This Week Health

Don't forget to subscribe!

April 23: Today on TownHall Lee Milligan, MD, Chief Executive Officer at Asbury HealthTech Partners interviews Larry Sitka, EVP & CSIO Enterprise Imaging and AI at PaxeraHealth. Larry shares his insights on the revolutionary integration of AI in medical imaging, along with its challenges and risks. What does it take to get an AI product through FDA approval? What are the expected benefits of introducing AI into medical imaging? What does Larry perceive as some of the security risks of AI in today’s world? Furthermore, we explore the broader implications of AI in healthcare, particularly in improving diagnostic accuracy and patient outcomes. This episode not only highlights innovative strides in AI but also prompts a deeper discussion on the future of healthcare technology.

Subscribe: This Week Health

Twitter: This Week Health

LinkedIn: Week Health

Donate: Alex’s Lemonade Stand: Foundation for Childhood Cancer

Transcript

This transcription is provided by artificial intelligence. We believe in technology but understand that even the smartest robots can sometimes get speech recognition wrong.

Today on Town Hall

you can't build a house, From the roof or the facade, right? You build the house from the foundation up. Why? Cause you can't change the foundation once you put the house on it. So when I keep saying we got to start and fix and solve the data problem, that's exactly what I'm referring to.

My name is Bill Russell. I'm a former CIO for a 16 hospital system and creator of This Week Health.

Where we are dedicated to transforming healthcare, one connection at a time. Our town hall show is designed to bring insights from practitioners and leaders. on the front lines of healthcare. 📍 Today's episode is sponsored by ARMIS, First Health Advisory, Meditech, Optimum Health IT, and uPerform. Alright, let's jump right into today's episode.

hello, everyone. Thanks for joining us on this episode of Town Hall. My name is Lee Milligan. I'm a former CIO for a multi hospital system in Oregon, and also the former CIO for the largest ambulatory medical imaging company in the United States.

Today, I'm very pleased to be joined by Larry Sipka. Larry is the Executive Vice President and CSIO for Paxera Health. Prior to Paxera, he was at Canon. Larry knows a thing or two about medical imaging, and I'm really excited to have this discussion today. For this episode, we're going to be specifically talking about AI in medical imaging.

Roadblocks, optimizations, aspirations, and Larry is just the right guy for this discussion. But before we dive in, Larry, welcome to the show. And can you tell us a little bit about your background and how you got here? Sure. Thanks. I'm a developer by trade primarily. Coming up through the ranks, I started my medical imaging career actually in 1988 was my first PACS deployment.

So I've been here, I've done development of a couple of PACS systems development of two archival systems, and then I actually built a company. And a VNA solution called Accio Technologies, which is now owned by Hyland. Having left that, the construct of adding AI into medical imaging is just absolutely ripe for AI.

It is so inefficient in what it does and how it works. It's just focusing purely on the optimization side of it. If you think about technology, if you look at every PAC system today, for example, Every PAX system the big boys, right? They all were built and created in the early 90s.

So if I correlate technology with that, you can look at, a Zenith TV, an IBM Selectric typewriter, a Wang Word processor. Well, you don't even talk to your, I mean, you don't change the channel on your TV anymore. You say, hey Google, hey Alexa. So there's a lot of optimization that we can do just to assist the physician.

actually dove right into my first question, which was supposed to be What are the expected benefits of introducing AI into medical imaging?

So I guess I really approach it much differently. Let's start with the problem that we're seeing in AI, that being the business model. There is no money to purchase AI algorithms.

AI has to be democratized and commoditized to run inside of VNA and PEX, right? It belongs as part of that. So the construct or an idea of creating a service, I call it VNAI, which is an inference engine for, it's like a marketplace, an inference engine for running AI and running them in essence on top direct access to the VNA.

If you watch every AI algorithm that exists today, You have to send data to it, and you spend 80 percent of the time sending data either to the cloud or even to on premise devices to get a 10 second response. If we reverse that trend and we move the algorithm to sit on the data, you don't have to, do all of that transmission.

Once it lands, it has access. So rather than have five copies to five AI algorithms, You now have one copy of the data accessed by five AI algorithms. So that's a pretty basic difference. It's different in a market than a marketplace, because the, this in essence is a marketplace that the expectation is you can run vendor AI algorithm, but also the ability to build your own AI algorithms without having to be a data scientist.

