AI Is Reading 15 Million X-Rays a Year With No Human in the Loop | Prashant Warier, Qure.ai

0:00Is the doctor's office of the future going to be a bunch of these systems and the doctor kind of guides you? The AI gives the doctor a cheat sheet on what he should pay attention to. You go in and get a body scan, the doctor gets a report on what he should pay attention to. Is healthcare moving in that direction? My personal belief is that primary care will be AI in the future, maybe 5 to 10 years from now. We have to be more proactive about diagnostics, and we'll see algorithms play a role in that. I think that's where definitely the world is headed, where diagnostics will happen much earlier through the amount of data that we're generating.

0:37And AI has a big role to play in that journey. So let's start with you introducing yourself to listeners. Hi, my name is Prashant Warrior. I'm the co-founder and CEO of Cure. I've been building AI algorithms for the last 25 years. I did a bachelor's in technology out of, in engineering out of India from one of the IITs, IT Delhi, and went to the U.S. Georgia Tech, did a PhD in operations research. My PhD was optimizing trucking networks for US trucking organizations and spent, I mean, did that, went on to work for SAP, where I was basically doing price optimization for retail. So price optimization, markdown optimization, demand forecasting, a bunch of retail and consumer products problems.

1:32This was all in the US. I came back to India about 14 years ago and set up an advertising technology startup, which was using AI to basically collect consumer behavior from a lot of e-commerce sites in India and then use that to target the right ads for customers. And then I got an exit for that about 10 years ago and started Cure slightly less than 10 years ago, focused on AI in the healthcare space. And yeah, we'll talk about the Cure journey today. Yeah. And so talk about what Cure AI was founded to do and how it has evolved. So when we started, our hypothesis was that image recognition algorithms, I mean, there's AlexNet, that was one of these neural convolutional neural networks that was released in 2012, right?

2:31And our hypothesis was that can we sort of take some of those techniques applied to a lot of radiology images, billions of them, and would that enable us to then identify abnormalities on an X-ray or a CT scan? And that was the hypothesis when we started and of course, I mean, we did that, spent a lot of time collecting data because getting access to anonymized, de-identified customer data, patient data is not easy. And so spent almost the first year just working with hospital systems, especially we started out of India. So working with hospital systems in India, getting de-identified patient data, basically radiology images in the corresponding reports and created a large database, more than a billion and a half images now and that went into training these algorithms and uh the what we sort of focused on initially was detecting abnormalities on chest x-rays now chest x-rays the most common imaging modality more than 1.3 billion chest x-rays taken around the world uh we focused on head cd scans head cd scans are one of those scans uh which require the fastest interpretation because you're looking for bleeds or stroke and you want to be able to treat the patient quickly so speed is important so high volume tasks we looked at things where real-time processing is required and then sort of went on from there built out solution portfolios for musculoskeletal x-rays chest cts and a lot more i mean over the years we evolved into an early detection company so we went beyond just radiology imaging we said radiology imaging or detecting abnormalities on a radiology image is not enough you have to actually uh be participate in the end-to-end diagnostic process from the time a patient comes in uh to all the way to the time they get diagnosed with the disease how can you infuse ai into that patient journey at multiple steps and so we have now ai which runs on emr data which runs on the imaging ai imaging vertical imaging radiology imaging data we have follow-up management tools which ensure that patients are followed up for the next test.

4:41So multiple tools that bring together the patient's diagnostic journey and help patients get diagnosed earlier, detected earlier with conditions like tuberculosis, lung cancer, heart failure, COPD, and so on. Yeah, and the customer for this are healthcare systems. Is that right? It's not a consumer-based product. It is not a consumer-based product. So one of the things that we are very careful about is that everything that we build and sell is a software. We only sell AI software, but it gets classified as a software, as a medical device. So it's a medical device. And when you deploy it at a hospital anywhere in the world, you have to go through a regulatory clearance.

