OR WAIT null SECS
Dr. Eliot Siegel discusses the role these tools play, as well as the challenges radiologists could encounter with using them.
As the volume of patients infected with COVID-19 continues to rise nationwide, so has the need for rapid patient management capabilities. To answer that call, many artificial intelligence (AI) tools manufacturers have taken steps to pivot their existing tools toward monitoring disease progression and severity.
But, what role can these technologies legitimately play, and what challenges can radiologists anticipate if they choose to use them? Diagnostic Imaging spoke with Eliot Siegel, M.D., professor and vice chair of radiology at the University of Maryland School of Medicine, to determine whether these tools could be useful during the pandemic.
Whitney Palmer: We are joined today by Dr. Eliot Siegel, professor and vice chair of radiology at the University of Maryland School of Medicine. He is also vice chair of research information systems for the Department of diagnostic radiology and nuclear medicine. He is here with us today to discuss the role of AI diagnostic imaging tools in the COVID-19 pandemic. Thank you, Dr. Siegel, for being with us today.
Eliot Siegel, M.D.: Sure, thanks for the invitation. It should be fun.
Palmer: Wonderful. Well, if we can start out what is the role of radiology AI tools in this pandemic?
Siegel: Well, this pandemic has created a large number of challenges. And, one of the major challenges is in respiratory illnesses. One of the opportunities is to try to determine whether or not one might be able to develop tools that would help in diagnosis and tracking, looking at the trajectory of disease, and particularly in follow ups. So, one of the things that's really important to consider is the fact that we're going to be experiencing COVID-19 in multiple phases. Right now, pretty much whether you're in New York or some of the hotspots or whether you're in a hotspot-to-be, it's really likely that you're going to be seeing large numbers of patients with respiratory illnesses that will be presenting and, in general, the reality is that they will be related to COVID-19. So, at least in the United States, diagnosis at this kind of first phase or wave of disease is probably a relative moot point.
The other thing that's really important is that the majority of studies will be done on portable chest radiography often outside the radiology department – either in the emergency department or in the ICU to keep patients out of the radiology department. That keeps them pretty much where they are in the hospital. At this point, making the diagnosis is really something that is largely and virtually entirely clinical.
So, AI tools to diagnose COVID-19 are probably a minimal value at this point. As time goes on, though, what we're going to be seeing are an increasing number of patients who have severe respiratory disease, and we're going to want to quantify the amount of disease that they have, and also follow and track those patients -- the ones that actually get admitted into the hospital. I think there will be a significant subset of them where we're going to want to quantify the amount of disease. Then, what's going to happen, eventually, is as this begins to pass, we are seeing an increasing number of reports of more permanent or long-lasting disease or pulmonary dysfunction or pulmonary disease associated with the complications of COVID-19.
So, to me, what is probably going to be the most important AI tool will be something that allows one to be able to track the progress of disease to quantify the amount of disease and, then, using AI tools to be able to quantify the amount of parenchymal lung damage that has been done. Those are a variety of different things. We're also seeing on patients who are coming in with multi-organ damage and, for example, renal disease. According to colleagues that I have in New York City, a significant number of patients are coming for CT scans of the brain to evaluate for mental status changes.
As far as AI tools, one of the really key and interesting questions is to what extent do we take tools that exist already and adapt them to allow us to be able to quantify the degree of disease and trajectory of change predominantly in the lungs, but also in the kidneys, in the brain, and other areas? And, to what extent do we, potentially, look at developing new tools? The other thing that's really interesting is that if you look at parenchymal lung manifestations of disease, particularly on CT, we're seeing some unique things associated with COVID-19. For example, peripheral ground glass infiltrates, reverse halo sign. We’re seeing something called crazy paving. So, some of those signs and symptoms may allow us to more confidently distinguish between respiratory symptoms associated with COVID-19 and other respiratory illnesses. This is proved fairly useful so far and Iran. I think it may be that every country has a different experience. But, in the United States, one of the really fascinating questions is, to what extent can we rapidly develop AI tools that might be specific for COVID-19? And, can we adapt really quickly to some of those challenges?
Palmer: All right, well, given that we're talking about kind of pivoting and using existing tools, what are some of the things that we need to consider in talking about using those tools? What kind of challenges exist or what needs to be done to make what we have now is usable?
