How is mama's mood? Should I come down to greet her or stay in my room? And that should not be happening. Like radiology or trigger practice should not be impacting your kids like that. And that is what I'm trying to solve. So I have daily reminder at home that how bad this has gotten, how worse it is becoming. The world has changed. Clinicians using AI tools will thrive. Clinicians resisting it will not, right? It's like cancel later. You're not gonna sit and do calculations with pen and pencil anymore.

Yeah, you can but that doesn't mean that's gonna be time efficient for you. We had a 40 times improvement in algorithm performance, not 40 percent, 40 times and 40 times improvement in GPU compute performance. That's 1600 times improvement in one year. So can you imagine, we're sitting in 1900, Model T just come out and you're asking me to predict what a thousand times better car would look like. I would have no idea.

Welcome to Signal and Symptoms Podcast. I'm extremely grateful today that my good friend, Dr. Khan Siddiqui, and my mentor and teacher is here with us. There's a whole episode that we did with other podcasts that how his journey came from being a physician in Karachi to come here, become a radiologist, and become basically one of the most renowned names in radiology AI. By the way, he is a serial entrepreneur, has had Two exits that I know of, maybe more. As a matter of fact, Hoppr, latest company, is probably doing one of the most exciting works in radiology AI. So going back to Dr. Siddiqui, sir, do you mind introducing yourself and more importantly, introducing what you're up to with Hoppr?

So as you mentioned, I'm a radiologist. I was faculty at UC Maryland at Johns Hopkins. I got into machine learning, computer vision AI in 2004 time frame. So before people even talking about this stuff. And based on my work in academia, I got a career at Microsoft to start their machine learning group. So all the work coming out of Microsoft, Myra and Manivision site, all this that group I started in 2008 worked on a of different things in healthcare as well as Xbox Connect and then eventually Microsoft Azure. Left to start a bunch of different companies, to do hypermation management and a company called Higgie had a successful exit of that and then decided to build a portable MR scanner, company called Hyperfine to become public on NASDAQ and have been investing in former employees and colleagues launching different entities and as you said, multiple different exits. From that, the Hoppr idea started at Microsoft in 2012 when we were kicking off our ResNet project. So the idea really was that the way I was thinking about is that as clinicians, when patient comes in, they don't come with a diagnosis. They don't come in and have pneumonia, find me pneumonia, or I have lung cancer, find me lung cancer. Like I have hemoptysis. Now you need to go figure out what is causing hemoptysis, right?

The thought process was that these model needs to do a lot of different things. And how do we build a platform to accelerate the AI development? Can we do it in a way that it doesn't take months and years to build a model? Can you do it in a day? That was kind of the philosophy in 2012. Microsoft didn't go in that direction at that time. So the philosophy was, can we build a platform to really accelerate the AI development? And can we eventually go to a codeless? Our low code development process. And as you know, the first thing you have to do is figure out how to get data and get data that can be generalizable across the place. The other thing you really need, even in CNN days, so the idea was like, what are those pre-trained models that we can provide as part of that aspect of it and all the tooling that goes on top of it. My philosophy always was that, it needs to be in the workflows and in the tool people are already using and not something distracting the clinician from their workflow, right? And then how do we provide easy tools to fine tune? So you don't need a ML machine learning PhD to build models. My eventual goal was the clinician you can build model to without, you know, knowing any coding kind of idea. So that's where Hoppr was. At this RSNA, we finally launched our codeless user interface. So you can now train the model for a classification task without requiring any knowledge of Python or anything like that. We also launched our delivery mechanisms, so findings indirectly into reporting applications called Presto, where a vision language model or any classification model findings can be integrated directly into PowerScribe 360, AI, any of the reporting parts that exist. So not building a reporting package, but solving both problems, getting models built to show in imaging solutions and then getting those results into technical workflows. So that's where Opera is.

So again, I'm extremely grateful. I'm gonna let Ed or Harvey, whatever, start with the first question and then we'll go from there.

So when you ran into obstacles, which I'm sure you did as a pioneer, Dr. Siddiqi, how did you overcome those obstacles? Like what was sort of like internal impetus to like keep pushing through?

That's a deep question. I'll tell you my internal motivation or my inner why is like, like to solve problems that people I love have and create a social impact at the same time. Right. I think we did this in the last podcast. So I'll show him a text message from one of my former fellows who's a professor at Penn and his text started, it was June, 2024, I think. They started like today. My 8 hour shift kicked my ass. I read 145 cardiac CT scans in one shift. That means he looking at 5.5 images per second. And that is insane. The max I've ever read was 80 in a shift and I had to sleep for two days to recover from it. Like he did almost double that amount. And I was like, how do you make sure you didn't miss anything? He's like, I have no idea. The turning point for me was when both my kids were, I think this is 2017, 2018. So we have a boys chat between the two boys and me and we heard the garage door open. So younger one was sitting in the kitchen, the elder one is room, I was in my basement office and the elder one texts, how is mama's mood? Should I come down to greet her or stay in my room? And the younger one goes, stay in your room, right?

