coming to us now asking about that AI colonoscopy they heard about. So without further ado, I want to go ahead and introduce our first moderators of this session Practical Applications of AI in the EndoLab, our two moderators, Dr. Amrita Sethi of Columbia Presbyterian Hospital, New York City, and Dr. Tyler Berzin of Harvard Medical School Beth Israel Deaconess Hospital in Boston, Massachusetts. Let me turn it just back to Pratik. Okay, thank you very much, Mike. It's my pleasure to introduce Irving Waxman, and in this session on Practical Applications, Irving's going to discuss recent advances of artificial intelligence for esophageal and gastric neoplasia. Irving is in Chicago, he's professor of medicine, as well as has a co-appointment in the Department of Surgery, and just moved to Rush as the chief of gastroenterology there, and I've known Irving for several years, and not just a great researcher endoscopist, but a wonderful person as well. So Irving, welcome. First of all, good morning. I want to thank Dr. Sharma for the very nice introduction, as well as Dr. Wallace, and the ASGA for inviting me to participate here this morning. I was thinking when they asked me to do this that Waxman artificial intelligence is really an oxymoron. What I'm going to try to review for you this morning is recent advances in AI for esophageal and gastric neoplasia, and potentially other applications. My slides don't run when we're talking about technology. Perfect. Thank you. You were never a chief resident, were you? Never. These are my disclosures. And I don't know how, I mean, I'm a science fiction fan, and I don't know how you remember Stanley Kubrick and Space Odyssey 2001, the monolith. And this is really how my interest in AI started. I definitely am one of these guys, somebody who, when he was a fellow, the biggest question was, do you use endoscopes for treatment of GI bleeding or not? And all of a sudden, I started seeing all of this information coming out about AI helping the endoscopist in detection of neoplastic lesions. And that really got me all excited and my interest started. So how does AI interact with the endoscopist? Well, in the present time, the way we do endoscopy, we find a lesion and have high versus low confidence, and we're the experts. Now, AI can interact with us in different ways. A, it can provide us a second read. We find a lesion, we take pictures, and then with AI, we interrogate the lesion. And it can, for example, in colon, it can tell us polycharacterization, size, histology. So that is one way. The second way is having a concurrent read where the endoscopist and the AI system are working at the same time. And you're probably all familiar with current available technology, like GIG News, or I know there's technology from Olympus, et cetera, Fujinon, in which right now you can do poly detection. So you have a second pair of eyes working with you at the same time that you're doing your procedure. And then the future is what is this black box question is, can you do AI alone? Meaning, can you have like self-driving cars? Can you have self-driving AI with a capsule in the colon, et cetera, that's doing everything for you? And that is probably the scary, exciting part, and that the future is yet to be seen. And this is one of the softwares. This is not Opera GI, but this is an area where you get detection and diagnosis with AI in the colon. You can see here that the lesion is detected. But the interesting thing about the next generation of AI colonoscopy is that it can also predict the histology and the size. So that is what you're going to start seeing in the next iteration of the software, the algorithms that are coming out for endoscopy. Let's talk a little bit about what's out there right now. This is a meta-analysis looking at artificial intelligence-assisted detection of Opera GI lesions. These were 23 studies in the literature that fit the criteria for the meta-analysis. Almost a million images were evaluated. And you can see that the area under the curve for AI detection of neoplasia in stomach, barrettes, and squamous esophagus, as well as H. pylori, were really very, very high respectively. So it really seemed to work extremely well in detecting pathology. In this particular study, when they used MBI, they also found that using MBI with AI was superior to white light endoscopy in detection of squamous cell changes in esophageal cancer. And you can see here again, it was very statistically significant. And the other thing that they found in these particular studies in the meta-analysis is that AI was superior to the endoscopies in the detection of neoplastic lesions or H. pylori as a whole in the Opera GI tract. So very, very interesting what the technology is doing. These are the receiving operating curves. And this is gastric neoplasia, barrettes neoplasia, and H. pylori. And you can see the comparison of AI and the endoscopies with areas under the curve really close to over 0.9, which obviously is very, very significant. And again, AI performing, if not equal, but better than the endoscopies. This is a study from Japan from Dr. Fukuda and collaborators comparing AI with experts endoscopies to diagnose squamous cell carcinoma of the esophagus. And this is an example of the images. This is MBI of the esophagus and magnification. And you can see here the box that appears in the area of the neoplastic process. And what the study showed is that using the AI algorithm, there was a significant higher sensitivity for detecting squamous cell carcinoma and a higher accuracy for characterizing what was dysplastic versus non-neoplastic in the images that were evaluated. And when you're comparing the experts, which are these red triangles, look at how well AI performed in comparison to the expert when interrogating the same images or videos. So again, AI performing superior to this particular group of experts. This is a study looking at the difference performance between expert endoscopies and computer aid detection system using magnification endoscopy and MBI. And again, this is an area that was identified as the potential neoplasm. You can see here, you're all familiar with the boxes that appear when you're engaging the AI system. And in this particular case, this system allows you to identify dysplasia versus normal tissue. And you can see here the early gastric carcinoma versus just normal mucosa next to it. What was interesting in this study was that the artificial intelligence algorithm performed exactly the same as the expert endoscopies. So why do you need these? Well, these are expert endoscopies in Japan. Imagine in the United States where we, quote unquote, do