Webinar Report 3

The Reality of AI in Radiology Today

Professor Eliot L. Siegel, MD, Professor Peter Chang, MD, Dr. Cindy Siegel, Professor Patrik Rogalla, MD, Tom Szostak

Cookies: We are sorry we can’t show you this video – as matter of your Cookies preferences. Accept here the Targeting Cookies to watch the video content.

Watch a YouTube video of the full discussion.

Cookies: We are sorry we can’t show you this video – as matter of your Cookies preferences. Accept here the Targeting Cookies to watch the video content.

Canon Medical works together in close collaboration with leading healthcare professionals to gather their experiences and their feedback as to what may help them and their patients even further when using AI. Our partnership with customers fuels our continual innovation. We invited four of the world’s leading radiology specialists to share their views on the AI landscape in clinical practice today.

The panel of experts included Eliot L, Siegel, Professor and Vice Chair at the University of Maryland School of Medicine, US, Professor Peter Chang, Assistant Professor Department of Radiology Sciences UCI School of Medicine, US, Dr. Cindy Siegel, Corporate Director Imaging Operations, UHS of Delaware, Inc., US, and Professor Patrik Rogalla, Professor of Radiology University of Toronto, Canada, Tom Szostak, Director of Healthcare Economics, Canon Medical Systems USA.

Do they think we need AI? What do they see being achieved with AI already? And what do they consider as the prospects for AI in imaging? You can watch the full webinar or read the summarized interview below to gain insights into these questions and more, in a fascinating and surprising discussion.

Question 1
What do you think of when you think of Artificial Intelligence?

"I think of simulation of human thought or decision-making processes derived from scrub data that's curated over a period of time. And that data is programmed to produce decisions that advise and guide humans with a suggested result. I think as time progresses and you curate more data, the algorithm evolves and matures and improves in its decision-making process," said Tom Szostak. "I believe AI doesn't actually bring us intelligence, but rather a critical component of intelligence, and that is prediction. And what I mean by prediction is that we are filling in information that we don’t have based on data that we do have. This general description or definition helps me better understand and direct AI in clinical practice," added Patrik Rogalla. "I think of buzzwords like machine learning, Deep Learning, predictive diagnosis, population health and data mining. However, it’s important to also look at the patient experience, workflow, turnaround time, reduced patient procedure time, and the ability to diagnose more quickly without additional studies. That's what I think of when I think of AI," added Cindy Siegel. "From the perspective of data science, I tend to think about Deep Learning. Other AI technologies such as traditional machine learning techniques are valuable and I advocate using the simplest solution, but if you have a problem that requires data from a million patients’ records to solve, no other algorithm scales as well as Deep Learning," concluded Peter Chang.

Question 2
Do we need AI in radiology?

"I think we do need AI in radiology. And from the patient perspective of coming into an institution, having a procedure done, having it done in less time, and not having to return for a callback – because you have AI embedded into your equipment – it is really a patient satisfier. If you, as a radiologist, are able to see things very clearly with one modality, that's also a patient satisfier," said Cindy Siegel. "I agree AI is needed. There is tremendous inefficiency in radiology and much of it comes from tedious tasks that should be automated. They fill up our day and don’t allow us to think critically about the things that are important. I think each of those tasks is perfectly suited for Deep Learning. Some work needs to be done, but without question, I think AI is needed. It's just a matter of time before we iron out the important details," added Peter Chang. "Yes, we do need AI in radiology. When Jeffrey Hinton said we should stop training radiologists at a CDL meeting in 2016, that really created fear in our specialty. It was provocative, and it made us think, but he was wrong. AI will make imaging and interpretation cheaper so they will be utilized more. In the medium term and short term, we'll need more imaging and more radiologists. I just can't see that imaging will diminish its role," added Patrik Rogalla. "I agree with much of what has been stated and agree that demand will increase due to an aging population, etc.. The real challenge will be having enough money to pay for the increase in demand. If AI can differentiate more complex and acute cases that need immediate attention and drive greater efficiency within radiology, it will help meet the rising demand and curtail rising costs," added Tom Szostak.

Question 3
What killer apps will compel further adoption of AI?