5:29It's an FDA clearance in the US. You have to go through CE marking in Europe and every other country has those clearances. And second thing I think is that these devices sort of are, I mean, you have to show evidence of the efficacy of these devices as well. So you have to publish research around this. So we have been very focused on that. So we have today more than 200 plus publications at Cure. And we have 26 FDA clearances. We have 65 findings that are cleared in Europe. And we are actually cleared for use in more than 105 countries. So it's scaled quite broadly across the world as well. So that's sort of our focus has always been on making sure that we get those clearances.

6:12And the product that we deliver is used by a physician in making the right decision. So it's a decision support tool. And at no place are we directly diagnosing something for a patient. And one place we do, I'll talk about that. There is one place where we are autonomous. And I can talk about that in a bit. Yeah, and the AI under the hood has evolved from supervised learning to integrating generative AI or language models. Can you talk about how the underlying AI has evolved? So our underlying AI stack, I mean, so we integrate language models, vision models. So we have a combined vision plus language model.

7:05And that's trained on, I mean, reports that we have. So we have images, radiology images and the corresponding reports. But also we have, I mean, for a large number of scans, we have annotations where a radiologist has gone in and marked out where the abnormality lies. So a lot of it is supervised. There is some unsupervised part, but it's a mix of both of those that goes into the vision language model, which is then what we use for, which is then what we fine-tune for specific use cases. We'll fine-tune that for detecting pulmonary arterial hypertension, for example, or COPD, or a bunch of other conditions, lung cancer, lung nodule, malignancy risk, and a bunch of other conditions.

7:46Yeah. And you were saying this is high volume. Can you talk about the throughput? And so what happens? A doctor takes an x-ray and then it's uploaded to the system. Can it just walk through the workflow from the doctor's point of view? so what happens is patient comes in takes an x-ray x-ray is automatically uploaded to our system right our patient comes in takes the ct ct is uploaded to our system we process it immediately it takes anywhere from a few seconds to maybe a minute to process those scans and the results are then available as a small widget on the radiologist workflow so as a radiologist reading a scan there will be a small queue at the bottom of the screen where they can see the results of what cure has um cure has processed for that particular scan what the abnormalities that cure has identified and they can click on that and they can go into further details it will show you the nodules that we have picked up you can go into details of those nodules you can um see compared to previous scans there is a previous cd scan you can compare to that so it then it gets into that workflow from the radiologist workstation.

9:06So this is one part, right? Now, what we also allow them to do is that they can add this patient into a lung nodule workup program. So basically, if they suspect that this patient could have a lung cancer risk, which is something that we can provide, they can now add this patient to a lung nodule follow-up program, typically called an IPN, incidental pulmonary nodule program, which many US health systems have. And so in that program, what they do is they actually call this patient back in for a CT and you do a low-dose CT scan to see if the nodule is potentially malignant or not, right? And sometimes what happens is that nodule is too small so then they have to call them back again three months later for another CT to see if the nodule has grown.

9:47If the nodule has grown then it is likely malignant otherwise it's benign. So we manage that process as well. So we have the solution called Q-Track which manages that follow-up. So not only do we interpret the x-rays and the CT scans using AI, we also manage that follow-up program. And the other capabilities that we have, some of the capabilities we have is that as the radiologist is reporting this nodule, for example, all the nodule characteristics are automatically inserted into their report. So they don't have to report it now. It's basically part of the report. And that's a very complex task.

10:21It takes time to measure each nodule, to measure, to identify where it is. And so we do all of that and also identify the characteristics of the nodule, right? We do all of that and we insert that into the report for the radiologist. Now, this is one use case. This is one use case. One of the other use cases that has become very big for us is ex-US. I mean, it's not in the US, but this is around tuberculosis screening, where tuberculosis still affects more than 10 million people across the world every year. So it's still very common, especially in the low and middle income countries. And tuberculosis screening protocols are that you take an X-ray.

11:04If the X-ray is positive, you would do a sputum test, basically a microbiological test of the patient's sputum. Now, the bottleneck for this becomes the fact that the X-ray interpretation, which requires a radiologist. and many countries that we work in, in Africa, for example, rural parts of India, there are not enough radiologists. So there is nobody reading those X-rays. It takes sometimes days, sometimes weeks to get those X-rays read. And in 2021, WHO came out and said that AI and specifically they called out Cure's algorithm can interpret those X-rays autonomously to detect TB on that chest X-ray.

