Jason Adleberg Official Transcript

Jason Adleberg – 00:00:02:

Welcome to Frame by Frame: Rethink Imaging, a podcast by Imalogix. Here, we explore the intricate world of medical imaging, aiming to dissect the field and inspire both professionals and curious minds alike. I’m your host, Chris St. John. Today on Frame by Frame: Rethink Imaging, we are super excited to welcome Dr. Jason Adleberg, a radiology resident and AI researcher at Mount Sinai in New York City. Jason used to work as a software engineer and later on had a career change to radiology. He’s written multiple AI papers, serves on the editorial board for the Radiology AI Journal, and led a workshop at RSNA on how to build AI models without any coding experience. That workshop is also available on YouTube. Today, we’ll delve into his insights on simplifying complex radiological topics, the importance of patient safety, and the exciting innovations shaping the future of radiology. Welcome, Jason.

Chris St. John – 00:00:59:

Chris, thanks so much for having me.

Jason Adleberg – 00:01:01:

Yeah, and just a little bit of background for our listeners. I actually went to college with Jason way, way back in the day. And as a relatively newcomer to imaging at this point, less so than I used to be, I’m just super excited to get to talk to somebody that I already know about all of this.

Chris St. John – 00:01:20:

Absolutely.

Jason Adleberg – 00:01:20:

Yeah, so let’s just start a little bit with your background. Jason, your career has evolved at the crossroads of technology and medicine. Can you just talk a little bit about your journey from AI and programming into radiology?

Chris St. John – 00:01:35:

Yeah, sure. See, what’s funny, because when I first started in medical school, I didn’t know much about radiology. We don’t really learn that much about it in like the first two years in the preclinical curriculum. It’s not really until we go into the hospital that we sort of see what radiology is and how, you know, how much it kind of weaves together all these different fields of medicine. So later on in med school, I discovered radiology. I, you know, decided to apply for it for residency. And, you know, with my background in computer science and AI, I realized my skill set could be a good fit in that setting. When I was in my last year of med school, I had some time to do research and I did some different projects, AI projects. There’s a lot of really great resources for people trying to get into this field. Some great resources like fast.ai, Kaggle, different sort of tutorials on the internet. And since then, I’ve had the opportunity to work on a lot more projects.

Jason Adleberg – 00:02:31:

Very cool. Once you entered the field of radiology and you really started developing a knowledge base, how did that influence your perspective on how AI could be incorporated? Right. Like, did you have an idea of what it would be before? And once you kind of went through your training, how did that idea evolve?

Chris St. John – 00:02:50:

Yeah, it’s a good question. And I’d say that like my views on AI and radiology have changed over time. I think as I’ve gone through training, I’ve realized it’s actually a lot harder to do this stuff than I had realized at first. You know, when I was a medical student, there were a lot of headlines about how there were kind of these like primitive applications that would look at an image and tell you if it was a dog or a cat. There are different, you know, hot dog or not a hot dog, right? There are lots of different, yeah, there are lots of different applications like that. And there were maybe a few sort of, I remember going to RSNA, the big radiology conference, and there were just a few booths of some companies that were trying to do that sort of thing for radiology. Pneumonia or not pneumonia, pneumothorax or not pneumothorax. You know, this was, I guess, 2019. And since then, you know, I’ve been pretty lucky to go to RSNA just about every year. And, you know, the AI section feels like it’s like the biggest part of the conference. So, yeah, I think people are kind of discovering new ways to use AI and realizing that this is probably going to really change our workflow as radiologists.

Jason Adleberg – 00:03:57:

Yeah. So you were just saying that it’s much harder to implement than you expected. What were the things that kind of stood out where you’re like, oh, this is like, what are the walls that you started to find that increased that difficulty or expected difficulty?