So that is the other construct. Our goal. Is to commoditize and democratize AI, making it available to those simply that can't afford it because nobody can and don't have access to it. So for free.

Yeah. So I give you a little bit of background on some of the things that we did at my former employment. So we would do just what you said.

We would package up our imaging data and send it across the globe. Our first foray into this was around long bone imaging x ray. We sent it to a company out of Paris. And then it would take a while and then it would pop out a result for us. The other challenge that you alluded to is the cost associated with this.

These are not cheap. And so a lot of these companies are struggling to know what to do about that cost. Do they pass the cost, down to the consumer, to the patient? Do they eat it as part of an operating expense or some mix of both? I, I know of a couple of companies that are passing it along, but it's a strange message, right?

If you're a patient, you're like, I'm going, I'm gonna get my imaging done, and I can either get it done decent, or I can get it done great if I pay extra. And that's, that's basically a new paradigm for patients, and so far, for the patients I've interacted with, they don't like it.

Yeah mean, I'm a patient.

Right? I pay so much money as it is. The expectation of me paying more for an added service, I scratch my head and I say, that should be part of my wellness plan. Plain and simple going forward. This is the problem, is that we've taken PACS and BNA and they haven't evolved. it's a repetitive cycle.

Again, my first PACS was 88. So in the 90s people bought a PAX, then in 2000 they bought another one, 2010 they bought another one, 2020 they bought another one, and it's the same source code, it's the same code base, right, and they just, they gotta get off this figure 8 track, right, that's causing this, I mean, Albert Einstein has a quote of doing the same thing over and expecting, a difference is insanity, and it is.

And if you look at the costs involved, they're incrementally going up. So we have to stop the construct of, adding something more and taking something more. We need to evolve what we have and take it forward, which is what. We can do so long as it starts way up front, though, because what I'm also watching is PACS systems never have an argument, right, with themselves.

But did you ever try and share data from PACS to another application? It's almost impossible. They can't scale, right, to that level. And they certainly lack interoperability. How proof? Try and migrate a PAX, right? They're, they take forever. They're expensive,

etc.

I did that.

I did that at Asante. It took two years.

Yep. It took two years. I did it

at my last place also, and it took, well, it started before I got there and it was, about a year and a half. Total, because it was so long. I want to go back to a little bit of what some of the stuff you were saying around, around the challenges. So, let's say you're a, you're an AI company and you've got this idea around a particular algorithm for a particular thing what does it take for a company to get an AI product ultimately up through FDA approval?

Like,

what does that look like? it's a long, very not only meticulous, monotonous task, really expensive task, their first challenge is data, they typically don't have access to data. So that's why the construct of linking together, not only a platform for running AI, but a platform for curating AI, and a research side of an organization, and then providing a.

A network that has access to differing cohorts so that the organization, the AI organization can actually run their algorithm against those cohorts. No, I'm sorry. One quick,

one, sorry, just one quick point on that. That's a great point. When I was at RS& A, one of the things that I discovered was that there's a whole market, underground market for anonymized images, and there's anonymized images, brokers.

out there who are in between companies that are generating images and companies that need images in order to run their tests. And it's a whole separate market right now.

Correct. And that's why if you look at the process of how do we monetize imaging that we've acquired, every organization actually has the ability to do this if they have the right service and part of what VNA actually does.

Is A, it normalizes data. B, it does all the labeling volumetrics right up front. So those algorithms don't have to do that. And then C it creates these cohorts, it starts right up front with data normalization, because what we're seeing also is DICOM data specifically and HL7 is really prone to error.

I mean, it is a mess, and garbage in is what garbage out. And that's why data inside these PAC systems, it's really never been clean. Just a few fields primarily. So the first AI algorithms are one that normalizes content. So you have content that's accurate, that's accessible. The second AI algorithm is body part labeling and volumetrics.

Okay. And these are happening under the hood. So now you have all the body parts identified with the volumetric information that's available. Okay. Then you take all the metadata that's involved inside of that and you extract it and save it. When you do this, not only are humans going to access that data, but other AI algorithms will access that data, especially if you move towards an inference engine with AI, right?

So now it's not just one algorithm with one finding. It might be five algorithms that do the same thing, and they're all asking for second opinion. Or two algorithms that do two different things on the same image set with a combined result. So those are the things and what we're going to start to see is AI being called by not only humans.

Right, by the physician base and by, so that's the other option, AI being called by the patient base, right, because they now have access through their EMR, and, to your point on a if there was a business model in doing that, if it cost you a dollar or five dollars, maybe some would do that.

