Well, thank you. Let's go ahead and begin session two, titled From Augmented Diagnosis to Autodocumentation. We have three distinguished moderators, Dr. Seth Gross from New York University Langone Medical Center in New York City, Dr. Nayantara Silho from Mayo Clinic Rochester, and Dr. Helmut Messmann from University Hospitals Augsburg, Germany. Thank you, everyone. It is my pleasure to introduce you Dr. Cadman Leggett. He's a clinical gastroenterologist and assistant professor of medicine at the Mayo Clinic in Rochester, one of my colleagues. He has a very strong academic interest in diagnostic endoscopy with various publications in the fields of advanced imaging and artificial intelligence. He is an active member of the AI and GI leadership group at Mayo, where he leads initiatives focused on the design and implementation of AI solutions in endoscopy. He's going to talk to us today on the role of CADx, and I'll hand it over to him. Thanks, Cadman. Thank you, Naina, for the kind introduction. Welcome, everybody, back from the break. So I've been tasked to talk about computer-aided diagnosis and why CADx, why do we need CADx to seal the deal when it comes to detection and diagnosis. So today, I'll talk a little bit about the ASGE-PV for characterization of diminutive colorectal polyps, which basically drives our need for real-time characterization of polyps. We'll talk a little bit about the barriers to date with regards to implementation of the ASGE-PV and where CADx comes into play with regards to the promises and its own pitfalls. So what is the rationale behind real-time characterization? So we know that diminutive polyps, those polyps that are less than five millimeters in size, are very commonly encountered during colonoscopy, accounting for approximately 60% of all polyps. These polyps rarely harbor any advanced histology, and it's very uncommon for these polyps to actually have any cancer. However, we do need polypectomy and histologic assessment to determine if these polyps are adenomatous or hyperplastic because this is required to establish a surveillance interval. However, this is associated with significant procedural and post-procedural costs, and it does lead to a delay in recommending a surveillance interval because we need to wait for histopathology to return before we can do so. So a shift in the paradigm would be to identify hyperplastic polyps in the rectal sigmoid. These do not carry any malignant potential, and just leave them in situ or leave them behind. This would lead to a reduction in the cost of pathology and a reduction in the risk of associated polypectomy. Also, adenomas, any diminutive adenoma identified across the colon, if identified in real time, could be resected and not retrieved. And in doing so, this would lead to cost savings with regards to histopathology, and it would be time-saving. So the ASGE came up with the following PV for characterization of diminutive colorectal polyps, and we'll just walk through this to have a framework for the discussion on CADx. So when we encounter a diminutive polyp, we first need to know its location, whether it's in the proximal to the sigmoid colon or in the rectal sigmoid. And then we need to identify real-time the potential histology or an optical biopsy. So we do this with different advanced imaging techniques, which we'll discuss in the future slide. And important to this step is the photodocumentation of that polyp. So we identify the polyp as being an adenoma or a hyperplastic polyp. And if it's a polyp in the rectal sigmoid that is adenoma, we perform a resect and discard. If it's a polyp anywhere else in the colon that is an adenoma or hyperplastic polyp, we do resect and discard. If it's a hyperplastic in the rectal sigmoid, we diagnose and leave behind. And in order for a technology to be able to assist in the characterization of diminutive polyps, these are the thresholds that were put in place by ASGE. For the resect and discard strategy, you needed a 90 or greater percent agreement in assigning a postpolypectomy surveillance interval. With the diagnose and leave behind strategy, you need a greater than 90 percent negative predictive value for an adenomatous histology. However, there's a small font to this, and this is where it's important to read the small font. Only when used with high confidence, right? So this is alluding to that the level of confidence is reliant upon the level of training. So you need to be trained in using a technology to be able to do polyp characterization in real time and to do so with high confidence. So the more you use the technology, the more confident you become in doing a diagnosis. And it has to be paired with photo documentation, and this is where medical legal issues come into play, right? We need to take a picture of that polyp, because if we are relying on a real-time assessment, we need to be able to trace that back. So I'll just briefly mention there are multiple advanced imaging technologies that can be used, and these imaging technologies have shown to be better at characterizing polyps in real time over white light endoscopy. There was a subsequent document by the ASGE in 2015 that determined that the PV thresholds that I just mentioned were met by narrowband imaging, which by far is the technology that is most widely used. Here in the States, we used it without optical magnification. In Japan, there are technologies that use optical magnification. However, this document is somewhat dated. I think that the technologies with regard to narrow spectrum technologies has advanced tremendously over the last few years, both in terms of resolution and brightness. The classification system that is used most commonly is the NBI International Colorectal Endoscopic Classification, or the NICE classification. And I won't go into the details, but it's sufficient to say that there's two types, hyperplastic and adenoma. And the purpose of using this classification is to look at the color, the polyp, the vessels, the surface pattern, to try and determine is this polyp a hyperplastic polyp, or is this polyp an adenoma. And some polyps may have features that are a little bit more subtle than others. So great, we have technology that helps us identify