Radiology Partners’ Nina Kottler, M.D., on AI’s Evolving Impact

Four years ago, I interviewed Nina Kottler, M.D., associate chief medical officer for clinical AI at Radiology Partners, the largest radiology practice in the country, with more than 3,600 radiologists. Last week, I spoke with her again, and it was fascinating to hear about the advances that have been made in AI in the interim. 

When I spoke to Kottler in 2021, she described two main ways that AI was impacting the radiologist workflow: (1) Looking at the EHR and summarizing information about the patient to provide a more robust history that radiologists hadn’t had before that; and (2) looking at an image and finding some kind of pathology in order to alert the radiologist or move the study higher on the work list. 

I asked Kottler whether these are still the main impacts or if there are new and significant ways AI is improving radiology. She said that the summarization is an easier task now than it was then, because of foundation models. Instead of using natural language processing, which requires the labeling of data and all kinds of work to get it to understand, now radiologists are using foundation models that automatically have that understanding. “It can extract information, it can summarize it, it can translate it for different audiences. It's actually far more robust now,” she said. 

In terms of finding pathology and moving the study higher on the list, that is still something that's available, but the market has changed, Kottler said. 

When Radiology Partners first started with AI in radiology in 2016, it began with detection and notification, which enabled radiologists to find more pathology than they did before. But she said the problems of today are different. “Now we not only need to at least maintain or improve quality, but more importantly, we need to improve the efficiency and workflow of physicians, because there are not enough radiologists to do the amount of imaging that is coming in. That is such a massive problem. There are hospitals closing down because their radiology departments or radiology practices that they're contracted with don't have enough staff to support them. It is such a huge gap right now because the volume of imaging has been growing at such a rate that we can't keep up. So we need another solution, and it's perfect timing for this tsunami of generative AI and foundation models.”

Radiology Partners has created its own “operating system” for radiologists called MosaicOS that it is rolling out to its affiliated clinicians as well as making it available to health systems and practices. 

One of the issues the company hopes to address is the fragmented use of technology by radiologists. I asked Kottler to describe this fragmentation, and she responded by describing her own personal workspace. 

“If you were here and were looking at my workstation, you would see that I have four screens,” she said. “Part of it is a viewer where I get to look at the images. Part of it is a work list that tells me what study is coming up next. Part of it is my dictation system or voice recognition. And those are superficially integrated because they're from different vendors,” she explained. “Tacked onto those are four different AI systems. And if I want to go to the EMR, that's a separate thing, and our chat component is separate. So I have about 10 or more components that I am forced to toggle between, because the integration is very superficial. We recognized that that just wasn't going to work if the problem is efficiency. If we've got to manage that, we have to completely rethink that system. The PACs archive is part of that. We have to rethink how all of that works together, because if you were creating that kind of technology from scratch today, instead of in the 1990s when it was created, you wouldn't do it the same way. It would all be integrated and have a central brain that has the data and the knowledge layer.”

She described Mosaic as a way to take all of those components and bring them together from first principles. “We don't have legacy tech that we own,” she said. “We can create something from scratch that's totally novel without supporting the old systems.”

I mentioned that it seems like a lot of health systems feel kind of wedded to their legacy vendors and their solutions. 

Kottler agreed that the turnover time for these contracts is quite long — about seven years. But she noted that in comparison to the EHR, it's a very small component of their budget. But most of the time, they will say “I just don't want to think about it. Let's keep it as long as it's working,” she said. 

She said that Mosaic is built so that it can be an overlay, so they don't need to get rid of their PACS system. “We can make upgrades to the system because it's cloud-native and AI-native. So all of these new innovations that are coming out, we can put through very quickly, so even though they still have their legacy system, we can still work efficiently on a very new-age system. I’s kind of the best of both worlds. If they're at a point where they want to get rid of their PACS, they can, and that will save money, but it's not a requirement.”

I asked her about the plan to roll this out across the Radiology Partners network of affiliated practices over the next year or two, and then go out to the broader market of health systems.

She said those things are likely to happen more simultaneously. “This problem is so acute we have to do things in parallel. The market need this more acutely, so we can do it together. We are rolling out segments of our practice one at a time. We call it an operating system just like your phone has an operating system, and you could have apps on your phone that you could add at any time. It can be modular like that. So we created the operating system infrastructure — that’s the hard part. We are adding a couple of apps at a time.”

Radiology Partners, which recently announced it is moving its headquarters from California to Nashville, is creating a new technology services division to roll Mosaic out. “We've always been a tech-enabled radiology practice, but being a division that then sells technology is new and we want to make sure that that's something that we can do well,” Kottler explained. 

Kottler closed by saying she is concerned that so many hospitals are using AI very superficially. “I think part of the reason why is not the technology. What’s really important about it is the integration of that technology into their systems and all of the change management that has to go into educating the end users to make sure they have know how to use it.”

She said if you see papers that talk about AI being better than the radiologist plus AI, that is just because the training hasn't been done. “The validation of the models and the continuous monitoring — all of these things are actually quite difficult to do, and they're totally separate from the tech itself. So the tech layer, while amazing, is actually only a small part of the problem, and what you need is a more holistic solution. This system makes it easy for us to do that.”

I look forward to interviewing her again in 2029 about how AI has evolved even further.