Chris St. John – 00:04:13:

Definitely. I mean, if you remember when you and I were in college, around that time, that was when Facebook was getting really popular. And I remember. They were like, they’re trying to recruit on campus and trying to, you know, get people to work for them. And their sort of motto was move fast and break things. That kind of felt like the mantra of the tech world. And I think that in medicine, this idea of moving fast and breaking things is kind of the opposite of what you want to do. So, you know, medicine, from my perspective, has always been pretty far behind in technology. I mean, even radiology, like we still use CDs to store images, right? If you go to the doctor to get an x-ray. And then another doctor wants to see that x-ray, like there’s a good chance you’ll get a CD with those pictures on it. You know, and like who even has a CD player nowadays, right?

Jason Adleberg – 00:04:59:

Yeah, they’re not even built in machines anymore.

Chris St. John – 00:05:01:

Yeah, you know, we’re talking on my laptop and I don’t have no way to open a CD. So I think part of this is like for good reason. I mean, you know, there’s lots of security issues and you don’t want to break stuff, right? But I do think that we could be doing so much more with technology. And I think one of the biggest challenges in this space really is having the right people to implement the technology.

Jason Adleberg – 00:05:21:

And what would you be looking for in the right people?

Chris St. John – 00:05:24:

Yeah. You know, I think that healthcare organizations have pretty complex systems when it comes to the technology that they use. I mean, you know, just in radiology alone, we have one system that stores all the images. We have another system that handles the reports or like the text documents that we’re creating. We have a different system to do what’s called protocoling, which is like figuring out how we’re going to do the studies before we’ve even done them. Of course, we have totally separate systems to manage the patients. There’s an electronic health record. There are systems to do like scheduling, like when are they going to get their appointments and kind of figuring out how all these systems talk to each other is actually really complicated and super complex. And, you know, this is just, I’m just talking about radiology, like in a large hospital setting. I mean, you know, there could be thousands or tens of thousands of doctors and hundreds of thousands or millions of patients. And, you know, there’s a lot of data and information moving around.

Jason Adleberg – 00:06:22:

So yeah, like what you’re saying is like that these barriers to entry is like, there’s just infinite workflow pathways, right? And it’s like, how are you going to integrate all of this, not just technologically, but also all of the human minds and all of the channels and the patients and the doctors, how are you going to get a smooth path through this insane process?

Chris St. John – 00:06:46:

Exactly, yeah. You know, I think when it comes to like implementing artificial intelligence in a healthcare setting, there are a lot of new questions that AI kind of brings to the implementation process. And we’re so early in this that we’re still kind of figuring out who in the healthcare system is going to manage the implementation of this software, who’s going to approve it. Once it’s installed, how we know that it’s going to work and it’s going to keep working, that’s an issue that comes up with AI a lot. And it’s just sort of every organization has a different way that they manage this.

Jason Adleberg – 00:07:23:

Yeah. And so if you’re cool, I’ve talked with various folks about implementation of AI into medical imaging, and we’ve always kept things relatively high level, right? Like we’re talking concepts, we’re talking where things could go, we’re talking collaboration, we’re talking suites. But I’m more curious, like you’re down and dirty. You know what I mean? Like you have a unique perspective in that you’ve actually been building these things and you’ve actually been reading these images. And so how does like the dual perspective of building these AIs, actually reading the medical images, how does that influence how you approach? Radiology in general and like radiological problem solving.

Chris St. John – 00:08:08:

Yeah. You know, one of the issues with these AI models is that they don’t always explain how they reach a conclusion. So that’s one issue that comes up. And another issue is you kind of want to know how confident the AI model is. So, you know, from a radiologist perspective, you really want to know why an AI is making the decision that it’s making. I think one of the things I’ve learned as I’ve gone through training is that things are not always as black and white as we think they are. A lot of times as a radiologist, you have to say something like, you know, I think this is probably this, but I’m not sure it also could be this. And a lot of the AI systems that we use right now in 2024 are basically just looking at one thing. They’re just looking at, is this a pneumothorax or not? And the answer actually might be like, not really sure. I think that’s one thing that I have an appreciation for having experience building AI models is that it’s really important to calibrate. The AI models. You know, a lot of times, again, we think about them saying a yes or a no, but really what they’re doing behind the scenes is they’re actually giving you a number between zero to one or, you know, zero to a hundred. And, you know, part of designing an AI model is saying, well, okay, what is like the threshold for a yes or a no? Is it 50%? So if it’s 50% positive that there’s something going on, like how do you use that information? And I think in the future, you know, there’ll be a lot more appreciation for the sort of confidence value that an AI is putting out. Because again, right now it just says yes or no for most applications.

Jason Adleberg – 00:09:41:

Yeah, for sure. And with that, you know, I hear very casually in conversations in the past and with different radiologists and different folks like, you know, there are people out there who are wary of AI replacing various aspects of their work. But, you know, from your experience, where do you see it as a tool to enhance rather than just like, you know, the WebMD AI version where you feed a medical image in and it says, okay, this is what you got.

Chris St. John – 00:10:09:

Yeah, so I think… This is a narrative that I’ve heard a lot. Especially in medical school, a lot of people would say to me, why do you want to go into radiology? The field’s not going to be there in 10 years. AI is going to take it over. I mean, look, one fundamental important thing is that the way that healthcare works in America is that we do have these AI systems that are starting to trickle through, but they cannot absorb medical legal risk. So you can’t just sue an AI. And I don’t think it’s really any time soon that the AI is going to completely take over the job of the doctor. I mean, for that reason alone, I don’t see it replacing us. I think that’s just like the most straight to the point thing. But the saying that I’ve heard is that doctors who use AI are going to replace doctors who don’t use AI. And in that element, I think that, you know, there are lots of ways that AI can help speed up our workflow, can automate repetitive tasks. It can, within radiology, it can kind of help us maybe look for sort of things that are like we perceive as being kind of tedious. And I think that if AI makes a clear argument that it will save time, I think that’s like really appealing to radiologists and doctors.

Jason Adleberg – 00:11:17:

Right. And not just time, right? Like, you know, I’ve been talking and doing a little bit of casual reading, but, you know, especially on like radiologist burnout, the way that I see it, AI could be a really helpful tool in just like getting the human through the day and preventing that burnout.

Chris St. John – 00:11:36:

Yeah. I mean, I think that, again, it’s like as doctors, we kind of know in the back of our heads that there are some processes that we feel like could be solved with technology, but for whatever reason, it’s, you know, not, maybe it’s a little bit slow or we don’t have the solutions. One particular example actually is with protocoling exams. So basically there’s different ways, slightly different variations that you can do the same exam. So for instance, if you want to do a CT of someone’s belly, right, you kind of want to know why the doctor ordered the scan to take the best pictures. So if they’re worried about a vascular issue, maybe, you know, they have a history of some disease with their aorta, then you want to do a study where the contrast is in the arterial phase, or it’s basically lighting up the aorta versus, you know, maybe if someone has a problem with their kidneys, then you want to do a totally different type of scan where you actually want the contrast to be in like different phases, you can see different characteristics of the kidney better. So the reason I bring this up is because in a lot of places, like, you know, there’ll be a whole list of studies that could get done that day. And it’s the radiologist’s job to go through each one and say, okay, this is, you know, aorta protocol, this is a kidney protocol, this is something else. And that kind of feels a little bit tedious because it’s a little bit mindless. It’s not really as interesting as like looking at images. So that’s to think a good use of how AI could automate something that’s like a little bit tedious.

Jason Adleberg – 00:13:02:

Yeah. And then just to totally pivot, but you know, you led a workshop at RSNA on how to build AI models without any coding experience. Can you talk a little bit about what that workshop looks like and how, you know, if there was another radiologist out there who wanted to start tinkering, how they would begin to tinker?

