The next section is entitled Practice Steps for Building an AI Healthcare Application in Your Program. So the last two sessions are more on practical tips and tricks of the trade as to how you get started with this, you know, if you're interested in doing AI research into your AI program or also in clinical practice. So with that, Yuichi, I'll turn it over to you for endoscopic imaging and the data that we gather with that. So thank you again, Yuichi, and welcome. Hello again, everybody. I think we have already learned the basics of AI in endoscopy. Therefore it's a really nice time for us to join or dive into the more practical steps, such as the implementation of AI in our GI practice. Here I'm presenting the three options, including the clinical practice usage, clinical research, or AI development for commercialization. I guess each participant has a different view on these topics, but I like to pick up these three topics in an equal way. Of course, clinical practice is in the center of our practice, because we are gastroenterologists. But if you want to use the AI device in clinical practice, you should confirm that the device has already cleared the regulatory approval. And in this slide, it lists the current status of the regulatory approval of AI in endoscopy. As you can see, there are roughly 12 devices which have already cleared the regulatory control, and among them in the States, only one AI device is available on the market, namely G.I. Genius, provided by Medtronic Corporation. But if you live in European countries, you have much more options, including the Olympus device or Fujifilm's AI device. But so basically it depends on where you live or where you do your practice. But why do we need a regulatory clearance when we use the AI devices? The answer is very simple, because it has a risk in clinical practice. So, for example, if you use the CAT-E device, it will provide a lot of the false positive findings, which may disrupt your concentration during endoscopy. This might be a risk. So the accurate assessment between the benefits and the harms of medical device is mandatory. And therefore, we need a regulatory clearance for most of the AI devices in endoscopy. Let us proceed to the clinical research, which is a really interesting topic to my understanding. If you look at the research field of AI in endoscopy, it is exploding in terms of the number of publications and also in terms of the variety of the researches. If you look at the left, the number of publications has got a big leap between 2016 and 2020. And of course, it contains a variety of research fields, including the identification of dysplasia in birth asparagus, as well as cancer detection or detection of gastric cancer. Even we have this scoring system for ulcerative colitis with use of AI. And of course, there are a lot of the AI devices devoted to capsule endoscopy. And nowadays, some researchers are working on the research on ERCP and EUS under the use of AI technologies. However, I would say that the AI for colonoscopy will play a really important role in this field because of the quantity of publications and also the quality of publication. And once you determine the field of the research, you should dive into more detail because you have the three options with regard to the areas of interest. One is a computer-aided detection. The other two includes the computer-aided diagnosis and computer-aided quality assurance. So there are a lot of the research fields which are waiting for you. But once you are determined with the research topic, next thing you should do is the construction of the research consortium and distribute the work packages to the appropriate persons. And there are roughly three categories in the research. One is collecting material for machine learning, and the remaining two includes the algorithm development and doing the clinical validation. And from my understanding, gastroenterologists should be engaged deeply in the first and third topics, namely collecting material and clinical validation. However, when it comes to the algorithm development, from my personal understanding, I would ask for help to the AI experts in this field. Of course, you can do it on your own, but I think the collaboration with the expert would be much appreciated. And of course, you can construct a nice research consortium. From my understanding, the collaboration between gastroenterologists, researchers in the computer vision, and industries is the optimal option. So once you construct the research team and start doing research, three things should be mentioned. One is the collecting learning material. The remaining two include the informed consent process and accurate labeling. And when it comes to collecting learning material, there are roughly three ways to collect the material, including the retrospective database, prospective database, and use of the other database. Of course, doing the research based on retrospectively collected data would be the easiest way. However, this may include a lot of the selection bias, and I think this kind of selection bias will contribute to our overfitting of the data or a model. Therefore, we would prefer using the prospective database, which will reflect the real world distribution of visions. However, of course, it will take a lot of time. So instead of using the prospective database, maybe you can use the open database, which is publicized in the internet, and you can use it very quickly under the agreement with the providers. And also, with use of this open database, you can compare the performance of your AI device with the others. Let me pick up some examples. So here I'm presenting the four kinds of database, including the CVC, ASU, Mayo, Kvasir, and SUN database. Among them, SUN database provides a huge number of the images, including 100 polyps. Actually, this database was provided by our research team. Therefore, I'd like to explain some more details about this database. So SUN database, or Showa University, Nagoya University database, provides a lot of material for free. If you request it. The strengths include the huge