11:44So today the protocol there is that X-ray gets taken, gets automatically uploaded to our so this is again we have to do local processing so we have laptops that are in those sites so gets uploaded to our laptop gets processed within 20 seconds they get to know whether it's positive or not and then if the x-ray is positive then they go to and go into a sputum test and that's how they continue testing and today we do this across maybe about 70-75 countries across more than 15 million patients around the world who go through screening with the cure algorithm yeah and from the doctor's uh point of view uh are there metrics about how this can increase uh patient uh coverage i mean uh how how many the time that it would take without cure ai as opposed to using cure ai to to detect some of these anomalies so you know i think in many ways, I mean, there is studies that we have done to show that cure plus I mean, and you have to do those studies for the FDA clearances where we show that cure algorithm plus the radiologist is way more accurate than the radiologist or more accurate than the radiologist alone.

13:04Right. And we do those studies to show that radiologist operating alone and radiologist radiologist operating with AI sort of what is the difference? And we have those studies. Again, I mean, reporting time, I mean, we have done studies which show that we can significantly improve the lung cancer detection. So almost by more than 60 days, I mean, patient would be diagnosed more than 60 days early, sometimes even more because of cure being part of the diagnostic pathway. And so we do many of these kinds of studies which measure accuracy, measure improvement in the reporting of the radiologist and how much earlier you can detect lung cancer, for example.

13:49there's a CREATE study that I've been told about. What is the CREATE study? And how should U.S. healthcare experts interpret the results? So the CREATE study was done. So I'll sort of talk about what the CREATE study is and the relevance of that, right? And then talk about the results. So the basic challenge with detecting lung cancer is that by lung cancer, about 80 % of lung cancer cases are diagnosed late, right? And the reason for that is, I mean, you have these nodules that grow in the lungs. And I'm not a doctor, so I'm also speaking from talking to people, right? But basically, these nodules, I mean, when they are small, you don't have any symptoms.

14:43So you get symptoms when it's already quite late, right? And that's why you have to do screening. So especially people who are at risk, people who are smokers, heavy smokers, above 50, 55 years of age, depending on where you are, different countries have different rules. But heavy smokers and older people go through screening programs, right? They do annual screening based on low-dose CT scans. Now, there are multiple challenges with that screening. One is that most people don't show up. That if you have 100 people who are eligible for screening, typically about five people show up. So people don't show up because they've, I mean, they feel like they're healthy, why should you go out and do the screening, right?

15:21The second thing is that you look at something like that, the capacity for those CTs does not exist in most places. I mean, in a system like the NHS, for example, in UK, I mean, elective procedures are already, I mean, difficult to schedule, right? And doing Lodo CT scan screening is, I mean, there is not enough capacity. I mean, they are creating that capacity in the NHS, definitely. but when you look at the low and middle income countries, there is not enough capacity at all to do those Lodo CT scans. So one of the things that we sort of looked at is we said there are 1.3 billion people who go through a chest x-ray every year, right?

16:00And on that chest x-ray, you can see nodules. Now, can you identify malignancy risk on those nodules and identify the right patients who should go and get a CT? That was the basic hypothesis, right? So can you sort of do a pre-screening for that screening? and we are not here trying to go out and screen people with x-rays what we're saying is just use routine x-rays all of us take x-rays for routine reasons we take x-rays for annual checkups you have some infectious condition you have a cough you have a cold you do a pre-surgery x-ray i mean n number of reasons that you would do an x-ray and we said on those routine x-rays can we detect nodules and can those nodules then be followed up right now that was the hypothesis now the challenge with with that hypothesis is that out of 100 x-rays about 10 will have nodules so not every nodule can be followed up because if you're sending 10 people for a CT out of 100 there is not enough capacity for that so then we said we have to find the right nodule that needs a follow up so we have to figure out which is a high risk nodule so we said let's figure out a risk score for identifying which nodules need a follow up examination that's how we created the lung nodule malignancy risk score now this is not cleared for use in the US right now this is outside of the US that we are doing this study But we said out of those 10 patients who have a nodule, maybe one has a high risk nodule.

