How Artificial Intelligence May Change Interventional Oncology

In this episode of the Society of Interventional Oncology (SIO) Corner podcast, Dania Daye, MD, PhD, Massachusetts General Hospital, Boston, Massachusetts, joins host Elena Violari, MD, RPVI, to discuss artificial intelligence (AI).

Dr Daye explains the different areas of interventional oncology (IO), and medicine in general, where artificial intelligence may be applied to improve patient outcomes through precision medicine. For all physicians, Dr Daye advises, “There are certain applications where AI does great, and there are certain applications where AI does not do great. Knowing what those applications are and what to do when an AI fails is really, really critical.”

Transcript:

Elena Violari, MD, RPVI: Hello everyone and welcome to the SIO Corner podcast. I'm your host, Elena Violari, and it's my honor to introduce our guest today, Dr Dania Daye. Dania, welcome, and thank you for joining us today.

Dr Dania Daye is a physician scientist and an assistant professor of radiology at Massachusetts General Hospital [MGH], Harvard Medical School. Dania is the interventional radiology [IR] division quality director, co-director of IR research at MGH, and director of the Precision Intervention Medical Imaging Lab. Dania, we are so honored and excited to have you here today and thank you for joining us today.

Dania Daye, MD, PhD: Thanks for having me.

Dr Violari: The main topic of our conversation will be your research regarding applications of artificial intelligence in imaging and interventions. But before we dive into our topic, why don't you tell us a little bit about yourself, your career path, and background?

Dr Daye: Sure, of course. In terms of my background, I'm an engineer by training. I did my undergraduate degree in bioengineering, then I moved on to do an MD/PhD at the University of Pennsylvania, and for my PhD, I was an HHMI/NIBIB [Howard Hughes Medical Institute/NIH National Institute of Biomedical Imagining and Bioengineering] Interfaces scholar, so I did my PhD between an engineering lab and a basic science lab in bioengineering, where I studied advanced imaging for breast cancer recurrence.

Through that work, I got very interested in machine learning and precision medicine. I had phenomenal mentors who really showed me the way and opened my eyes about all the exciting things we can do in that space. I then went on to MGH, where I did my residency and fellowship in IR and stayed on as faculty, where I started my lab, the Precision Interventional Medical Imaging Lab, where I focus on using machine learning for precision medicine.

Dr Violari: So wonderful and inspiring journey. Thank you so much for sharing with us. My first question is, what is precision medicine and how do you define it?

Dr Daye: Of course, precision medicine is important. It’s topic that has been very hot in the field for a while now and probably the whole concept started more than 10 or 15 years ago and started with the concept of picking the right management strategy or treatment for the right patient for the most optimal outcome. It used to be called “personalized medicine” and then got to being called “precision medicine” and it really goes around this entire concept that, right now, a lot of the decisions that we make in medicine are based on population-level data. For example, in breast cancer a lot of the screening decisions are based on population-level data that has been available for a while, if we look like, for example, at the Gail model and some of those models are available out there. And really, the more people looked into this, the more the question comes out: for this specific patient right in front of us, how do we make a decision for the “n of 1”? For that specific patient, what is actually the risk, not the risk of that population that really fits some of their characteristics?

And this really gets down to a couple of different things. For example, for us in medical imaging, we know that medical images have so much data that we're currently not using in our day-to-day clinical decision-making. We usually look at the images, for example, a patient has a kidney stone or does not have a kidney stone. But, what about the rest of the imaging series and all the pixels that are telling us a lot more information that we're currently not harnessing to be able to predict certain risks or certain outcomes for our patients. This is where the whole concept comes, how can we really use all of this data to build models that are specific for specific patients for a better outcome for them.

Dr Violari: That is a very interesting perspective. How did you get interested in researching this this topic?

Dr Daye: When I was an MD/PhD student, I had a phenomenal mentor, Dr Despina Kontos, who was very interested in this question. And she taught me a lot about how, do we really think about expanding our current and existing risk models for cancer prediction, to be very specific to the patient in front of us. And at that point, she was using texture analysis, one of the first projects I ever worked on in this space. And how do we use texture features to really enhance the predictive performance of some of the cancer prediction models that we have. We did our first project on this, we got some really great results and it really opened my eyes that there is something beyond what we currently have in medicine and our current standard of care to improve outcomes for our patients. This is really what drives me to do this research.

Dr Violari: Sounds like a great experience. You've published, and lectured, a lot regarding the application of machine learning in personalizing treatment decisions in IR. Can you please talk more specifically about the different applications of AI in IR and specifically in interventional oncology?