data set, high quality labeling, and also it includes the clinical data. And also the quality of the data is assured because we use it for a regulatory clearance. During the data collection process, we should take care of the informed content as well. But here you have also some options, including the non-necessity or opt-out opt-in processes. But when it comes to the usage of the open access data, I think you don't have to get informed content because it has already been acquired by data providers. And when it comes to the non-commercial retrospective studies, I think getting the informed content in an opt-out way would be sufficient. However, when it comes to the regulatory related studies or randomized control trials, of course, we do need opt-in based full informed content processes. And finally, labeling is a very important topic to be addressed. Here we have a lot of questions. For example, who should label the images? Should we include experts or non-experts? Should we include one or more experts? That's a question to be addressed. And also we have got a lot of questions with regard to the gold standard of the labeling. Should we use the pathology-based data or can we use the optical biopsy-based data? I'd like to present some examples to explain and clarify this point. Here you can find one endoscopy picture where you can find some areas that are suspicious for a polyp. However, nobody truly knows if it is a polyp because this area was not removed. So that's why the inclusion of the multiple experts is strongly recommended to construct a labeled database. And the next example is much more interesting because the polyp in the image appears neoplastic. However, a pathological report shows that it was a hyperplastic polyp. But this is really troublesome. Therefore, I think the adoption of the optical biopsy-based result might be considered when we construct the database. And when it comes to the clinical validation, we should take it in mind that there are large number of retrospective studies while a very limited number of prospective trials. Of course, doing a prospective trials or a randomized control trials is preferred. However, in consideration of the huge amount of the workload to collect the data to do a clinical trial, I think doing the kind of benchmark test may be acceptable in the early stage. But once you determine to do a prospective trial, you should take care of two things. One is how to report the study result. And the other thing is how to construct the protocol in compliance with the requirements of the AI research. And during this process, you may refer to two guidelines. One is a spirit AI, which is a guideline for the protocol provision. The other one is a consort AI extension, which is the reporting guideline for RCT, which includes the AI as the intervention R. So finally, some people may have a very strong interest in commercialization of AI device in endoscopy. Actually, I have some experience in this area, so that I'd like to share some of my personal experience. So firstly, it's my personal comment, but I'd like to recommend the three things. First, before you start anything, you need to have some experience. I would recommend that you contact the experts with regulatory clearance, technology transfer, personal information protection, and research ethics. It is really important to avoid a failure of the commercialization. And the next thing is doing pilot studies on your own to get a nice result. Otherwise, everything is going to be failed. And the next step is to get a nice collaboration with the industrial partners, or even you can launch your own startups. And finally, you should secure budget, because getting a regulatory clearance takes a lot of money. And this includes the getting the funding from the public funding body, getting some investment from the investors, or you can use some kind of in-house budget. And lastly, you should secure the budget of the large corporations. So here, I'm presenting the four AI devices that have cleared the regulatory approval in Japan. And actually, I was involved in these four projects. And fortunately, we have cleared the regulatory approval between 2018 and 2020 for these four items. And some of the experience will be shared in this presentation. So first thing we did is to construct a nice research team, including the medical experts, technology experts, and industrial partners. And also, we asked the Olympus Corporation to provide instrumental support. And also, we have secured a really big amount of the research funding, most of which are used for regulatory approval process. There, here, you can find a timeline to get the regulatory approval. This is just an example from my experience on end-brain series. Actually, everything regarding the technology part was completed by the end of 2015. However, it took about three more years before getting regulatory approval. And most of time was consumed for consultation with the regulatory body, which is the PMDA in Japan. And after getting the consent from the regulatory body, we could proceed to the benchmark test, which took around six months. And after that, our application was assessed or reviewed by the regulatory body for around eight months. Then we got a regulatory approval. So in total, we took three years, which is a very long way. And also, we need money for regulatory clearance. Actually, the off-shelf fee for regulatory assessment is not that expensive. It was around $80,000 in Japan. However, when it comes to the trial costs or the cost for a benchmark test, it requires a lot. Usually, it is above $1 million if you do a prospective evaluation of the item in compliance with the regulatory clearance. So securing money is a priority. Actually, we are very lucky that we could collect this money from the public research funding. However, of course, we have the other options, including the investment or the use of the in-house budget. So ladies and gentlemen, thank you very much for your attention. Thank you.

AI healthcare application

endoscopy

implementation options

regulatory clearance

clinical validation