Chris St. John – 00:13:20:

Yeah, definitely. I think one of the cool things that’s going on in the AI world that is sort of new to me is that people are developing these tools for you to like do your own AI, even if you have no idea how to program. So, you know, when you’re building an AI model, an AI application, what you really need is you need to have the data. So let’s say for instance, that you want to build an AI model that will detect a pneumothorax.

Jason Adleberg – 00:13:46:

Sorry, forgive my naivete. You said pneumothorax a couple of times right over my head. What is a pneumothorax?

Chris St. John – 00:13:52:

A pneumothorax is a collapsed lung.

Jason Adleberg – 00:13:54:

Gotcha. Okay.

Chris St. John – 00:13:55:

You know, sometimes if you’re in a car accident or in trauma and maybe you have like you break a rib, sometimes the rib can, you know, hit that lung space, your pleural space, and your lung can be unhappy as a result of that. If you do have pneumothorax, the reason I bring that up is because that’s sort of an emergent condition where someone needs to intervene. So that’s something that like the provider really needs to know and kind of quickly. But, you know, it’s also just like something that comes up relatively common on an x-ray and it’s something that’s important. So what you need is you basically, as a radiologist, you just, all you need is a collection of x-rays of images that have a pneumothorax. And then you need a collection of x-rays that don’t have a pneumothorax. And, you know, 10 years ago, you would also need to write all the code from scratch. You would have to, you know, either you yourself or someone else would have to write all this, you know, Python code or whatever. And that would take a really long time. But now a lot of that stuff has actually been automated. And as long as you just dump the positive cases, those pneumothorax cases in one folder and the negative cases, the control cases in another folder, you can just click a few buttons, wait a few hours, and your AI model will be ready to go after that. So I think that’s a really exciting space. And I think that, you know, there are a lot of people working on that. I think that will really mature in the next five to 10 years.

Jason Adleberg – 00:15:10:

Yeah. And are these like AI inception, right? You’re like using an AI to build an AI.

Chris St. John – 00:15:15:

Yeah, sort of. I mean, I think it’s like people have created these pipelines and they just kind of like, you know, do it in a way that you, the user, just needs the data.

Jason Adleberg – 00:15:24:

Right. And so these are online services, they’re programs. Like, is this like something you can download from GitHub? Like what?

Chris St. John – 00:15:30:

Yeah, there’s a few different websites where you can just build your own AI models. And there’s like, as far as I understand, there’s a lot of different companies kind of with, you know, different stages of maturity in this field.

Jason Adleberg – 00:15:40:

That’s so wild, man. Just insta-AI within, and like, is there from a developer perspective, right? Like I’m a radiologist, I have a bunch of images, like what are kind of like the foundational principles for what you want to do? Like, is there a number of images where things start to get much better? Like, are there tipping points? Are there like general rules that you should be following when you’re starting this process?

Chris St. John – 00:16:05:

Yeah, that’s a really, really good question. And I think it’s kind of an open-ended question. Like, I don’t think there’s really a clear answer, but some general principles are that the more data you have, the better. And the more like representations you have of a disease, the better your model will perform. So that’s actually something that I think having a clinical experience can really help you out as a radiologist. Because, you know, we keep talking about pneumothoraxes, right? But there’s like, you know, it’s not really just one disease. There’s different ways it can present. So for instance, you can have an apical pneumothorax, you can have like, you know, a relatively small pneumothorax at the top of your lung. You can have a tension pneumothorax, which is like really bad. Like your entire lung is just like shriveled down and it’s just all air inside your chest. You know, a lot of times pneumothoraxes are associated with rib fractures. So that’s also something that comes up a lot. And, you know, the more images you have, the better your AI model will perform. If you just create an AI model with tension pneumothoraxes, for instance, it may not be able to diagnose an apical pneumothorax. And then one other kind of issue that comes up a lot is that different people have different looking chest x-rays. So like between adults and kids, for instance, if you build a pneumothorax model that was trained only on adult chest x-rays, like that may or may not work. You know, kids’ chests are like smaller and, you know, look a little bit different. There’s a really interesting paper where someone built, I believe it was a pneumonia AI detector. They basically determined that there’s this really popular dataset online that people were using, like an open source dataset to build models. And this paper is really interesting. It discovered that like patients who are represented in this dataset tended to be mostly Caucasian, and they determined that patients who were not Caucasian, like the AI model didn’t work as well on them. And so that is something that comes up also a lot with AI. And this is, it’s such a new field that there’s really like, we’re still kind of figuring this stuff out. You know, if you’re buying an AI model or developing your own AI model, you want to be sure that this AI model has been trained on different demographics of people. So young people and old people, different ethnicities of people, you know, of course, male, female, and, just that everyone is sort of well represented because an AI is only as smart as the data that it receives. So if you’re leaving people out, it may not be able to make diagnoses on those people.