17:18And then the question was, how accurate is that risk score, the lung nodule malignancy risk score? How accurate are we at that risk score? That is what the CREAT study studied to see if that lung nodule malignancy risk score was effective at identifying the right patients for a follow up. right now the key result one of the most important results of this is that out of 100 patients who were identified as high risk nodule high risk by the lung nodule melanin series scored 54 patients were actually positive on a CT so all of them got a CT and 54 were actually positive on a CT which means that it was extremely accurate at identifying the right patients for CT screening and just to give you an example if you do a ct by screening program the number of patients who are positive on that ct will be about two so we are going from two out of hundred to 54 out of hundred with this lung nodule melanin series score and the other thing is what is the false negative rate i mean are we missing any nodules right and there also are we missing any cancer cases and there also we identified that we are more than 95 percent uh positive i mean so sensitivity is more than 95 percent and in fact i mean there are some people who might be positive on a ct but in the create study there was nobody who we recommended as low risk uh who was actually then diagnosed with lung cancer so everybody um uh did not i mean so any anything that was negative was finally negative from a cancer perspective so that was the high level results from the create study which basically is now creating a scoring system to identify the right patients who should get a follow-up ct and based on routine chest x-rays which enables early detection of lung cancer our goal is to detect more than 50 percent of the lung cancer cases early so that 80 percent late 20 percent early you want to make that 50 50 50 percent early 50 percent late and we are seeing that movement happening actually as we're looking at data coming out of multiple sites we are seeing that we are enabling that stage shift and i can't talk today about some of those results because you're going to publish them.

19:24But we are seeing that more and more cases are getting diagnosed early with our solution. Yeah. And this is in the centers that are using Cure, obviously. This is in the centers that are using Cure. This is about maybe about close to 700, 800 centers around the world who are using Cure for a lung cancer product, lung cancer pathway. And how do you, if you're a patient and you get an x-ray, how do you encourage your provider to run it through Cure? Is it, is, is you have to find a provider that is using Cure or, or can your provider, is an easy thing for them to adopt? Isn't it an expensive thing for them to adopt?

20:12it's not it's not incredibly expensive to adopt right but the providers have to adopt this system um the i mean so again i mean we sort of have to go through a deployment process we have to integrate with their so what we do is we collect imaging data but we also collect medical records so we are able to then combine the medical record data with imaging data to make the right prediction so that integration with the amr systems that integration with the months, right? And the providers have to go through that process. So a patient cannot use this. I mean, they can, of course, go to a provider who's using Cure.

20:47They can ask the provider if they're using Cure. But the patient cannot. I mean, right now, we don't have a patient program where they can send us a scan and we can process it. And I don't think it's also legal for us to do it right now. So I think based on FDA regulations, we cannot do it. Yeah. I'm just thinking for, I'm at an age where everyone's getting tests of all my friends. If you want to find somebody with Cure in their practice, is there a directory somewhere that Cure provides? You know, that's a good idea. What I will do is I'll ask my marketing to put up a list of sites that we deployed on our website.

21:28So I think we should do that. Yeah. That's one place where you can look it up. Yeah. Yeah. So this is only one cancer that you're addressing, lung cancer. Can you talk about how many different cancers you've addressed? And is it all through x-ray? Or are you covering other kinds of tests? so we also have a ct solution for lung cancer where the same nodules that i spoke about we can detect them on a ct we can characterize those nodules we can measure their progression measure the size so that's again a solution we have again focused on lung cancer we don't currently do other types of cancers but we have similar solutions for tuberculosis like i spoke about we can detect signs of heart failure on chest x-ray and we can follow that up we detect coronary artery disease from routine chest ct scans so on a chest ct scan you can see the calcification of the coronary artery and that is basically a indicator of coronary artery disease so again when you do these lung cancer screening programs you can actually find signs of coronary artery disease and then those patients can then be followed up for heart disease and so they they can then be tested for that right so we do that then we have a solution which can detect bleeds and multiple abnormalities and stroke on head cd scans and cd angios and then we detect fractures on musculoskeletal x-rays so about 16 different kinds of fractures that we're able to detect now some of these are not clear in the u.s so i mean again just wanted to clarify that some not not all of the capabilities that we have are currently cleared by the FDA.

