← Scrum Master Toolbox Podcast: Agile storytelling from the trenches
Scrum Master Toolbox Podcast: Agile storytelling from the trenches
BONUS The Future of Seeing—Why AI Vision Will Transform Medicine and Human Perception With Daniel Sodickson
00:37:18
Loading summary
Transcript51 lines
- [00:04]
Vasco
Hey there, agile adventurer, just a quick question.
- [00:07]
What if, for the price of a
- [00:09]
fancy coffee or half a pizza, you could unlock over 700 hours of the best agile content on the planet? That's audio, video, E courses, books, presentations, all that you can think of. But you can also join live calls with world class practitioners and hang out in a flame war free and AI slop clean slack with the sharpest minds in the game. Oh, and yes, you get direct access to me, Vasko, your Scrum Master Toolbox podcast. No, this is not a drill. It's this Scrum Master Toolbox membership. And it's your unfair advantage in the agile world. So if you want to know more, go check out scrummastertoolbox.org membership. That's scrummastertoolbox.org Membership. And check out all the goodies we have for you. Do it now. But if you're not doing it now, let's listen to the podcast.
- [01:12]
Hello everybody. Welcome to another AI bonus episode. And for this bonus episode we have joining us from New York, I believe, Daniel Sodigson. Hey, Daniel, welcome to the show.
- [01:24]
Daniel Sudiksen
Hello, Vasco. Great to be here.
- [01:27]
Vasco
Absolutely. It's a pleasure to have you here to explore a very interesting topic. So let me tell you a little bit about Daniel. He's a physicist in medicine and chief medical scientist at Function Health, previously at NYU and also a gold medalist and past president of the International Society for Magnetic Resonance in Medicine. He has pioneered AI driven imaging in medicine and he's the author of a recent book called the Future of Seeing. And I think this book is interesting because it might be that the next leap in AI is not about the size of the models or how to apply it to reason, but rather how to apply it to things that we are very used to and often depend on to do our jobs. Seeing. So today's guest argues that we are on the edge of a vision revolution that will change medicine, technology, and even human perception. Pardon me, itself. So stay with us. This conversation is perhaps a turning point for many of us in understanding the impact that AI can have. So, Daniel, back to you. You've spent your career helping people see things we literally could not see before. From inside the human body to patterns in hidden data. And in the book, the Future of Seeing, you argue that we're now entering a vision revolution powered by AI. So I would love to start at the beginning. When you look back, what was that first, I guess, career fork in the road that moved you towards seeing. The process of seeing as your life's
- [03:08]
Daniel Sudiksen
Work well, Vasco, I'd love to say it was this carefully planned trajectory that led me there. In fact, it was a complete and utter accident. And my father, who's a physicist, who has sort of followed his nose throughout his career from one interesting thing to another, has always told me to make room for serendipity, to keep your eyes open, as it were, for that happy accident. For me, I was doing a rotation at the end of medical school. I had finished graduate school, and I was in medical school. I was doing a rotation with a cardiac imager, a guy named Warren Manning, whose job was to image the heart. And he told me to write up a paper at the end of my month with him on anything I thought was interesting and let me loose on the literature and said, go study. And as I was studying, I realized that one of the real challenges, or in this case, MRI of the heart, was speed, because it's sort of considered bad form to stop the heart while you're imaging it.
- [04:13]
Vasco
Yes, indeed.
- [04:15]
Daniel Sudiksen
I started wondering, because I didn't know any better, well, what limits the speed of an mri? Why can't we go arbitrarily fast? And. And I settled on the fact that it's because we're actually gathering one point and one line of data at a time. We call our MRI machines scanners because they operate almost like digital scanners, like old screens, right?
- [04:36]
Vasco
Like CRT screens that drew one line on the screen at a time.
- [04:41]
Daniel Sudiksen
We sort of raster our way through the data. And it occurred to me, well, maybe the best way to image faster would be to gather multiple lines of data at time at a time to image in parallel. And I was doodling, literally on a napkin. I mean, I know it's a stereotype on a napkin in a piano bar in Boston, and came up with a way that I thought would work to hit multiple lines at once. Ran to Warren and said, hey, I have this idea. Never mind my paper. Can I come work with you when I'm done? And he said, who are you again? Sure, why not? And I did, and it worked. And that, for me, was the beginning. So I came up with this idea called parallel imaging and was inducted into the mysteries of imaging and all of the ways that we can sort of see what was once invisible. And, in fact, one cool thing for me was, in a way, this was a way we could make our artificial imaging emulate the eyes a little bit more. Because, of course, as you know, the eyes capture an entire scene all at once. And that was part of the inspiration. So this connection between natural and artificial vision, I think, caught me, and I was caught for life.
- [05:55]
Vasco
And we'll dive into that in a little bit. But I'm interested. Do you remember the first time that you used that, let's call it acquired superpower to actually look at an image and discover something that changed the outcome for a real patient?