not see early gastric cancer with something like this. A, it may allow us to find it. And B, for us who are not so familiar with recognizing early gastric cancer, this may present a huge improvement. But not only these algorithms allow you to detect new lesions, it can also allow you to diagnose. For example, in this particular study from Nagao, they looked at AI systems to predict invasion depth of gastric cancer. And in this particular case, I mean, the system was developed to recognize submucosal invasion or deeper in gastric cancer. Why is this important? Well, when you are selecting patients for endoscopic therapy, obviously up to down to SM1 is something that you're going to remove endoscopically. If it's a lesion that SM2 or more, that patient should undergo surgery. And you can see here how the AI system performed in diagnosis of SM2 disease. And that's why the oxymoron part, I cannot even work slides. You can see here the accuracy in detecting SM2 invasion more than 500 micrometers. It's over 95%, no matter if you're using white light, MBI, or just indigo carmine. So again, you can see the potential of this technology, not only in detection, but diagnosis. What about, this is a newest paper on artificial intelligence algorithm for detecting endoscopic picture of eosinophilic esophagitis. I know that I was just talking on neoplasia, but this seemed very apropos. And you can see here how AI performs against beginner endoscopists, fellows, or even experts. And AI performs better in not only detecting, but diagnosing esophagitis when you use the endoscopic eosinophilic esophagitis reference charts. You can see here that AI performs much better. These are features detected using gradient based visualization, which makes it, again, very, very interesting. You can see here the furrows being in eosinophilic esophagitis, the areas of exudates. So again, where does this help? Potentially for beginners, for trainees, being able to detect this without really indicating to you who needs to do biopsies, et cetera. You can see, again, the power of the technology. The following slides, I really want to thank a colleague and a friend, Professor Jacques Bergman from AMC in Amsterdam. He was kind enough to lend me some of his slides for cat detection for Barrett's neoplasia. We actually participated in this study as one of the expert groups. I suspect Dr. Sharma also. And bottom line is that the cat diagnostic system performed in regards to sensitivity of neoplasia was 90% and specificity of 80%. This is, again, if you... Okay. It's alive. Sorry about that. Do you guys know why it's doing that? I'll jump that slide. This is the data on expert endoscopies against a general gastroenterologist. And then having the algorithm in regards of detection of neoplasia to the general gastroenterologist as an aid. And what you can see, these are the image test sets. Endoscopies will just show images, the experts, the general gastroenterologist, or video test sets. I think we had to review over 100 plus videos, which was interesting. And what you can see here is that for the algorithm, for AI, the sensitivity, as I mentioned, for detecting dysplasia or more, high-grade dysplasia or more, was 90% with an 80% specificity. How did the experts do in comparison to cat, to AI? Well, we did relatively okay. Sensitivity of 87%, specificity of 86%. Very similar. But this is where it actually really helped this technology. When you look at non-experts without the AI system, the sensitivity was 74%, much lower than cat, and 89% specificity. Now look what happens when you add the technology to the non-experts. Their sensitivity goes dramatically up. So they got much better at diagnosing advanced neoplasia and barrettes when using this particular algorithm. Now what happened, that was still images. What happened with the video test set? Well, again, the AI algorithm showed a sensitivity over 90%, specificity of 80%. Experts were very, very similar. The non-experts, the sensitivity was about 65%. And look what happens when you engage AI algorithm in these particular videos. Their sensitivity goes dramatically up. So this totally really illustrates the fact that it can actually have a direct impact when you're actually looking at barrettes esophagus in identifying dysplasia without having to be an expert, which I thought was a really extremely useful study. This is examples of a detection of dysplasia and barrettes. These are images that we had to evaluate. And you can see here the AI accurately predicting advanced neoplasia and dysplasia with barrettes. Another case of the test series, you can see here barrettes esophagus, and with the AI software you can identify the area of advanced neoplasia. So in summary, AI-assisted endoscopy is really an evolving field with an exciting future in my opinion. While a computer-aided diagnosis allows us for detection, differentiation, and characterization of neoplastic and non-neoplastic disease processes, there's a lot of false arguments concerning AI in endoscopy with risk of implementation, particularly de-skilling the endoscopies. We just had Dr. Parasa actually lecture at the University at Rush on the GI ground rounds for us, and actually the question from the head of the fellowship program was, do engaging AI for detection of colon polyps makes the fellows less skilled endoscopies? Which, because they become over-reliant on AI during their colonoscopy, I kind of think that it doesn't. It actually allows them to learn better and faster, but, you know, again, it's still to be debated, and we need trials and more data to make those conclusions. There's question of the AI systems being very heterogeneous on representative training databases. What happens in the West doesn't happen in Asia. Are they universal? There's questions about hacking, and all of this needs to be considered as we implement AI in endoscopy. So I actually tend to think, where does AI fall in endoscopy in my field? I think I follow Sarah Turnbull. For all of you who don't know Sarah Turnbull, when you're wearing your N95 masks, she was the integral part of the design of N95 masks, and this is like 25, 30 years, actually like 40 years ago. She worked for 3M, and she actually then became the head of the design school at Stanford, and this is what she said. If you don't stretch it, you don't know where the edge is. So I think it's a very exciting time for AI. I think it's here to stay in the endoscopy room, and I'm going to stop here and thank again Pratik and Mike for their invitation.

AI systems in medical unit

practical applications of AI in endoscopy

Dr. Amrita Sethi

Dr. Tyler Berzin

Dr. Irving Waxman