Jason Adleberg – 00:18:36:

Yeah, absolutely. And honestly, that gets me thinking too, like about, and once I still don’t know exactly what I’m talking about here, but you know, I’m thinking about like noise and image quality and how there is obviously like, there’s going to be variability in your images, right? You know, patients are going to be centered differently. They’re going to have different doses. The image quality is going to be higher. It’s going to be lower. Like in these early stages of building your own AI models, like are there concerns about image quality or noise affecting you know, the AI’s interpretation of whatever it is you’re training it on?

Chris St. John – 00:19:10:

Yes. I mean, that’s a really great point. That is something that comes up a lot too. A lot of times like AI, there’s this concept called data drift where you might have an AI model up and running at your institution. And then maybe you have a new x-ray tech who comes in, who does different parameters for capturing x-rays, or you have a different type of x-ray device, right? A different x-ray machine. You install a new one. And all of a sudden the AI doesn’t work as well. This is something that comes up all the time. You know, if an AI was trained on really clear images and you start giving it noisy images, it may not work as well. And honestly, vice versa too. If it’s trained on noisy images, the clear images, it just might get confused. That’s one of the big issues with AI is that it doesn’t really sort of right now still a black box. It doesn’t always explain how it reached a conclusion. And people are working on different explainable methods like heat maps, and, you know, other ways to explain what it’s thinking. But, you know, especially if you have a black box, like it’s not always clear why it got something wrong.

Jason Adleberg – 00:20:13:

Yeah, for sure. This is just me asking you for help at this point. Is there a good follow-up question to expand on data drift at all? Like, is that worth diving into a little bit more or not really?

Chris St. John – 00:20:22:

Yeah. I mean, you could say like, how do organizations avoid data drift?

Jason Adleberg – 00:20:26:

Oh yeah. Great.

Chris St. John – 00:20:27:

Yeah.

Jason Adleberg – 00:20:29:

In your experience talking about data drift, how are you finding organizations are adapting to it and handling it?

Chris St. John – 00:20:36:

Right. And I mean, I think, again, we’re still so early with implementation of AI. But this is a question that, you know, healthcare organizations, you know, should be thinking about is once an AI is installed, like, how do you make sure that the performance of the AI is just as good, you know, on day 100 is on day one? And how do you make sure that it’s agreeing with the radiologists or not agreeing with the radiologists? Like, what is the that number? What is that percent? As far as I understand, there’s some thoughts that the FDA might start to incorporate some of this stuff into regulation of AI. But I guess that’s also another issue is that because it’s such a new field, you know, the way that the FDA currently regulates some of these products may not be the way that it regulates it in 10 years.

Jason Adleberg – 00:21:23:

Yeah, of course. I mean, you know, what’s the curve? Moore’s law, right? You know what I’m talking about? Where like the number of…

Chris St. John – 00:21:29:

Yeah, like exponential. Exponential.