I don't think so. The unwillingness to pay, given the fact we're paying tens of thousands of dollars already, is a huge business hurdle for AI vendors. But if you can make it part of that. of the existing application, there is money, there is a financial benefit in PACS and BNA that's part of the existing establishment.

  📍 📍   In the ever evolving world of health IT, staying updated isn't just an option. It's essential. Welcome to This Week Health, your daily dose of news, podcasts, and expert commentary.

Designed specifically for healthcare professionals like yourself. Discover the future of health IT news with This Week Health. Our new news aggregation process brings you the most relevant, hand picked stories from the world of health IT. Curated by experts, summarized for clarity, and delivered directly to you.

No more sifting through irrelevant news, just pure, focused content to keep you informed and ahead. Don't be left behind. Start your day with insight at the intersection of technology and healthcare. This Week Health. Where information inspires innovation.   📍 Increase

So, so one of the things that came to mind when you're, when you were saying that is, I'm thinking about it from the practical perspective of each health system who over the course of the last 25 years have either stood up or acquired multiple PACS. So they have, they don't have, the health system PACS.

They have a radiology PACS that's current. They have a prior radiology PACS that's still maintaining. They have a cardiology PACS. Sometimes they have a neonatology PACS specific to that particular department. And then you have all these acquisitions. And one of the challenges that, that I experienced at my last job with the ambulatory imaging company was that as we acquired medical imaging centers, which was part of the imaging strategy acquisition around acquisition, the PACS data.

That was there was suspect and we didn't want to mix that data with our main data And so what ended up happening? We ended having to maintain multiple packs

and so ten vertical service lines, right? Yes,

it just gets bigger and bigger and then you know around FTE? Well, of course I need a new FTE If we keep acquiring and get all this, these additional packs, it's crazy.

So the way to solve that issue, right, is that's the sole purpose behind VNA technology. So in other words, you can preserve that logical separation by service line. And it's, I call it, pulling the A out of packs. So use packs as a viewer. I'm an old guy. So the ISO model is really stuck in my head.

So you have your application level. If you can pull the application away from the data management side of it, that's PACS and that's VNA. Yes. That's the way to handle that scenario. Then you can actually migrate all those service lines into one common location. You can preserve the existing PID schemas, session numbers, et cetera, whatever those departments or organizations actually do, and then create a logical enterprise class ID, right.

That sits directly on top of it. So, and that's the way I've always rolled in from VNA perspective, these great big organizations, right. The Sutter Health, the Piedmont's. The United States Department of Defense, those are organizations that I brought together in just through this concept, that wasn't my AccuAdase, that was its sole purpose.

Now you need to take that service and we need to get smart about it. If it's storing this data and managing this data, orchestrating HL7 and DICOM. We need to clean it up, we need to label it, and then make it available because there's a new user. That new user is AI. Everything that we've done and built, the perception of that data has been done by humans.

Now the perception of this data is being accessed through AI. AI is data hungry. There was a site that I worked with about three years ago out of New Orleans Oshter Health. They do about 17, 000 net new studies per day. Inside of those studies, there's about 400 radiologists and cardiologists.

They have two AI algorithms. Each algorithm pulls 20, 000 plus studies. Every single night within two hour time frame. So, 40, 000, 50, 000 studies getting pulled to the cloud out of, you know, the enterprise by just two users, two AI applications. There is no PAX that can run at that rate.

Yeah, I want to go back just for a second to FDA approval process for a moment.

So, in, in addition to needing all of the tests To be able to test it out properly, all the samples you can't put forward a product specific to a body part and a imaging modality only. In other words, you can't just say chest x ray, right? Or head CT. It is a diagnosis by diagnosis approval process.

So for example, if I'm going to, put forward a chest x ray AI product, It will look specifically for pneumonia, pneumothorax, cardiomegaly, whatever it might be specific to that. And I think that is part of the, I think it's the right thing to do, but I also think it's really slow.

And when people get products, they have to realize that it's approved for the following 6, 8, 10, 12 diagnoses. But if it doesn't fall within that, in that category, then you may have to look elsewhere.

And the other problem that it, they've actually created is there's a version 1. 0. And they'll never evolve.

But yet the software itself, if it's really a learning software, right, where it's supposed to get more accurate, is changing every day. So that's why the, again, I'll go back to this, VNAI, an inference engine that runs these AI algorithms that you can pull in from vendors or you create. But you need the ability to actually manage that.