Dr Daye: Of course. When I give lectures on this, I always start my first couple slides talking about the 5 different buckets of the applications of AI in IR, and more specifically, IO that that we can get into it. The first bucket that I always talk about is in patient selection. And this is the first area where we're seeing a lot of products currently available on the market. Many of the products available on the market allow us to identify patients that are good candidates for IR procedures, and also help us do care coordination. There is some interesting data coming out from some of those products right now that is starting to show us an effect on some key outcome metrics in terms of length of stay and even I just saw an abstract recently affecting mortality in one of the cases. So very exciting data in that space probably the most developed of the applications of AI in IR.

The second bucket that I talk about usually is pre-procedural planning which is very applicable for us in interventional oncology. I think most of us today use some software for segmentation for Y90 dosimetry. A lot of the softwares that are available on the market are actually AI-based and you have FDA clearance for that component. Most of us are using some version, many of us may not be aware that these are actually AI-enabled softwares. There is a lot of work also being done on endograft sizing, other pre-procedural planning components.

Now the third bucket that I like to talk about is intraprocedural support. There are a couple of products available in the market but I think this is probably the area where we're still mostly in the infancy of what can be done in that space. There are some applications relating to biopsies and a fusion of pre- and intra-procedural images to allow us to better target lesions for biopsies or ablations, for example, in IO. Probably those are the most developed, but in terms of applications specifically for angio applications, beyond some of the guides that we see, for example, EmboGuide is AI-based, there's still a lot that we can do. We’re just scratching the surface in that third bucket.

The fourth bucket, which also is very relevant for IO, is predicting response to treatment. And there have been many, many papers that have been published in the last probably 5 to 10 years in the space of predicting response to TASE, predicting response to Y90, etc. And I think the sky is the limit there in terms of what we can do. There are lots of very exciting opportunities in that space, to really be able to pick the right treatment for the right patient to have the best outcome, so we can really preselect who's going to respond to what treatment.

And the final bucket that I like to talk about is really the new applications of large language models, like ChatGTP, and augmenting our patient-facing care to patients. We're starting to see a lot of papers come out about how ChatGTP can really enhance the patient experience. Many papers being published on chatbots, for example, lots of papers being published on decreasing the readability level of some of the instructions that we provide to our patients, or making our results reports more patient friendly, etc. Lots of opportunities in that space, as the fifth bucket of application. I think this is really going to change how we practice over the next couple of years. Now, there are a number of applications that are really improving workflow, et cetera. I won't talk about them, but I think AI is going to touch pretty much every single area of IR in the next 10 to 15 years.

Dr Violari: This is an amazing response and the way you described it with the different buckets, it really allows us to organize all the different applications and where they should be applied. It’s very exciting. I think it's going to be wonderful in the future, being able to apply all these applications.

There's an increasing body of literature on radiomics and deep learning, artificial intelligence applications in medical imaging. There is also a steadily increasing number of, like you mentioned earlier, FDA-cleared AI applications in radiology. Today, how many FDA cleared applications are we using in IR?

Dr Daye: Yeah, that's an excellent question. I will start by saying that if we look at the list of AI-enabled medical devices that are currently on the market, the last report that was put out by the FDA in October was a little bit over 690. Today, I believe I just looked recently, it's a little bit over 700. Now, those devices, so they can be SaMDs [software as a medical device] or SiMDs [software in a medical device], I won't go through the details here, but I will say most of them are in diagnostic radiology if we look at the applications.

However, if we dig a little bit deeper. I believe it's a little bit over 500 in radiology, very few of those are actually in IR. I will say probably a handful or something along those lines. And we still have a long way to start having more products that are approved specifically for IR. As a specialty, IR presents a lot of challenges for moving forward with applications of AI that are suitable for our specialty. And I'm more than happy to talk about why that is. But we're starting to see some push in that direction, which is very exciting, but we're still in the infancy compared to diagnostic radiology and there's still a lot of work to be done.

Dr Violari: I'm very interested to hear more about the challenges, why it's more challenging compared to diagnostic radiology.
My next question would be, how can we use AI and interventional oncology to identify the right treatment for the right patient and to ensure the most optimal outcome? And do you think we're close to putting this kind of technology into clinical practice? And if not, what do you think needs to happen before we can and how can we overcome these challenges?