And that should not be happening. Like radiology or trigger practice should not be impacting your kids like that. And that is what I'm trying to solve. So I have daily reminder at home that how bad this has gotten, how worse it is becoming. Can you imagine detecting pneumonia and things like that six days after imaging is done? This is insane what is happening. If you go to the ACR's career website right now, they 2,500 radiology jobs open.

We only train thousand rounds every year and CMS just reduced that number to be missed in the cuts that just came in. And we also got another 2.5 % reimbursement cut on all non-time based imaging procedures. So this is not sustainable, right? We don't have the human workforce to do this. At the end, patient is getting impacted, right? Yeah, you can talk about whatever it is at the end, who suffers the patient who suffers, right? And know, delaying diagnosis on pneumonia or things like that.

These are acute conditions and the more you delay, the more complications you have, the more patients suffers. So that's the motivation, inner motivation. And I always feel like, you know, I don't ever solve everything. always, my philosophy is hire people who can teach you and not tell them what to do.

Yeah, I love that. And a couple other questions is, so the first question is for clinicians, but then I have another one for you on the administrative side that's similar. If you're a clinician listening and you're like, oh my God, that's me. What he's talking about, what Dr. Siddiqui is talking about, that's me. I come home all ruffled up, I'm stressed. What should they do? So these are practicing physicians, not entrepreneurial physicians like that are represented right here. What would you say might be the first step that they might think about doing? Action taken.

I think action taking like literally what we just did, right? What is happening in the industry? Yeah. The world has changed. Clinicians using AI tools will thrive. Clinicians resisting it will not, right? I always, you know, when I talk to folks, it's like cancel later. You're not going to sit and do calculations with pen and pencil anymore, right? Yeah, you can, but that doesn't mean that's going to be time efficient for you. How do you do this? in a more efficient way. And that's what needs to happen. And again, without educating yourself about what are the limitations, how to do this stuff, what is going on, listening to podcasts like we are doing and others, it is really hard, right? All of us have our own massive intelligence in our pocket now. That is kind how to think about it, right? This is democratizing intelligence that now you have your own PhDs in your pocket that you can use. So the question becomes like, who do you trust? How do you...really be able to use those things, right?

Yeah, that's sage advice. And of course, from my perspective, know, go talk to your CMIO, your CNIOs, your CTOs, your CTOs, whatever, C. But if you do that and run into a barrier, you got to keep pushing and doing whatever you need to to bring this to practice. So the second, so the corollary is you're at the boardroom talking to a health system. So I sit on a couple of boards for health systems. Most are not clinicians. and many are from finance, banking, lawyers, all great people. So I'm not saying anything negative, but there may be a little bit of a lack of understanding of AI and technology and things that you bring to bear. So what would your message be like? What would you sort of encourage them to do?

That's also a very great question. The way I think about this, to be honest, what you just asked, I have actually the board folks ask me also, what do we do? Average time a commercial consumer model stays state of the art is nowadays seven days. Was 53 days in the summer, but with Opus 4.5 coming out in Gemini and seven days later, 3 comes out and OpenAI 5.2 comes out.

It's crazy how fast things are moving. So by the time you approve a tool, and if you look at like, I mean, even look at medical imaging related stuff, There were 52 medical imaging foundation models that came out in 2025. By the time you will approve one through the government process, it'll already be obsolete. One of the things to think about is that the application built on a model today is the worst version of that application. Because the next model coming out in a month from now will outperform that model, right? So in my philosophy has always become make sure you have a very good validation and continuous monitoring practices in place. It's not upfront governance, it's but monitoring as it's going on. So you understanding any aspects of drift that is happening. So that discussion is also happening at the regulatory level side of things. Still to change that factor because it's going to be hard to approve models by the time you get that approval, then it's already obsolete in this day and age.

Obviously the next question is, where are we going? What do you see five years from now? I know that's predictability, who knows, but I would love to see from your eyes, based on what you're doing, based on what's coming and what you can share, where do you think we'll be in medicine? At least in your area.