So you need like an IDE environment for curating algorithms. Meaning being able to pull in other data models and other algorithms, being able to generate your own, and do that periodically because you can, having that IDE environment that allows you to get those statistical accuracy measurements is extremely, you know, what is your loss?

Has this drifted over time based on the same cohort that you run? So you have a set that's been pulled through inside VNA. You build and curate your algorithms and vendor algorithms, and you periodically then pull those and run them again to make sure they just haven't drifted too far.

I used to sit on

But quickly, and also you can set for many of these, you can set your sensitivity and your specificity, and it's a two edged sword, right? If you set the sensitivity too high, it's going to over call things. You set it too low, it's going to under call things.

And, it's hard to make doctors happy

and

getting that

right is a challenge. And typically is to build that algorithm requires a Python expert with a Python scripting. What we've done is built an IDE environment that lets you do that through an interface GUI. And then it also lets you connect directly to the EMR to add other ancillary events.

There's a FHIR service directly to the EMR that says, Is this a smoker? A non smoker? A shipbuilder? Coal miner? Are they Northern European? Are they African American? Are they, Asian? Cause there's differences that these AI algorithms, take. When, sat on two AI organizations as an advisor.

And both of those are in the market now. But I remember one that came from South Korea when it was here and it first ran on the U S market. It was horrible and it was horrible because it was trained. on an Asian population set. Yes. But it became very accurate quickly. my take is, let's take a single algorithm, train it across multiple different variations.

And when data comes in, we have the inference engine that's smart enough to know that this person is Northern European. He's overweight, right? He's over 60. He's a non smoker. And then run that out. That's me. I'm again, I've really approached this from a selfless standpoint. It's we're all patients.

Yeah. The older I get. The more important this becomes to me.

One of the, one of the challenges that exists and going back to what we're talking about health systems with acquisitions is that finding an environment that is, is clean enough to be able to add this, let this additional layer of complexity can be a challenge.

I started a company last year called Asbury Health Tech Partners, where we focus almost exclusively on imaging. Project management primarily focused on optimizing the end user experience. In other words, getting rid of latency times, improving really all of the experience around looking at the image.

In addition, it looks at some potential AI options that are specific to that individual company. What I'm finding with the companies I'm interacting with is that around their network specifically, they have this layered approach where it's very kind of, clunky and has been built over years.

And, during the course of the process, folks haven't been able to step back And look at the entire schema and be able to say, okay, for our current needs and our future needs, which include our AI needs, what should our network look like? So we can move this data in a, in an efficient manner throughout the system.

And I feel like that element is a little bit missing in some places have gotten out a little bit ahead of their skis when it comes to some of this AI stuff and their teams are really paying the price right now.

You nailed it. The biggest problem right now. is legacy, right? We're pulling, again, 1990s tech, when we're trying to do things, that scale by VMs, we're trying to do things that have to scale by Kubernetes, you know, those kinds of things, scale by service, in order to do that with AI.

So, you can't build a house, From the roof or the facade, right? You build the house from the foundation up. Why? Cause you can't change the foundation once you put the house on it. And that's the analogy. So when I keep saying we got to start and fix and solve the data problem, that's exactly what I'm referring to.

Part of what VNA services do is normalization, is accuracy, is consistency, is that migration strategy. The other thing is some of the data is so old. When I first started out, DICOM had, the first releases of ACR NEMA 2, right? And the first DICOM 3. 0. They had 30 DICOM tags. Now there's thousands of DICOM tags for every image that comes across.

And most of them are actually private tags generated by the modalities their PAX vendor, right? So GE has theirs, Agfa has theirs, Canon has theirs, and it becomes really hard to interoperate when that happens. So having a standardized base, right? For accuracy first, then find the information, make sure that information is absolutely available, not only to those humans, because what we're finding also is.

Very monotonous tasks. \ these radiologists become bound to a particular PAX because of the tight integration of the keyboard and the mouse. Yes. So in other words, hotkeys. Yes. You have 50 hotkeys. Yes. They right click, left click, drop, whatever. Yep. Why not take a step back and like we talked to the remote controls were starting to get that crazy too.

And finally, Alexa and Google came along and I just say, Hey, Alexa, find me a, find me the NFL game, this weekend, right? Yep. And it does it and finds it for me. Well, why can't we have a radiologist actually say, hey, Aerobot, can you tell me what spine labeling is?

