Maryland's Eliot Siegel discusses AI's potential for disruption in medical imaging

Implementing artificial intelligence (AI) and machine learning into clinical practice is a primary topic for debate for healthcare. At this year’s SIIM Conference on Machine Intelligence in Medical Imaging (CMIMI) in San Francisco, industry leaders came together Sept. 9 to discuss the impact and disruption of AI in medical imaging and healthcare.

Moderating the panel was Eliot Siegel, MD, a professor of radiology at the University of Maryland School of Medicine and chief of imaging services at the Veterans Affairs Maryland Health Care System. Health Imaging spoke with Siegel before the panel to learn more about challenges and opportunities related to artificial intelligence (AI) in medical imaging, how AI and machine learning should be incorporated into clinical applications, and what the future holds for AI in imaging.

Health Imaging: How has the University of Maryland School of Medicine and the VA used AI and machine learning in medical imaging and clinical applications?

Eliot Siegel, MD: I think overall we have taken very little advantage with healthcare—not because of the lack of interest but because applications related to deep learning have, for the most part, not been available to us. We are not using convolutional [neural] networks or machine learning, and a lot of the things that we get as far as dashboards and reports are based on relatively simple descriptive statistics rather than utilizing predictive models. Healthcare tends to be 15 to 20 years behind some other areas in the industry, and I think that most of those tools at this point are not available to [those of us] who work in healthcare. Even in 2018, there are disappointingly few real-world applications that are available to me to implement and I'd love to see that change.

How has the lack of having AI and machine-learning technologies at the University of Maryland and the VA impacted patient care?

Patient care, in many ways, is suffering because of the fact that, despite our transition from a paper to digital environment, electronic medical records are still just digital versions of paper. The electronic medical records don't support a type of simple query, much less support predictive analytics or models. Healthcare is so far behind, it’s almost as though when I walk through the doors of the hospital or an outpatient center I'm walking backwards in time to an era where there were far more limited capabilities. With electronic medical records, I can't do so many of the things that I want to do to be able to improve patient care, to be able to improve communication, to be able to improve safety. I think we’re on the cusp of being able to start to do that and certainly machine learning and specifically deep learning offers the potential to be able to take that to the next level.

What is challenging about implementing machine learning and AI into healthcare?

I think part of it is the culture; healthcare systems tend to be very slow to adopt new software.  Another part of it is that a lot of the folks who have been the stewards of information technology within healthcare were traditionally—before the transition to the electronic medical record—more focused on billing and lab results and weren’t really data scientists.

Another part of it is that the larger electronic medical record companies have been focusing more on just getting their software out and running. Look at a lot of the major added value related to not just billing, but real clinical predictive models and analytics, which is where I think we should go next. We’ve all rushed really quickly into an electronic medical record environment in order to be compliant with regulatory requirements. I think now we're trying to figure out how within the constraints of the larger systems—Epic and Cerner, for example—we can add these types of advanced machine learning algorithms to all sorts of different applications.

How should machine learning and AI be implemented to improve imaging analysis, decision support and other non-imaging based healthcare applications effecting workplace efficiency and communication?

What I'd love to see would be a lot of the companies that are spending a lot of time and resources into doing that task to be able to optimize algorithms ability to predict what a radiologist or other experts have said. I also don’t think we have that many people who are currently up to speed in the healthcare environment that understand hospital workflow, hospital efficiencies, clinical operations, etc., and who are also data scientists. I think we need to attract them into healthcare and I think we need to educate them about some of the domain specific challenges.

What challenges are AI and machine learning currently putting on clinical applications now and in the future?

The real disruption is going to come fundamentally in the way that people end up consuming algorithms. In general, we are constrained to pick a particular PACS system from a particular vendor. But with an increasing number of AI software being developed that are really interesting and creative, the potential to be able to consume data with a system or engine that delivers different types of algorithms is going to be super disruptive in radiology in a positive way.

We've got to be really careful with AI to be sure that we test algorithms. AI algorithms that are developed with one patient population may not work particularly well with another patient population. Making sure that software is being developed directly from data and that we're really careful we don't end up having bias in the software that would be deleterious to a subset of our patient population, that's really important.

You also don't want to have overreliance of the software with people who don't understand the limitations of AI. It's also making sure that any of the software that comes out is actually clinically relevant and tested; we need to make sure that the FDA clearance process is one that allows software to be able to get out to people who really need it and use it, and yet also we want to make sure that it's very thoroughly vetted and tested.