הבינה המלאכותית שתציל חיים: מהפכה באבחון רפואי

My name is Idan Basuk and I lead the R and D&AI for ADOC. And I think we are really right now at a really transformational time, not only for AI, but also for AI in healthcare specifically. And I think there are a lot of huge opportunities for applying AI in healthcare. And I want to give what we are doing right now at ADOC as an example for that and discuss both what our AI and products are doing and also how technically we're doing it and making it work. So let me see. Okay, great. But before that, let me speak about the problem that we want to solve. So imagine a relative or a friend of yours going into the hospital. And many of us know that, right? You go in and you run through, through some tests and then maybe you're told that you need to get a scan, like a CT scan. But when you get the scan, you start encountering several very well known and challenging problems of the medical diagnostic process. There are three difficult problems of the diagnostic process. The first of them is triage. When you get scanned, you are put in, you are added into the queue of a radiologist, let's say at the 30th or 50th place in their work list. So it will take you hours, if not days to get even looked at by someone. And in that time you might have a life critical condition that's developing inside your brain or inside your body. So you are really at risk during that time. And this is one problem, the prioritization and triage problem. The second problem is the misdiagnosis problem. Since doctors are flooded with information, it's very easy to misdiagnose. So from what we've seen about without AI, about 20% of the patients are being misdiagnosed, meaning that they can be sent home without the correct diagnosis, even for very life critical conditions. And it's not because doctors are not very qualified. They are extremely qualified and work extremely hard. It's just because the amount of data doctors need to deal with, just as an example, a single CT scan contains 1 billion pixels. And the finding that could risk your life could be 50 pixels out of these 1 billion pixels. So it's a, it's an unbelievable amount of data to process for a human and it's quite literally unbelievable that they are able to do this without AI. And the third problem is the treatment delays. So let's say for medical imaging, even after you get a CT and after a radiologist looks at it, the radiologist only writes a diagnosis reports, they don't actually treat the patient, they don't give the medication or do the surgical procedure. And even if the radiologist gave the right diagnosis, there could be many answers for whether this patient needs a surgery or medical treatment of another kind, like medication. So the problem is that loops don't close fast enough. And this is another problem. And even after you got diagnosed in your image, it could take hours, even 10 hours until the decision is made of what clinical action needs to happen for you. So these are the problems that we are trying to solve. And these are, they could sound like, okay, interesting problems, but is it rare? How rare is it that, you know, radiologist or physician in general misdiagnosed, or that someone with an actually life critical condition waits for hours for his diagnosis or her diagnosis? And from many, many studies that were performed, like this study from Johns Hopkins University, one of the leading medical institutions in the U.S. just in the U.S. alone, about 1 million people every year either die or suffer from irreversible medical damage because of these three problems, because of these three fundamental issues of the clinical diagnostic process. And we believe that in order to solve these problems, you cannot fix the system without putting AI not in the periphery and not around it, but, but in the very most basic sense of how you do medicine today. And that's what we're trying to do. So we try to tackle it from three different angles within our product. Each angle is for a different part of the different part of the three problems that we discussed. So the first angle is the prioritization. So every doctor works with a work list and goes through one patient after the other. So instead of remaining the work list order as it was, if the AI detects a clinical and life risking condition, we bump that patient into the top of the work list. So that's one thing. And from what we've seen, this can literally reduce the amount of time sick patients can wait by up to 95%. Okay, so these are crazy numbers. The second thing is how we impact disease detection. So for example, for radiology, when the radiologist looks at a patient, they open their image, right? So for many of our products, when they open their image and in their native viewer, it's not an additional viewer in their native viewer, we point them to the finding that our system found. And thus we significantly reduce the chances that they will misdiagnose and not look at the finding that is actually present there because of this needle in a haystack. And the third problem of the treatment delays, we said that eventually a radiologist writes a textual report, but it doesn't treat the patient. Right. So the people that do treat the patients, the surgeon, the surgeons, the emergency department doctors and the nurses, they are not on desktop computers all day. They need to see patients go through their beds and talk to them. So we literally developed a mobile application that they can hold in their pockets. So for example, a surgeon that is right