"I'll take a controversial stance. One of the most compelling things you could build is a tool that would to some degree cut the radiologist out of the picture: such as a very high sensitivity algorithm for detecting abnormality. If the algorithm didn’t see anything, the exam would be essentially negative and the radiologist wouldn’t need to look at the images. Without a tangible benefit that obvious, the adoption of AI may continue to be incremental as in previous years," said Peter Chang. "Yeah, one of the most important lessons for residents, fellows and physicians in general, is Know what you don't know. So your killer app would be able to identify a subset and then essentially do those autonomously. How would you know when it's really ready to do that?" asked Eliot L. Siegel. "For this application to be successful, it would have to be orders of magnitude better than a human. If you look at the types of instinctive emotional reactions we get when machines make mistakes, an algorithm that performs only as good as a human will face tremendous resistance," said Peter Chang. "I would like to see a killer app that predicts the chances of developing a disease at a really early stage, before it can be seen. And by the way, I have to respectfully disagree that imaging will increase. I think it will decrease when AI provides more definitive diagnoses," added Cindy Siegel. "So for you the killer app is really population health, where we can screen a subset of patients, identify them and have a real impact on mortality and morbidity associated with that interest," commented Eliot L, Siegel. "Exactly, because that's something chronic and you may not mention that or identify it from a chest X-ray when looking at the vertebral bodies because you're focused on the lungs or ribs," added Cindy Siegel. "I couldn't agree more with Peter on what type of app will succeed best in AI. Sorting out the disease from no disease is indisputably the best application we could ever get. As much as I love X-rays, maybe it's time to get rid of abdominal X-rays and chest X-rays. If an AI application could rid our institution of the burden of 300 X-rays per day, radiologists could concentrate on tasks they actually want to do, such as MRI, CT, cross-sectional imaging and radiomics," added Patrik Rogalla. "So you want your killer app to kill X-rays, conventional radiographs. I like that. There's an irony there also," commented Eliot L, Siegel. "I think that if you really want to serve the radiologist community with a killer app, then create a killer app that curates all the quality reporting metrics that Medicare requires. It takes a lot of person-hours to do and if you can develop an app that eliminates the human interaction of curating that data and uploads and sends it digitally to CMS in the Quality Payment Program, I think that would curry a lot of favor from administrators and radiologists," added Tom Szostak. "I've got to step out of my moderator role for just a moment and give you my opinion of killer apps. I think what exists today with regard to incorporation of Deep Learning into reconstruction algorithms to reduce noise and improve image quality in CT – even for conventional radiographs, MR and PET scanning – is really incredibly exciting. And I think there will be many related killer apps in the future. We're already seeing an incredible amount of incorporation of those technologies that will completely change the way we end up forming images moving on in the future. So for me I think that's another one to consider," concluded Eliot L, Siegel.

Question 4
What is AI most likely to enhance in the short-term?

"In the short-term, I think the things you just mentioned – clinical outcomes, improved efficiency and improved patient experience – as well as non-imaging things like reconstruction, are the ‘low-hanging fruit’. The technology needed to solve those problems is well-defined and we know how to do it. Now it’s simply a matter of execution and really seeing how these algorithms act in the wild to figure out where things are headed," said Peter Chang. "I agree. In the short term, as well as the long term, I think AI will enhance clinical outcomes, the ability to diagnose something quickly, the patient experience of not having to come back for a repeat study or additional study, turnaround time for the radiologist and getting the reports to referring physicians," added Cindy Siegel. "I would put patient experience in the forefront and I agree with Cindy. In the end, we are physicians, right? And we care for the patients. I think AI can really help integrate radiology in the circle of care more deeply and make us relevant in this whole continuum of care. I think that's the ‘low-hanging fruit’. and I believe that's the most important one as well," added Patrik Rogalla. "Do you see a particular barrier toward us getting there at this point. It's already the end of 2021 and we talked about the emergence of Deep Learning in 2012 or 2011. People keep saying this is right around the corner. Why aren't we there yet," commented Eliot L, Siegel. "Hierarchy, meetings, fear, risk aversion… It's a general aversion to innovation. In particular, the whole healthcare system is by definition conservative – for good reasons. Organizations have become complex and difficult to navigate. It’s a complexity problem that AI is best suited to solve," commented Patrik Rogalla. "I like Professor Rogalla's statement and similarly found resistance to innovation over the years. I keep thinking about the role of AI in value-based care and what will it bring when all of these risk share arrangements come into play and the payment formula changes. That's where I think AI will really be embraced even more to help people improve those efficiencies that will help to improve the bottom line and patient experience. But it's going to take a major event to get everyone to move in that direction," added Tom Szostak. "So you're not really predicting that there will be major changes in implementation of AI in the next two or three years unless there is some other significant sea-change?" commented Eliot L, Siegel. "Correct. Technology gets ahead of the regulatory environment and then the regulatory environment steps in and slows the momentum. It stifles innovation to an extent. Maybe that challenge needs to be articulated when standards are created," concluded Tom Szostak. //
Professor Peter Chang, MD
Assistant Professor Department of Radiology Sciences, UCI School of Medicine, USA
Professor Eliot L. Siegel, MD
Professor and Vice Chair at the University of Maryland School of Medicine, USA
Professor Patrik Rogalla, MD
Professor of Radiology University of Toronto, Canada
Dr. Cindy Siegel
Corporate Director Imaging Operations UHS of Delaware, Inc. USA
Tom Szostak
Director Healthcare Economics Canon Medical, USA
Contact Us