real-time histology. We have the ASGE guidelines to tell us to go ahead and use that technology, and that we are able to use it with high confidence to perform resect and discard or leave behind. However, this is not widely used in community GI practices. And the reason why it's not universally supported is that there are medical legal concerns about making a diagnosis. And in truth, there's lack of a financial incentive. Some community practices do pair themselves with a histopathology practice. So there is a financial disincentive to actually not resect and retrieve. There is suboptimal performance. People don't feel comfortable necessarily making an accurate diagnosis with a real-time assessment of an optical biopsy. And there's a survey that determined that 25% of gastroenterologists assessing polyps only reached 90% accuracy, which is the threshold that we're wanting to meet. There's no standardized training when it comes to the use of these technologists, and there's no performance monitoring. So these are other barriers to implementation. And interesting enough, there is somewhat of skepticism from patients with regards to an optical diagnosis. There was a survey that was performed in 981 patients. Only 60.8% supported PV1, 49.2% supported PV2. And that concern for medical legal concern from physicians is actually well-founded. 21.2% of patients would seek financial compensation if a polyp is left behind and leads to cancer. Again, that would be very rare, but there is that concern. So in terms of the three main barriers to implementation, the fear of an incorrect surveillance interval assignment, fears of possible medical legal issues, fear of making an incorrect diagnosis is what's driving the barriers to implementation. And we'll discuss how CADx has the potential to resolve some of those issues. So a computer-aided diagnosis would provide a real-time assessment of that polyp. It would classify that polyp as a hyperplastic polyp or an adenomatous polyp. It's not reliant on the user expertise, at least not as heavily, or the confidence of the diagnosis. It's almost a second opinion real-time. It requires minimal training to use, right? So you identify the polyp, you turn on the CADx system, and it tells you it's a hyperplastic or an adenomatous if it's able to make a prediction. And it can lead to cost-saving strategies by reducing the number of polypectomies and pathology assessment. CADx is expected to standardize the practice of polyp characterization, right? It will be a universally used tool, ideally. But it is still to be determined if it will abate some of the medical legal concerns, right? Who is ultimately responsible? Is it the CADx system? Is it the endoscopist? Is it both? Is it the hospital? It should improve the adoption of the PV strategies among gastroenterologists. We don't know that. It's too early to tell. And the hope is that it would improve the perception of patients to using the PV strategies. And still, that hasn't been studied yet. So regarding cost analysis, I just want to mention this. It's a modeling study that was performed on a previous CADx study using the endo-brain system on 270 patients that had 250 diminutive rectal sigmoid polyps. And out of these, 145 polyps met criteria for diagnosed and leave behind. And that led to an estimated cost reduction here in the states of 10.9%. And from a population standpoint, that would be 85.2 million. So it's a significant cost reduction. This is a table just summarizing the studies to date with regards to CADx systems. And as you can see, there is variability in terms of how these polyps are being classified. And one thing that is salient to me is that CESL serrated polyps are not universally classified in their own category. And in all fairness, the PV document does not state specifically how these polyps should be classified. But it is somewhat unfair to compare them across the board. So take this with a grain of salt, the sensitivity, the specificity, and the accuracy of these systems taken together. This is the range, right? But most of them do perform pretty well. I want to mention two studies that are, I think, of importance, two systems that are being approved or currently approved in Europe. One is the CHANGE study, Characterization Helping the Assessment of Colorectal Neoplasia in Gastrointestinal Endoscopy. That's why it's called the CHANGE study. It uses the third generation GI genius system and implements both the CADI component and the CADx component. And the CADx here, interesting enough, it works only on white light endoscopy. So it does not use any of the narrow spectrum imaging techniques that we previously mentioned. And it predicts whether it's an adenoma versus non-adenoma versus no prediction. And the characterization sequence of the study started with the endoscopist identifying a polyp, letting the CADx algorithm determine a diagnosis for that polyp. Then the endoscopist would use MPI or blue light imaging to determine a diagnosis and provide a level of confidence. And then the endoscopist would have to use the information from the CADx system and from his or her own diagnosis to do a final diagnosis. So let's walk through the study here. There were four experienced endoscopists trained in optical diagnosis, 162 patients, 494 polyps, diminutive polyps were identified, 295 were in the rectosigmoid, 39 of them were adenomas, so they should be resected. And seven of those were actually left in situ. 256 were non-adenomas, and those were correct. 