now doing surgery can, and it actually happens a lot, receive an alert in our mobile application while they are doing a surgery. Hey, there is another patient here that it looks like, according to the AI, that they have like a clot, meaning a blockage in their lungs, arteries. Do you agree that we need to start preparing the operating room and flying them with a chopper through different side of the country. And with a click of a button, the doctor can decide while he is in his current surgery that the chopper needs to take this patient 100km to his hospital so we can operate on him. And when he finishes the first surgery, the patient is already waiting in the, in the surgery room. So this is critical shortening of the time that patients have to wait. So beyond the numbers and the statistics of 900,000 people a year, I think it's worth looking at, you know, the human and the day to day perspective of patients and doctors that use such systems. So these are two feedbacks, examples of feedbacks that we receive actually on a daily basis. So as you can see in the top, feedback and a text message that we got from a doctor that is using us at Israel. It talks about a girl that underwent a car accident, a very severe, very serious car accident. But the fracture in the spine was so tiny that without adoc, the radiologist would have missed it. Think they would have missed this fracture and sent this girl home with a spine fracture? Okay, imagine if it happens to your daughter or your child or to yourself. And the second example is a patient that came to hospital with actually a stomach ache. And because the suspicion was for a problem in the, in the abdomen, they scanned the abdomen and the pelvis, and just by chance, they scanned a few centimeters of the lungs. Most radiologists would not look at the lungs because this is a scan of the abdomen. With AI, it takes you no time. You just deploy the AI to look at the lung pixels as well. And we found that what actually explained the pain that this patient was experiencing was not a problem in the abdomen, it was a blockage in the lung arteries. And the doctors acted upon that problem. So these are the kind of problems that we see every day and that AI can help with every day. So I just want to give you a little bit of background of where we were in the beginning of this year. So it's not something that we started working on today. You don't just found a company and you here you have a mobile application and desktop application, all the integrations. We were founded about 10 years ago, almost 10 years ago, and today we're installed in almost 2,000 hospitals. So the medical world today is really open to AI. Innovation is really embracing it much more than ever before. And in the beginning of this year, we had about 30 clinical solutions that diagnosed about 30 different problems. And these solutions were, these solutions were based on supervised learning eventually, that we built over the years different supervised learning solutions for each of these 30 clinical conditions. But three years ago, of course, with ChatGPT, we understood that there is a real chance to build an model, foundation model or a frontier model that actually understands the image. And we understood that we are accessible to the amounts of data. Specifically in our case, we have access to about 100 million medical imaging scans, which is an unprecedented amount of data to train such a model on. So we decided to develop our own image encoder. And that was three years ago. And in the beginning of this year, together with this foundation model, it took us from 30 clinical conditions to 45 clinical conditions that we already sold this solution. So within one year, within the last year, we jumped in 50% with the amount of the clinical conditions that our product is covering. But this was actually only the first generation of what this foundation model is capable of. Because the next step, we understood in every scan there are hundreds of different findings that you can look for just in a single scan. And we understood that providing clinicians with the AI that analyzes hundreds of different findings, it's not just an AI problem. How do you even let a clinician consume such a vast amount of AI alerts and data without disrupting their work? Because it's a huge amount of data. So we understood that to take it to the next step, we need to provide the results not through a user interface that points at the findings, but in the native way that the radiologists are used to looking at diagnosis of images, which is the radiology report that they write. We understood that we need to write the radiology report covering hundreds of different findings in the same way that a radiologist would write it. And this is what we did. And for our first model of that kind, we took chest X ray scans, so two dimensional X ray scans of the chest and we mapped more than 100 different findings. And this is nearly all actionable findings that can be found in an X ray scan of the chest. We mapped them to different categories. Problems of the heart, of the lungs, other problems, even orthopedic problems such as shoulder dislocations. We mapped all of these problems and we trained an image encoder and a generative model that can generate end to end radiology reports for these. But it's as I said earlier, it's not just the model, but it's also the workflow of how you let the radiologist use it. So how do I make the movie run here? Do I do it or. Yeah, so this is how it looks like from the radiologist