Jason Adleberg – 00:21:31:

Right. Yeah. But you know, the exponential growth of technology, your chips are… But you know, you can fit more and more and more information onto smaller and smaller and smaller devices. And with that, and with AI, it’s like the curve and the rate at which technology is increasing, it’s just getting faster and faster and faster and harder to keep up with. You know, and as someone, you know, you’re in radiology, you understand that things are increasing in the tech world at an exponential rate. Are there policy concerns? You know, you’re talking about the FDA. You know, do you have larger concerns about regulation and the speed at which they move versus the rate at which technology is increasing?

Chris St. John – 00:22:08:

I think one thing on the horizon in regard to AI technologies, this idea… So like right now, a lot of AI models just say yes or no, right? They look at an image, whether it’s an x-ray or a CT, and they say, yes, there is a blank, pneumothorax, intracranial hemorrhage, whatever. Yes or no. But there’s this idea that… In the future, it’ll actually produce a full report. It’ll be like knowing everything, and it’ll say there’s no pneumothorax, and there’s a small pneumonia, and, you know, this patient has an enteric tube or whatever. And, you know, if that’s the case, the way that the FDA currently regulates a lot of these products is that they basically want to see documentation just for one finding. And like every time you submit to the FDA, everything has to be centered around one thing. So if you’re trying to create a chest x-ray model that can detect 100 different things, that’s like 100 different submissions. And in the future, I mean, kind of seems like, you know, we’re moving to this world of AI, which is getting smarter and smarter every day, and it’s going to be able to produce a full report. Well, like… No one even really knows how you would like submit that to the FDA. Like how would the FDA even approve that?

Jason Adleberg – 00:23:24:

Yeah. I mean, it also just seems wildly inefficient, which, you know, I’m not going to touch that one too much. But, you know, I’m thinking about what you said about a chest X-ray. If you want to check for 100 different things you’re submitting over and over and over again, like, what do you think about the way that this policy like policy is being handled surrounding this? Is it remarkably frustrating? Like, do you understand, like, the political red tape of it all? Like, coming from a health care professional, but who also understands the value and relative accuracy of the tech, right? Like, what do you see as a smooth path forward?

Chris St. John – 00:23:59:

Like, how should the FDA regulate this stuff?

Jason Adleberg – 00:24:02:

Yeah, I mean, I certainly don’t have the answer.

Chris St. John – 00:24:04:

Yeah, I mean, to be honest, like, I don’t know either. I mean, I’m not going to argue that, like, they should be hands-off and let everyone run wild and just make any AI models that they want. Obviously, there’s an argument that they should be moving a little bit faster because you also don’t want them to be too slow. I think all of these issues, though, like what we’re talking about, pretty thoroughly convinced me that, like, AI is not taking over anytime soon. I mean, there’s just so many issues. And it’s one thing to design a model, but to implement it is another thing. And then to, like, get certification that this actually works is another thing. And then to make sure that after you install it, it is still working. That is even another thing. So there’s so many different topics as to why AI is going to take a while to, you know, have a huge impact the same way in other industries.

Jason Adleberg – 00:24:54:

Right. And it kind of like, honestly, it reminds me a bit of the cycle of academic research and papers, right? Like, you know, you want to write a paper on something. Okay. It takes, you know, X amount of time to gather your data and then X amount of time to analyze, write the paper, peer review. And by the time you’re getting published, you’re talking about data that sometimes is like almost 10 years old, depending, not always that long. But there’s this lag or drag that while understandable and the reasoning for it is sound, but at the same time, it feels like you’re kind of dragging your feet through this process.

Chris St. John – 00:25:32:

Yeah, I mean, I think the FDA was created because there was this medicine thalidomide, right? Have you ever seen the pictures of this? Thalidomide? Let me just Google this to make sure I’m not talking nonsense.

Jason Adleberg – 00:25:44:

Yeah.