So here's, let me give you some to think about. The fastest car that we have today, what, 1,000 horsepower or something like that. So from Model T, which was one horsepower car, Ford Model T that came out in 1900s, to now we've been using 1,000 Ford increase in horsepower, right? Now, if you look at the architecture performance improvements in 2025 and the new chips Nvidia is putting together right, the black well and going forward. We had a 40 times improvement in algorithm performance and not 40%, 40 times and 40 times improvement in GPU compute performance. That's 1600 times improvement in one year. So can you imagine, we're sitting in 1900, Model T just come out and you're asking me to predict what a thousand times better car will look like. I would have no idea. Airplane hasn't been invented yet, right?

So models coming out next year or in late 2026 would be way better performance than what we have today. And how do you imagine that is really hard for humans? So that's how I see it. That's why this whole idea of how do you monitor continuously validate is way more important than worrying about, you know, what is good today and not. Like lot of the things that I can't talk about that are happening are just insane. What I'm seeing is just insane among the performance and models.

We went from paper charts to EHR to Epic when I was in St. Louis as a resident. I saw a lot of my senior physicians who were close to retirement, but not retirement age, just took early retirement not to go through that process. And now we're changing from what we call AI transformation. How do you address the human behavior and human adoption issue?

So there's a whole discipline for this called change management. So when you're implementing technology or new workflows, new things, you have to address change management. If you don't, then you're not exactly talking about what is going to happen. So if you have not thought through how to bring these folks on board, I was lucky to be part of the early digital revolution, converting film to filmless and guys in your health system and went through the change management process and had amazing mentors who kind of guided me through that whole process to do this. I mean, since then, you know, I've done a bunch of epic implementations at Hopkins and Maryland, PACS implementation, all this stuff, and didn't have anybody quit because we addressed it. How do you think about change management? How are you understanding the workflow? Where are the disruptions going to happen? And how do you address those aspects of and where...where the frustration comes from, right? So understanding those aspects really becomes important. I mean, that's my philosophy that AI needs to be almost like, if you look at our product Presto, you don't even know it's running, right? So it's exactly to already using same functionality of your existing reporting tools, it's just getting the model results right into your templates that already exist, right? So you're reducing the barrier for change management on that aspect.

Now, Dr. Siddiqui, you probably do sit with, I don't know, way bigger people than I do. What do you think that the regulation of FDA will look like three years from now, given all of the advancement? Because it took us five years to get to multi-diagnostic detection. This is the first time it happened, like, late last year. But what is your prediction in terms of regulation and how important it is?

So yeah, a couple of things. So first of all, FDA clearance depends on the indication of use, as you mentioned, right? That's all that is, right? So indication of use is governed by the application that the model is being used in. So they don't approve models, they approve applications being used for clinical care of any sort, right? Be it segmentation, as you said. So it's all about the indication of use and various different guidelines and regulatory process have been defined based on how those approval happen. The way it happens is there's a completely brand new indication, a brand new use case, usually goes through a de novo process. And de novo process then leads to a creation of regulation on how they will be assessed going forward. If you talk to any of the FDA senior leadership, they will agree with you 100 % that this model was initially designed on top of a drug approval model in the early days of CADE, which was very deterministic. know, weren't machine learning models, breast CAD is how all things started. And what has happened is because of the risk and cost and all this stuff, lot of people have built applications using those old frameworks to get it through the FDM, right? So now the problem is for this to change, requires Congress, you know, legislation changes. And I don't know, with all the pressures, a number or MTA, this is the right time anybody is going to go through and try to change the regulations.

Harve, what do you think? mean, do you think that Trump is gonna sign an executive order and get rid of some regulatory hurdles? That's possible. I'm just messing. We are, by the way, Dr. Siddiqui, we talk very openly. mean, people need to learn from us. so what do you think? I mean, I really think the administration's push on AI might end up rather than an act of Congress that executive order.

Yeah, my commentary was just what we all know, especially those of us, all of us who've been around for a while. People will find ways to optimize their workflow. And you can do it through a structured way, which has a lot of good benefits of doing it, very ideal, know, make sure everything's safe, all that kind of stuff. But if that takes too long, which is historically… the problem, which is why I asked the question about the boardroom. know, we put up all these committees on slowing things down to the point of detriment. And then it sort of forces people to do kind of their own maverick things and which again helps push innovation. So I'm not against it at all. helped mavericks.

And now the question is, because I have to compete in the stupid market, which is a market problem, right? Because if I was going to ethical, which now I feel like, should I?

It's lesser evil, right? What do I do? I have all these doctors and in the freestanding world, it's sort of worse because at least in the main hospital, you have a big mega hospital system and there's all these radiologists and if push comes to shove, I can walk over to the radiology department and be like, dude, hook me up, read this x-ray, I'm worried about X. When you're in the freestanding, as far as I know, this could be like another side of the planet, it could be somewhere else, I don't know who's reading this, I don't have that relationship. And so I do have to make a cold read.