Siegel: We have a number of different companies with different offerings. So, we have company that has a patent, for example, on detection of ground glass lesions, and one of the questions is to what extent could that be adopted in such a way to be able to be a more general tool for finding ground glass lesions? Will that software allow us to be able to see other manifestations that may not be ground glass associated with COVID-19? Or, do you use software that was developed for tuberculosis that essentially is trained on a different pattern of disease, or software that was developed for diagnosis of pneumonia?
There have been a number of reports recently on using that sort of software and repurposing it for COVID-19. I think what we're going to find is that each one of those are able to make individual findings, but what we need is a database of patients to determine if we're making a diagnosis and differentiating it from other respiratory illnesses. What's specific about COVID-19? What is the pattern? And, people have talked about more than one strain of COVID-19. Does COVID-19 in China and Iran and in Europe differ from the manifestations that we're seeing in the United States? Another really interesting question is the potential to be able to do a triage where one might determine which patients might be at greater risk.
For more severe complications, we don't have a lot of data. I have a data set that I'm getting from colleagues in the Middle East that is stratified according to how the patient initially presented on CT, and how the patient ended up doing ultimately as far as their course. So, being able to look at some of those complex predictions may require multiple different variables. Just finding ground glass lesions or just finding lung nodules or just looking at the pattern of pneumonia or just looking at other findings may not be enough to be able to be specific for COVID-19.
I think what we need to do is use a mechanism of transformation learning where we take software that's been developed for something close like ground glass lesions or like pneumonia, or potentially even lung nodules, and create a way to use the knowledge in those algorithms. Then, further train those to be able to recognize additional findings and to be able to use a priori information. For example, in an environment where COVID-19 is really unusual, maybe before a wave of disease. Then, a certain series of findings may be relatively unlikely to represent COVID-19. Somebody where there's a large surge of COVID-19, most of the findings that one might see from a CT perspective may likely be associated with COVID-19. So, you need to know the Basian a priori probability of disease. Determining that really is based on how many patients in your city or in your hospital are positive.
So, what is the relative prevalence of disease? It may be predicated on the symptoms that a patient is having. And, it may even be predicated on the lab testing. PCR testing, for example, has been suggested to have a significant number of false negatives, maybe a 70-percent sensitivity, maybe 90-to-95-percent specificity. So, it's not definitive, even when one has sort of the gold standard lab test, and being able to take multiple different variables and put them together and, then, come up with a probability of disease and, potentially, the ability to stratify patients may be really important.
Overall, I'm not suggesting that we do AI currently and have a large number of thoracic CT studies that are performed in the United States because I think that we don't need those in general right now with the surge. We want to keep patients out of radiology. Thoracic CT probably has a minimal impact with regard to diagnosis. And, potentially, even just radiography may have limited impact. But as we get farther and farther, as we start looking at second and potentially third waves in the United States, as prevalence ends up changing to significantly lower prevalence, I think these tools may be really important.
I think also that the complications of the disease, which is something that, you know, we're only beginning to see in some of the patients from China, and following up on those, we'll be able to use AI tools. I think the major role is going to be in quantifying disease, having a trajectory of change of disease and, then, looking at complications. I think it's really exciting in a manner analogous to the way the virus itself changes and mutates and is able to adapt as time goes on.
The potential to be able to take AI software and algorithms and be able to potentially modify those in ways that would be most effective for diagnosis are really important. The typical cycle for turnaround for development of software may be many months or years, so the potential to be able to develop something within weeks is really exciting. And, hopefully, as time goes on, we'll be able to rethink some of the regulatory issues associated with the use of some of these algorithms. In many ways we're relaxing some of the other policy and regulatory issues in general, to be able to deal with the surge of the virus. And, it may be really interesting to see whether or not we might be able to use these algorithms, potentially, without FDA clearance or to get a rapid or tentative clearance. Those are all unknowns.
Ultimately, I think we're going to learn a lot of lessons from COVID-19 with regard to software. Questions about can we rapidly develop and deploy software? Can we get around some of the regulatory considerations? And, to what extent can we, in this new environment, that we're moving forward with where people will be much more aware of pandemics and much more aware of the potential to be able to need to develop software and deploy it rapidly? I think we're going to learn a lot of lessons from that. And, there are many efforts being published already. Many research scientists, and now I'm involved in a couple of computer science seminars that are happening virtually, where they're really interested in lessons learned associated with developing and deploying algorithms rapidly.
Palmer: Wonderful. Dr. Siegel, I appreciate your taking the time with us today. Thank you so much.
Siegel: Sure. It’s my pleasure, please stay healthy. And, and it was great talking with
you as well.
Palmer: Thank you.