Chris St. John – 00:25:44:

So my understanding is that one of the reasons the FDA was created was because there was this really popular medicine in the fifties that was given to pregnant women for nausea, thalidomide. And unfortunately it actually caused like pretty severe birth defects. And, you know, they like, of course they didn’t know that, but from what I understand back at that time, like they didn’t really have a robust process for doing like drug trials. So if they had done a trial, you know, they probably would have figured out that, you know, this causes birth defects, but that wasn’t even like something they were thinking about back then. So then the FDA was created and, you know, of course the FDA slows down like development of drugs, but I would argue that this is something we have to do, you know, as a society to make sure that the medicines and technologies that we are creating are safe and effective. The pictures are pretty wild. If you haven’t seen this, sorry.

Jason Adleberg – 00:26:35:

Oh, no, no. I mean, I love a little pivot. Granted, it’s not typically one that involves birth defects, but oh, yeah, that’s no joke.

Chris St. John – 00:26:42:

For sure, there are lots of very good reasons that the FDA exists. And, you know, with AI, the issue is that, you know, say you develop like some AI model that detects a disease and it’s just totally getting it wrong and there’s no person in the loop to officiate to make sure that it’s actually producing the right output. I mean, it’s not hard to imagine that would be a really big problem.

Jason Adleberg – 00:27:04:

Yeah, for sure. Let’s just talk to about, you know, the, for lack of a better word, these kids, you know, the future radiologists, kids in medical school, coders, what advice would you give to medical students or, you know, programmers interested in working specifically on radiology AIs? Like, where would you direct them if they were looking to get a foot in the door? Or where should they start playing?

Chris St. John – 00:27:29:

Yeah. So, you know, I think if you are listening to this podcast right now, and you are someone who is motivated enough to discover this podcast and wants to learn more about radiology AI, I think that you are probably motivated enough to learn some of the basics of programming, just the really simple parts. And, you know, I think I have a pretty good tutorial. I’m happy to take any questions or point people to that. But okay, if you don’t want to learn any programming, and that’s not something you’re interested in, I think it’s still worth it to understand the concepts. So, you know, how do AI models get created? Like, what are the things that you have to be thinking about when they’re getting created? Like we talked about different demographics, making sure different appearances of a disease, making different demographics. So if you’re someone who is not interested in doing programming, I still think that there’s lots of concepts that you should understand and that are relatively easy to understand. So things like we talked about with making sure that different types of patients are represented, different demographics, older patients, younger patients, men, women, different ethnicities. You want to make sure that different appearances of the same disease are well represented. So not just pneumothorax, but apical pneumothorax, tension pneumothorax, obviously both sides, left, right. I mean, lots of different appearances for the same disease. And I think it’s also could be really valuable to have an understanding of how these things get implemented. And to develop an understanding of how this gets implemented, you know, you should get some familiarity. What are the different systems that a radiology department uses like PACS? You know, your reporting system, often PowerScribe, protocoling system or RISC, lots of different systems that all talk to each other. And if you can understand like how these pieces all fit together and talk to each other, that I think is really, really valuable.

Jason Adleberg – 00:29:12:

Yeah. And let’s say that they do want to start messing around with coding. Are there like specific languages that you would recommend as like an entry point?

Chris St. John – 00:29:22:

Yeah. Everything I do is in Python. I think a lot of AI stuff happens to be in Python, but I would say like, you don’t even really need to get super into the weeds, like starting from scratch. I think that there are lots of good tutorials and lots of good YouTube videos that will explain like, here are the basic concepts of how to build an AI model. And here is like a notebook, like a collab notebook or some online resource where you can see exactly what I did and just click on the play button and like watch it go.

Jason Adleberg – 00:29:50:

Yeah. Once again, forgive my naivete, but, you know, I’ve been hearing my friends casually talking about working in Python for, what, 15 years or something now. Does it evolve or change at all? Or, like, is Python at its core like a stagnant beast?