91.8% were correctly identified. So based on the preview thresholds, the system met the threshold for leave in situ with an NPV of 97.6. And in the same way for the resect and discard, a total of 212 polyps out of 544 polyps, so 39% of the polyps were resected and discarded. And again, they met the threshold for the 90% for surveillance intervals, which is shown in the bar graph up in the right corner. So how about experts versus non-experts? So for the change study, the negative predictive value of the CADx system alone was 97.6, as I mentioned. The experts alone also reached this threshold without the system. Same goes for accuracy, they were both equally good. A subsequent study that looked at experts versus non-experts, this was done offline with 513 polyp videos, showed that the system actually performed better in non-expert hands in terms of it led to improvement in their diagnostic capability, as you can see from the ROI curves, the experts versus the non-experts. This is the second study I want to mention, the ABC study or AI for blue light imaging characterization. This is using the CAD-I Fujifilm system. And the aim of the system was to study PP1 and PP2 thresholds. This study actually compared experts to non-experts, and the polyp characterization sequence, interesting enough, is the opposite of the change study. So this sequence started with the endoscopist looking at the polyp, providing a diagnosis and a level of confidence, followed by turning on the CADx system, providing a diagnosis from the CADx system, and then a combined final diagnosis. The system in expert hands met the PV1 threshold as well as the PV2 threshold when it's assisted, so when the endoscopist is assisted by the AI system. It did not meet the threshold when the AI system was working alone. And here's the data with regards to experts and non-experts. So interesting enough, this study didn't really demonstrate that there was a benefit to non-experts like the change study did. And to me, it brings to light maybe the order does matter, right, whether it influences the final diagnosis or not. But the results of this study do highlight the importance of diagnostic experience, right? If you are well-versed in the diagnosis of polyps, you are better served by the AI system than if you lack that experience. So I think there's going to be a conversation about training and AI. I think this is a good example of how training and AI should not substitute training, right? You still need to train in the use of optical diagnosis, even if we're using AI. So interesting enough in this study, as you can see in that curve, the last 50 polyps that were evaluated by non-experts did meet the PV2 thresholds, meaning that there is a learning curve where the AI and the endoscopist, so the endoscopist is learning from the AI system as they evaluate these polyps. I had mentioned serrated polyps. I just want to briefly mention this study that looked at serrated polyps using a different classification than the NICE classification. And it reached 84% prediction of serrated lesions. So it is feasible. What's interesting also about this study is that they did external validation on a different set of images using a different technology, blue light images. So they trained on MBI images and did validation with blue light images. And there was still a little bit of a drop in performance, but still it's good performance. So what are some of the pitfalls with regards to CADx and some of the future directions? Like I mentioned, I don't think that it should be a substitute for endoscopic diagnostic skills or training. We should still be training the future generation of endoscopists with regards to optical biopsy and how to use these different classification systems to identify whether a polyp is metanormal or not. There's that interplay, which is interesting between AI and the endoscopist. And I bring to question, does the order matter, right? If you turn on the system and it tells you a prediction, are you going to go with what the system is telling you, or are you going to let it kind of bias your own decision? Let's not lose sight of the classification that is being used. I showed that table with the different ways that these bonds are being classified. The PV documents say adenoma versus hyperplastic, but there's also serrated lesions. There's also normal mucosa, right? Normal mucosa is being classified as non-adenomatous, but we don't want to be resecting normal mucosa, right? That's a polypectomy that shouldn't have happened. So let's not lose track of how these systems are classifying normal mucosa. I mentioned the training for serrated lesions. I think further training has to be done on these systems to specifically identify serrated lesions. And also an unanswered question is, how do we resolve discrepancies between the endoscopist classification, the CADx, and histology? And I just briefly mentioned a study that looked at 1,169 polyps. 16% of these polyps were rated with high confidence, but did not match the histology. And on re-review or deeper sectioning of the histopathology, 25% of these discrepant cases were resolved. So although we consider histology the gold standard, there is also inter-observer variability with regards to the interpretation of histopathology among pathologists. So that's just something to keep into account. And I think explainable AI is going to be of medical legal importance, right? If we can trace back how the decision was made and that algorithm can explain why a decision was made, then it will I think abate some of the medical legal fears that gastroenterologists currently have. And finally, we don't know what is the perspective of the patient with regards to CADx, and it will be interesting to inquire whether patients are more open to having that interaction between the computer and the endoscopist when it comes to an optical diagnosis. So in summary, we reviewed how CADx is capable of meeting PV threshold for both resect and discard and diagnosable need-behind strategies. It will, and it is, an important aid in diagnostic endoscopy, especially for non-experts. It should not be considered a substitute for training or the diagnostic skills of the endoscopist. I don't believe that it's going to compromise training. It might actually, as I showed in that learning curve, it actually may aid in training endoscopists. It will likely lead to practice standardization and definitely, you know, if we can implement the PV thresholds or the PV strategy, it will definitely lead to cost benefits. And it is to be determined whether CADx will generate sufficient financial incentive to drive adoption in community practice. There are still questions about how is it going to be paid? Is it going to be, is there a disincentive to not remove a polyp? And with that, I'll end my presentation. Thank you.

Augmented Diagnosis

Autodocumentation

Computer-aided Diagnosis

Polyp Characterization

CADx Systems

Colonoscopy