Dr Daye: I think this is a great question. I will say that the future is bright. There is a lot of potential here. However, we're not there yet, unfortunately. There's still a lot of work to be done. And I will point out that in IR today we have very small data sets. And we don't have standardization about the data that's collected or how we collect data interprocedurally or for different patients, etc, and that makes it very difficult to be able to combine datasets across different sites or different institutions to be able to develop proper robust machine learning models.

A number of things have to happen before we can start developing large scale models. The first one is building large multi-institutional datasets, which has been probably the biggest barrier for us to be able to see larger applications being implemented in IR. The second being the standardization and being able to have standard data collection across different sites, or even if we start looking at interprocedural images for doing any procedure, having specific views, et cetera. Because I mean, we all know here that there are very few IRs that will collect the exact same views for 2 different procedures, much less across different IRs.

There are certain things that probably will have to happen to allow us to have access to such large datasets, but there are a number of efforts underway, so I'm optimistic, but I think there's still a lot of work that needs to be done.

Dr Violari: I'm optimistic too. You were recently awarded the SIR Gary J Becker Young Investigator Award, regarding your research, Beyond MELD Score. Congratulations, that's a great achievement. Specifically, the title is Association of Sarcopenia with 90-Day Mortality Post TIPS [Transjugular Intrahepatic Portosystemic Shunt] Placement. Now, how was the idea conceived for this? I thought it was a great topic.

Dr Daye: Thank you for mentioning that. I will say that this really goes back to my interest in precision medicine and how can we really personalize the prediction of mortality for these patients that we all see. The MELD score has been the standard of care for many, many, many years. And I know over the years, there have been a couple of different attempts at trying to make the predictive performance of that score better, like MELD-Na and among other things.

When we start looking a little bit deeper — there is a developing area of research around CT body composition that I don't think we have really incorporated much of that into IR yet. CT body composition has been very hot in a number of other areas of medicine, especially by our surgical colleagues who are starting to use that to predict response to certain surgical intervention, long-term mortality, among other things. Seeing some of that data really got me thinking about how can we really start exploiting some of those metrics into our IR interventions, to be able to better select patients who would benefit from our intervention. Based on that, the idea was born for this project and we already had a CT body composition algorithm in place that is based on deep learning that we were able to apply to a patient population that underwent TIPS in our institution, use those metrics to incorporate into a risk-prediction model and see how adding those metrics into a model that included the MELD score, how that improved performance in predicting a mortality at 90 days.

Dr Violari: Wow, that's amazing. And I'm pretty sure this is something that will be applied in the future for these patients. How did you develop this machine algorithm that allowed you to evaluate this 90-day mortality of TIPS? Can you explain, in simple terms, what your algorithm does?

Dr Daye: We'll first have to touch a little bit more on the CT body composition and how the algorithm works. There is quite a large body of literature published on using machine learning for CT body composition metric computation. In a nutshell, what we do is we take a slice from a CT abdomen/pelvis, specifically at the L3 vertebral body, an axial slice. We then use a deep learning-based technique for automated segmentation of the visceral fat, subcutaneous fat, and the muscle. Our algorithm will calculate a number of metrics, based on those 3 segmentations, that allow us to really quantify the amount of muscle, subcutaneous fat, and visceral fat. We then incorporate all those metrics that we extract into a model that also incorporates the MELD score. And we assess the predictive performance in terms of predicting 90-day mortality, because we do have the ground truth data in our model, about which patients died and which patients were alive at 90 days at our institution. And based on that, we can assess the performance of the model using our ROC analysis and confusion matrix to be able to see how the model performs. In the paper, we talk a little bit more about how we compare the performance of just the MELD score alone versus the MELD score plus the CT body composition metrics, and we show that there is a significant improvement in the performance, which was very exciting when we saw that.

Dr Violari: That is very exciting. That's awesome. And briefly, can you talk to us a little bit more about the results of the study.

Dr Daye: Of course. We were very excited to see that there are a number of CT body composition metrics that are indeed correlating with survival at 90 days when we do simple logistical regression analysis. And then, once we incorporate them into a model that includes MELD score, when we do the comparison of the performance between the model that just has MELD versus the model that has MELD plus the CT body composition metrics, we see that there is a statistically significant improvement in the performance in predicting 90-day mortality when we add the CT body composition metrics, which we were very excited to see.

Dr Violari: Did you guys see any correlation of worst MELD scores in patients with sarcopenia?

Dr Daye: Yes, absolutely. And we were happy to see that. And this was published also by another group, that sarcopenia is indeed correlating with worse outcomes after TIPS. We were happy to reproduce that as well. I think this is something that we all should start looking at as we're selecting patients for TIPS procedures.