Chris St. John – 00:30:05:

Yeah, it kind of changes, but, like, the basic principles don’t change. So little tiny things won’t change. But the basic grammar and like verbs and nouns are still there. Maybe they just add a few verbs and they add a few nouns every year.

Jason Adleberg – 00:30:21:

Right. Are you this is an analogy, right? Like when you’re saying nouns, verbs and adjectives, right? You’re using that as placeholders for like segments of code that achieve specific tasks.

Chris St. John – 00:30:31:

Yes. Correct. Yeah. Yeah.

Jason Adleberg – 00:30:33:

Okay, you know, the voice of this podcast is that I know nothing and I am trying to learn. So forgive my silly questions at times.

Chris St. John – 00:30:40:

I mean, that’s the whole thing. I really think programming is like, I think it’s an amazing thing to learn. And I think it kind of makes you see the world a little bit differently. Like at its core, programming is just about like being able to give instructions to someone about how you would do something. You know, I mean, I think it’s honestly a lot like cooking. Like do this, like this is the big thing you’re gonna do, which is broken down into little steps. And then while this is in the oven, also you should start cooking, I don’t know, start cutting up the vegetables or whatever.

Jason Adleberg – 00:31:10:

I don’t know. Right. So like, yeah, like larger workflow concepts for just being like the most effective version of yourself that you can be.

Chris St. John – 00:31:18:

Yeah, exactly. Thinking about like, what are the different processes, the different things that you’re trying to get a computer to do? And like, what things can be done at the same time? What things have to be done in order? It think program is really cool is just a way of thinking. I feel like it’s something I really like doing just for fun, even if really going anywhere.

Jason Adleberg – 00:31:37:

Yeah. And we are nearing the end of our time and I don’t want to keep Jason late today. But just as we wrap up, I’m just kind of curious, what are you most excited about in terms of the future of implementation of AI and radiology? And how are you hoping to contribute as you continue to grow?

Chris St. John – 00:31:55:

Yeah, definitely. I mean, I think that every year, like AI is going to get smarter and smarter. And I think that there’ll be a need for people who are able to think about how we’re going to manage an AI system that is being used in their practice or their hospital. So that’s something that I definitely would like to do in the future. And, you know, I mean, like every few years, there’s some new big thing that gets discovered, some new trend in the AI space. Like I mentioned, I mean, people right now in 2024 are working on these like image to text models where, for instance, you give it a chest x-ray and it produces a whole paragraph. So I don’t know, at this particular moment, that’s something I’ve been playing around with. And there’s like different open source implementations of that. But yeah, I mean, I just think it’s like a really cool time to be involved in radiology. And I’m really excited to see where the future goes.

Jason Adleberg – 00:32:46:

Absolutely, and before I let you go, Jason, where can everyone find your stuff, right? You know, you said you’ve got some tutorials online, you have some stuff. Where can everybody find you and continue to learn from your expertise?

Chris St. John – 00:32:58:

Yeah, sure. So you can certainly find me on LinkedIn, Twitter, and my homepage, which has links to all of this stuff, including the YouTube video, is pixelstopatients.com.

Jason Adleberg – 00:33:09:

Great. Awesome. Jason, thank you so much for joining us today on Frame by Frame. It’s been lovely to have you. And I’m sure we’ll be talking again sometime.

Chris St. John – 00:33:17:

Chris, awesome chatting with you. And thanks so much for having me.

Jason Adleberg – 00:33:20:

Awesome, thank you. Frame by Frame: Rethink Imaging is brought to you by Imalogix. Here, you’ll find engaging interviews with thought leaders, experts, and patients, sharing stories that showcase the transformative power of medical imaging. To discover how Imalogix is rethinking imaging in healthcare, visit imalogix.com. Be sure to subscribe to Frame by Frame: Rethink Imaging on Apple Podcasts, Spotify, or wherever you listen. And from all of us here at Imalogix, thanks for tuning in.