perspective. Using this AI, they open their workflow just as they would always. And when they open their workflow, they see the image and they see the template of the report. It's still an empty report. And what they would do without AI generates the report is just look at the entire image. You can see how much data there is. There are like dozens of organs within the different organs that they need to look at within this image. And then they need to write the entire report. So now, today with our product, they don't just look at the image, but they can look at the image and see that we pointed to all the different findings that we found in this image, like for example, the kidney injury that we pointed to. And then they approach writing the report. So what you. Oh, sorry. What you'll be able to see in a few moments is that they approach a view that enables them to look at the different findings that our AI detected. And so let's say that. And now we're going to see it. Now let's say that they want to populate the report. Let's say that even you have an AI look at how many words, how many different sentences. It's really overloading. It really can disrupt the radiologist workflow more than it can help it. So what we did was literally develop an application that as you can see now, you can review on the right each finding and decide that you don't agree. For example, with the first finding with a spleen injury, all you need to do is cancel it and it's deleted from the report. And you can trust that your report is coherent and consistent and just populate it into your regular reporting software. It's not a side window, it's your regular reporting software. So this is generally the product and how it works and why it's so critical to build things in the medical domain that are so deeply integrated with how the radiologists work usually, and we already have a few dozens of clinicians in the US using this AI in real clinical practice under experimental workflow and according to their feedback, this is performing nearly on an experienced radiologist level. And just want to clarify, it's for more than 100 different findings, many of them are super rare. This is performing on nearly a radiologist level. So this is highly accurate. We call it first read. The FDA also acknowledged it as what we call a breakthrough device, or the FDA calls a breakthrough device, is an acknowledgement by the FDA for products that could have very critical impact on the medical world and their regulatory approval needs to be expedited. So we believe that we'll see this product in the market relatively soon. But what's more important than that, and I think stems from both the FDA acknowledgement and the really good feedbacks and really unprecedented feedbacks that we're getting from our first users, is that we truly believe that AI today for the medical domain, if you have the data and you develop the appropriate workflow around it, is today in the level of accuracy that enables you. And that I believe in one or two years from now we'll look back and understand that today is the day that AI in healthcare reached its cloth code moment. It's the moment in which AI was comprehensive enough to touch almost every everything that a radiologist or a clinician could every finding, every clinical decision that they could encounter in their day to day, but also do it in a way that is accurate enough and native and seamless enough to really accelerate them and not only advise them. So this is, I believe, the moment we are undergoing today in clinical AI and we don't have a lot of time. I just want to touch two technical challenges for how we made it work. So as we can see here, the model itself, I really simplified it to be mindful of the time, but it has several layers, an image encoder layer, of course, we have an LLM which literally writes the report itself, writes the text. But eventually, because the LLM tokenization does not use, is not trained on the image encoder embedding space, we needed to train a translation layer between them. And this is something that we train end to end with the mistakes of the radiology reports. And one thing that I wanted to one thing, one interesting technical challenge that I can show here, and I'll be finishing just a minute after I show it, is let's look at just each of these components that I've shown has very deep technical challenges that are unique to the clinical world. So, for example, the image encoder, which is based on contrastive learning, which I guess many of, you know, contrastive learning was shown to be really good at, you know, images and scenarios like this, in which you have sort of like Facebook, Instagram, like images, and you have a very large and central component in the image, and you have a relatively short caption which talks about this central object. But imagine a CT scan of the brain. The central object is the brain itself. It's the most uninteresting thing you can say about the image that it has a brain. You need to find something of the size of 50 pixels. It's like a minutiae in that image. And that is what you need to find using this method. So in this analogy, what we are looking for is not the kid with the balloon, it's the fly on the balloon, right? So this is what we need to find, and we won't have time to get into it today, but we had to come up with a lot of variations and adaptations of the contrastive learning method to make it tuned to find the subtleties in these images. So thank you so much. It was really fun talking to you and have a great evening.