Dr Violari: For sure. Now, how do you see this project affecting clinical practice? Specifically, how can we use the results of your study to optimize patient selection for TIPS and most importantly, on when to offer TIPS? Do you think offering TIPS to patients earlier, prior to them undergoing large-volume paracentesis weekly, would that prevent sarcopenia and therefore better outcomes following tips?

Dr Daye: I have to say, I think there's going to be a lot of opportunities for us to really think how to incorporate the results of studies like this into our day-to-day practice. We know that in surgery there are a lot of practices that will optimize the patient medically before they do a surgery. And we know that many of the TIPS procedures, at least the TIPS procedures that we studied in the study, were elective TIPS procedures. They were not really because of bleeding, etc. Because of the elective nature of the tips, it really begs the question of should we optimize our patients medically before we do certain procedures? And this is an area in IR that we have not really put a lot of emphasis on before and I think that as we care for more patients and we become more and more of a clinical specialty as the direction is today, we're probably going to see more and more attention being put into optimizing our patients for our procedures for the best outcome.

Dr Violari: I agree with you and I really think that this was a great and very important study. What do you think is the most important thing for IRs to understand about AI applications in IR practices? Do you think that physician IRs who don't apply AI in their practice will be disadvantaged versus IR physicians who do use AI in their practice? Also, do you see these technologies being available and put into use in community practices?

Dr Daye: I will say my personal prediction is that AI will definitely change how we practice over the next 10 to 20 years. We're starting to see a lot of signals of how AI can really improve how we practice in community practice, our efficiency and potentially in some cases patient outcomes from some of the emerging data.

We might get to a point, although it's very early to predict this, that if we start seeing significant effect on patient outcomes by some of the AI algorithm, because of quicker response or about our patient selection, etc., we may get to a point where IRs who use the AI tools may eventually replace IRs who do not use these AI tools, because obviously patients want the best outcome. Once we start touching on outcomes with the AI tools, that will become eventually the standard of care.

The message I want to get across is that AI is definitely coming, if it's not already here, in many applications and it's very important for us as a specialty to embrace it fully. Especially given some of the trends we're starting to see, because it's going to change how we practice in the long term.

I will also add a plug to large language models. Large language models have been here for maybe about a little bit over than a year. And all of us here can attest to how these models have changed how we think about certain things. These models haven't even been trained on medical data. I'm just going to leave this comment here, that once we train these models on medical and clinical data, the sky is the limit. I'm very excited to see what they're going to do and how they're going to help augment our practice. Obviously with some caveats, we're seeing some situations where they fail. But being in tune with some of these developments is really key for us as a specialty, to keep advancing, as AI is advancing.

Dr Violari: This is so exciting. Besides embracing it, is there any other advice that you give to IR physicians and how they can include applications of AI in their practice? How do we keep up with this advancement?

Dr Daye: I have to say that one of the key things I recommend, I talk to a lot of people who ask this, is really educating yourself a little bit more about AI is important. When I say educating yourself, I don't mean you need to know how to program an algorithm. That's absolutely not what I’m saying. But educating yourself about what questions to ask of industry when you're trying to evaluate an algorithm and whether or not to buy it. Knowing when an algorithm fails, knowing some of the pitfalls, and some of the nuances beyond just seeing what happens when an algorithm does well, I think all of these are important areas for all of us to educate ourselves on. Because there are still a number of limitations for AI and there are certain applications where AI does great, and there are certain applications where AI does not do great. Knowing what those applications are and what to do when an AI fails, I think is really, really critical. Some of these are still active areas of research, but just having some familiarity with those I think is very important.

Dr Violari: Wonderful. Dania, this has been very educational for me and I'm sure for our listeners as well. I've learned a lot, and it has been a pleasure doing this podcast with you. Thank you so much for being here today, and once more, congratulations for your research, for the Gary J Becker Young Investigator Award. Again, I've learned so much just by talking to you today. Thank you so much for sharing your expertise and knowledge. I really appreciate it.

Dr Daye: Thank you, Elena. Again, thank you for having me.

Source:

Elhakim T, Mansur A, Kondo J, et al. Abstract 212 - Beyond MELD score: Association of sarcopenia with 90-day mortality post transjugular intrahepatic portosystemic shunt (TIPS) placement. J Vasc Interv Radiol. 2023;34(3):S96. doi:10.1016/j.jvir.2022.12.272.