23:2126 of them are. Yeah. And on the skeletal fractures, I have a friend who's an orthopedic surgeon, and he's talked about certain fractures, one in the wrist, that are extremely hard to diagnose. and I know that there have been groups working on developing a diagnostic tool for that. Is that, do you address sort of esoteric fractures? We do, we do. Yeah. Are you familiar? I mean, can you talk about any of those? So this is something which is still being launched and we are sort of getting the FDA clearance for this solution right now. But detecting fractures, wrist fractures is common, I mean, across different body parts.

24:15And the idea is simple. I mean, if you take a musculoskeletal X-ray, I mean, the algorithm will sort of mark out that fracture for the radiologist to see before they report it. Very similar to how you do not use or other conditions. Yeah. Where is this headed for you? I mean, are you building out, are you covering more and more types of anomalies, either fractures or malignancies or, you know, other kinds of disease? or is your focus on integrating other kinds of AI into the system to make it easier to use from the doctor's point of view? It's both, actually. It's both. I mean, we have an LLM research team that is looking at research on what can you interpret from the patient's medical records, which is mostly text data, report data, pathology reports, right?

25:20What can you interpret from there? Then we have a vision research team, which is looking at vision models or basically vision language models. But then they are looking at what more abnormalities can you detect? And as we sort of are growing, what we're realizing is that the real value is in detecting conditions that are very hard to diagnose, right? I mean, COPD diagnosis is very poor. or lung cancer, like I said, early diagnosis is bad, right? So those are areas where AI can play a very important role. And so sort of we are expanding our capability there. But one thing that we are very clear about is that we focus on diagnosis of a disease, right?

26:01And the biomarkers for that can come from pathology tests, can come from radiology, can come from medical records, right? And we also have a tool which basically scribes doctor-patient conversation. so we have a note-taking tool as well so again that allows us to get access to some some medical record data but finally all kinds of data have to then flow into helping make that diagnosis faster and and that's that's the focus so we we sort of do end-to-end diagnostics for a disease and that's that's what we have been doing for the last two and a half three years before that we were more of a before that we were just doing what you were asking for that asking about right that can you do more and more findings on a scan?

26:44And we were doing more, but then we realized that I think doing more findings on a scan is good, but then we sort of have to do the, make sure that we are enhancing that patient journey and finding that patient earlier, diagnosing that patient earlier. And that's where the value gets created. Yeah. I have to ask because, you know, early detection of lung cancer is so important. The other one that requires early detection and is notoriously asymptomatic until it's too late is pancreatic cancer. Have you worked at all on pancreatic cancer detection? And if not, is there a particular issue that makes that difficult?

27:35We have not, actually. We have not looked at pancreatic cancer. But we are looking at more types of cancers and seeing where we can sort of invest into and build those technologies. We have a team of about 200 data scientists and engineers who are working on research in many of these areas. But right now, nothing on pancreatic cancer. Yeah, yeah. This is fascinating. Do you think the day will come when all medical records, all tests, whether it be x-ray, CT, or MRI, or whatever it might be, will be run through something like your AI to help doctors with their diagnosis? I think that, yeah, we are not too far away.

28:35Probably we are already there or getting there very soon. I think AI can add tremendous value to diagnostics. I think overall, I mean, we collect so much data about the body, right? We have sensors, I mean, heart rate sensors. we have got BP sensors, we've got watches, rings, I mean, and there is so much medical record data. So I think in general, the diagnostics has to move earlier and earlier. So we have to be more proactive about diagnostics. And we'll see algorithms play a role in that. I mean, how can algorithms ingest all the data about the body that we are producing and then sort of come to a diagnosis very early so that we can treat, Maybe you don't even require medication.

29:21Maybe it can be treated by lifestyle changes, right? So I think that's where definitely the world is headed, where diagnostics will happen much earlier through the amount of data that we are generating. And AI has a big role to play in that journey. Yeah. I mean, currently, is your rollout primarily in countries that lack provision, that have a shortage of doctors, or are you targeting the more affluent countries like the United States where there's plenty of doctors?

