Challenges With Artificial Intelligence in Interventional Oncology

At the Society of Interventional Oncology (SIO) 2024 Annual Scientific Meeting, Judy Wawira Gichoya, MD, Emory University School of Medicine, Atlanta, Georgia, presented on the challenges of utilizing artificial intelligence within interventional radiology.

Dr Gichoya states that more representative data-sets can be built when clinicians work together to design better systems through larger cohorts.

Transcript:

Hi, my name is Judy Gichoya. I am an associate professor of interventional radiology and informatics at Emory University School of Medicine. I just came back from the Society of Interventional Oncology, SIO, Meeting in LA last week. And I had a great opportunity to connect with peers. I was also presenting at the AI session. We had 4 panelists; my specific topic was on practical and ethical challenges of artificial intelligence (AI) in radiology.

I want to tell you that it's very difficult to do AI in interventional radiology. Everyone shows you things around stroke and cardiology, but remember, those people go to very specific vessels. If we had 5 interventional radiologists placing a right internal jugular chest port, we would all do it in very different ways. Some people tunnel fast, some people tunnel last. Some people assemble the port outside.

I started off by reminding people why this is very difficult. One, we live in a place where we have few data sets and our art that makes us very different in terms of how we operate all the time. My talk started off with showing my failed project — failed not in terms of technical achievement, actually this was an IVC filter detector that was inspired by a patient that I had taken care of and it worked very well, but we've never implemented it here at Emory. I thought this was going to be an easy lifesaving intervention. This is because there are many technical challenges that govern our work when we deploy artificial intelligence in the real-world clinical settings. And we are all learning what those look like. But more importantly that there's quite a big role for AI in intervention oncology. In fact, this is where we've seen quite a lot of proliferative applications.

Earlier on during the day, before the AI session, we had a lot of talks about the use of immunotherapy, as combination therapy with transarterial chemoembolization (TACE) and Yttrium-90 (Y90). This is an exciting time for us, but most commonly you'll hear that we need to understand the tumor microenvironment. One thing can change the environment and that will make something else more effective. This type of microenvironment is related to something that I work in and that I discussed, which is that pictures tend to have more than what radiologists can see. We know this from AI because, for example, if we showed you a chest x-ray of any patient, you can tell their sex, you can tell their image-based age, you can tell their ICD codes, you can tell their healthcare expenses. You can tell even a little bit about the type of area they live in — not because they can tell you the exact zip code or the area deprivation index, but it can tell you if they live in a deprived area or not, and if they're black, white, or Asian. If this metadata can be acquired just from an image, it tells you that we have potential, especially when we join imaging, for example, CT scans or MRIs, with the digital slides, for us to really understand the tumor microenvironment. Very, very big potential to understand whom should we treat with what earlier on, in this era of precision oncology.

What could go wrong is that if you think about how our practices are, for example, during training at the Dotter Institute [of Interventional Radiology at Oregon Health & Science University, Portland, Oregon] I did very few Y90 cases, most of them were case interventions. As a resident at Emory University, we did quite a lot of Y90 cases as an attending. Now, we do quite a lot of Y90 cases. And to me, what I have realized over this time, your pattern of referral affects the type of care you get as an interventional oncologist. And this is because your tumor board may decide, well, this patient should go down to this pathway. These patterns are very easily picked up by AI, which is going to be one of the biggest challenges for us. Because if you trained an AI system at Emory, then it would only work well for patients who their decision-making is to start off with Y90. You may see earlier disease. Well, if in other places, for example, for cholangiocarcinoma, they would get this at a later stage. Our patterns of care are encoded in the datasets that we have. And these AI systems tend to be very lazy, and they pick up the easiest pattern. And that's going to be one of the biggest challenges for us to overcome. Moving forward, I encourage us to think about how patients come to us, to go back to the source of the river, where we see, how do we design better systems and work together in larger cohorts so that we have more representative patient practice patterns.

But I remain optimistic not just because of AI for intervention oncology, but in other domains, for example, with the ChatGPT systems that we are seeing that patients can now access information in the languages that they speak. I don't speak English as a first language, and you don't only need to speak English to understand your medical information. And in the era of robotics, where we see quite a lot of innovation. There were some robots in the exhibit hall, and those are powered by AI. And I remain optimistic about our ability to better take care of the cancer patients.

Source:

Gichoya J. “Practical and Ethical Challenges with Artificial Intelligence.” Presented at the SIO 2024 Annual Scientific Meeting; January 25-29, 2024; Los Angeles, California.