Turn it on. Yes. Yes. Instead of drop down, right click, select, on, and then run it? I mean, seriously. These are things that we can do. Really easy, really simple.

Yeah, I hear you. We're almost at time. A couple more questions and we've got to wrap this up. Number one, what do you perceive some of the security threats are associated with AI in today's world?

This is where I wanna bring AI to the data. That I think sending data to the algorithm. Is a really big security hole. . In other words, especially if you have 50 vendors and you have 50 holes to the cloud, if your data's up in the cloud. Let's just bring those algorithms into, to sit on the inference engine right in your cloud instance.

Yes. If the data is on premise, let's bring the algorithms, as a container that run on a docker that sit directly on top.

Yes.

Yeah. Okay. I'm with you. 100%. It's funny that you asked that question because I was just asked to write a white paper on cyber security and AI. So. I finished the first draft that will be coming out shortly.

I'll share that with you.

Okay. Yeah. Look forward to it. Okay. We got to wrap this up. Any last thoughts for our listeners around AI in medical imaging and where the industry is going?

It's happening. It's going to happen. The business model isn't there and won't be there to support it in healthcare. It wasn't there in the financial industry either. We're going to gain AI by optimization and accuracy. That's the biggest thing that it can actually bring to bear. And as a patient, I'm very selfish.

The human eye can only see so much data. AI sees every bit and byte of data that you can possibly create. These modalities are exceeding what the human eye can see.

As you were saying that, what I was envisioning is some of those pictures from space. Where they use x ray technology to see things that the human eye can't see.

And I think you're right. I think these, um It's no

different. Yeah. It's really no different. Yeah. We have our limits and that's why we create technology to leverage it forward. We created, telescopes and then we created telescopes in the sky and then we created software that runs on the images from the telescopes in the sky.

The same is true internally, right? We used to cut people open to look and see what was wrong. Then we came out with, x rays and digital modalities. And now they're going deeper, deeper, deeper. And we're creating AI that runs in the modalities to enhance the images and then bring them forward.

Well, and the clinical implications obviously are huge. I think about, scenarios, for example, pancreatic cancer, right? This is something that is is traditionally very difficult to diagnose, frequently diagnosed late. For folks. And traditional imaging has done only a so so job of actually picking it up.

Being able to pick it up a year or two earlier, by virtue of having some of these advanced algorithms doing their thing, it would be huge for patients and for society.

Yeah, enterprise imaging is a three legged stool. We used to think of it as, as pixels, right? But it's pixels. It's discrete content, right?

EMR stuff. And it's the, what I call the biology side of the business the genetics, the genomics, the geography, the liquid biopsies, all of this on one common timeline has to be used and leveraged and humans just can't do that. It's just like a stockbroker, right? When I first started out, my career in 81, I had a stockbroker.

He managed, 10 stocks. Now a stock, there is no such thing as a stock broker, right? They're a portfolio manager. They manage tens of thousands of stocks and EFTs and how can they manage that volume and quantity of data? They were the first users of some of the AI algorithms.

Oh yeah. And they still exist. The models. Filtering out the semantics of what the responses are and giving a best guess, accurate, that the human can filter through the whole, that's where healthcare is going.

I totally agree. Well, listen, Larry, this has been awesome. I very much appreciate you carving out time to talk to us at Town Hall really informative information.

I hope there's a part two in the future. And I'm looking forward to that white paper.

Thank you, sir. Appreciate it. Thank you. Bye

bye.

 Thanks for listening to this week's Town Hall. A big thanks to our hosts and content creators. We really couldn't do it without them. We hope that you're going to share this podcast with a peer or a friend. It's a great chance to discuss and even establish a mentoring relationship along the way.

One way you can support the show is to subscribe and leave us a rating. That would be really appreciated. And a big thanks to our partners, Armis, First Health Advisory, Meditech, Optimum Health IT, and uPerform. Check them out at thisweekhealth. com slash partners. Thanks for listening. That's all for now..

Thank You to Our Show Sponsors

Our Shows

Newsday - This Week Health
Keynote - This Week Health2 Minute Drill Drex DeFord This Week Health
Solution Showcase This Week HealthToday in Health IT - This Week Health

Related Content

1 2 3 259
Transform Healthcare - One Connection at a Time

© Copyright 2024 Health Lyrics All rights reserved