30:08So for 90 % of our portfolio, the market is US and some European countries. There is a small part of a portfolio, which is what I spoke about, tuberculosis, which is more prevalent, I mean, in the low and middle income countries. That product, again, has its own market. It's scaled up across Africa, Asia, LATAM. But the rest of the product portfolio is more developed country focused. Yeah. Yeah. And so in the U.S., what are the barriers to getting this into every medical practice? I think getting it into every practice is just about being there, I mean, being in front of them, presenting this portfolio to them.

31:00What we are seeing is that most times when we are in front of the right stakeholders in a hospital system, they see the value in our lung cancer portfolio. and we are seeing very high win rates from a deal perspective. So if we are in front of them, so again, we sort of started our US business only about a year and a half ago. So we are still scaling up. We are still sort of scaling up across the country. But one of the things we are seeing is that everybody that we speak to is interested and they want to deploy the solution. So it's a matter of time, I think, in the next one or two years, they should be scaled up across most of the US systems.

31:36Yeah. You know, one of the frustrations, you mentioned NHS, but one of the frustrations in the U.S. as well is getting appointments and getting tests scheduled or getting insurance to cover tests. And there's sort of the high end, this trend of testing, sort of elected testing that people can pay for directly. Why not make this available to the consumer? because, yeah, the consumer can go to their doctor and say, hey, you should get Cure AI because it'll give you an earlier and more accurate detection of anything that's going on with me. But it would be great if I could take my medical records and scans, upload them to a system that was trustworthy, invented, as yours is, and get flagged that you need to see a doctor, there are nodules in your lung that should be taken, should have a professional look at.

33:03Why does it have to go through the medical establishment? I think where we operate, I mean, especially with scans, right? I mean, people don't have access to their scans. I mean, if you have taken a chest CT scan, you may not have access to that, right? And getting access to those scans is hard for people, for patients. And then if they have to give it to us, we have to process it. Of course, if they have it, we can process it. But then there is the regulations also, FDA regulations, which don't allow us to directly interact with a patient or provide a diagnosis to a patient. So I think multiple steps, I think there is definitely tremendous value in making this available to patients, right?

33:45And I think if you look at chat GPT for health, I mean, they launched a direct patient facing healthcare tool. Again, they're very careful about the disclaimers there because they cannot, I mean, they cannot provide a diagnosis to a patient. But yeah, I mean, I think there is value in providing these healthcare tools to patients. And I think we'll see that happen. Right now, I think one of the bottlenecks is that patients may not have access to their imaging information. And so we have to process that. We have to take it from the provider where they took that X-ray or CT rather than from the patient.

34:22Yeah. Yeah. It just seems that with all of this, the trend is to put more and more power in the patient's hands or more and more knowledge in the patient's hands for them to direct their care. there is one more point there actually which i want to talk about i think the point that you made about knowledge right that is also a difficult one because if i go out and tell a patient that you have a nodule which could be potentially lung cancer that communication has to be done by a human by a doctor right because they communicate in the right way that this could be lung cancer but there is 90 chance that it's not lung cancer right but you should go to a CT.

35:07When that patient gets that information from an AI algorithm, that could scare them. I mean, and that, again, I think the way this information is communicated, how the patients react, I think there is more to it than meets the eye. So, yeah, I think there is more research that needs to be done to make these algorithms patient-facing, but definitely patients should have the power to get a second opinion from AI. I think, And for sure, that should be available to them. I mean, I had an interview recently with a guy who had a skin cancer scare. And he's developed a system that scans. It's a hardware.

35:53It's a scanning room or portable scanning room that a doctor's office could set up. And you go in and it does a 360, very high definition imaging. And then that's fed through an AI system that identifies potential cancerous lesions. And is the doctor's office of the future going to be a bunch of these systems? And the doctor kind of guides you. You do this x-ray. It's the AI, you know, gives the doctor a cheat sheet on what he should pay attention to. You go in and get a body scan. The doctor gets a report on what he should pay attention to. I mean, is health care moving in that direction? My personal belief is that primary care will be AI in the future, maybe five to 10 years from now.