- [06:13]
Daniel Sudiksen
Well, I have to say that one of the beautiful and powerful things about medical imaging is that those examples are routine, not rare. And for the book, I studied the history of imaging, and essentially every time a new imaging modality, like x rays at the end of the 19th century and then cross sectional imaging, MRICT, PET in the 1970s came along, they were instantly adopted because basically, when you can see something that you couldn't ever see before, it adds value. So in that first job I was telling you about, we were imaging the heart all the time. We would encounter people with narrowed coronary arteries who are at risk for a heart attack, and they would immediately go in for treatment with a cardiologist, maybe to get the arteries opened up with a stent or maybe to go on medical therapy all the time. Now I work with people who have potential tumors that are found. And one of the key things for cancer is to find it early. And so I know many people who have been on these screening programs, for example, for prostate cancer, to check year by year, do they have anything, Is it growing? And not just once or twice, but routinely, we find something that then tells somebody, okay, now's the time to go in and get it treated. So really remarkably routine for me.
- [08:02]
Vasco
What's interesting is that as you studied this, I think it was 4 billion year arc of the history of vision, right? As you studied this, one starts to recognize that the power to see something differently allows us to discover not only the diseases and potential problems like you described, but also the right questions, right? Because imaging a heart for the first time, I wasn't there. I don't know how it went, but I can only imagine that the first time somebody could image a living heart, because dead hearts we would have been able to see many times, right? But the first time somebody could image a living heart without opening the patient because it's an extremely dangerous surgery, that created a lot of questions because now we could do prospective imaging, right? Like looking for things, we could do studies based on imaging. How do you feel that? Because you were working on this imaging aspect of medicine for so long, I bet you must have had these points where it was like, oh, now this is possible, right?
- [09:07]
Daniel Sudiksen
Absolutely, absolutely. I mean, I'll start with the historical story, but I'll end with me. The Copernican revolution, the fundamental revolution that shifted us from sort of an egocentric, geocentric worldview where we were the center of the universe, to the notion that we're on this tiny little rock orbiting something else. That basically arose from imaging. It was tracking the path of planets through the sky using first our own eyes, then enhanced vision with telescopes. That led us to that fundamental change in perspective. And I think every time a new imaging tool arises, not coincidentally, that's when discoveries arise. So I'll give you one small example from my own career. We were imaging with stronger and stronger magnets, because in mri, the stronger the magnet, the better the signal you get. And we kept finding that there were these perturbations, these artifacts, these sort of distortions in the image that were bothering us. And we were trying to get to the root of them. And eventually we realized it was because when you get high up enough in frequency, the body's electrical properties start interfering with the imaging process. And so that was annoying. That was a problem to deal with. But at some point, I realized, wait a second, the body's electrical properties are affecting the image? Can we make an image of the body's electrical properties? How well it conducts electricity, how well it stores and separates charge? This launched a whole research area where we did electrical imaging to try to map out conductivity, the ability of the body to carry current throughout the body. And sure enough, we were able to create maps of new information that we'd never been able to see before by turning errors into signal.
- [11:14]
Vasco
And actually, that leads me to the next question, because when you start thinking about it, those weren't errors. They were errors from one vision technology perspective, but they were information from another vision technology perspective. And when people hear about AI vision, they probably imagine that somehow the AI system is looking at something similar to what we would see. But that's not the case, right?
- [11:38]
Daniel Sudiksen
Like how.
- [11:39]
Vasco
How do machines actually. AI systems, how do they see things that humans simply can't?
- [11:47]
Daniel Sudiksen
Well, that's a very deep question and a very interesting one. What I would say is they can see in ways that are analogous to the way we see. So when the first convolutional neural nets were developed, when Yann Lecun and others first came up with these, they actually modeled them on the visual system of animals, like the cat. Interestingly enough, this sort of convolutional structure was meant roughly, to emulate a biological vision system. And so in many ways, we can take an image or machine neural nets can take an image, distill it down into kind of higher level features, distill those down into still higher level features. And there's sort of this hierarchy, like what happens in our brain. But at the same time, we have been beginning to realize now that we don't need to start that process with images the way we see them. We can start that process with the raw signals that are coming off of the MRI machine, which normally we need a, you know, mathematical transform to convert. We can start it with raw signals coming off of sensors or signals coming into telescopes or microscopes. We don't actually need to construct an image. In other words, to feed into the neural net, like the retinas feed into our brain. We can start with raw data. And one of the even more exciting things that my group has been exploring in recent years is sometimes we need dramatically less data than we think. We don't need a nice, crisp, perfect image. A properly trained AI system is capable of basically filling in the gaps. And so we can do all kinds of things that we never would have imagined with AI by letting it sort of run free. We can, some ways, do imaging without the image.
- [13:55]
Vasco
Yeah, and actually that is the point that I wanted to get to, because when you start thinking about this,
- [14:02]
Daniel Sudiksen
the
- [14:02]
Vasco
example I usually talk about when we talk about vision is the Matrix, the movie, right. Where you needed to learn to see something that was beyond the image that the eyes could see in order to see the Matrix in the movie as a metaphor. But the same thing happens to us a lot. Like, for example, in my field, where we do a lot of process work in software development, we need to learn to see the process, but the process isn't visible through our normal optical means. The process is only visible through tools like value stream mapping and boards and graphs and all of that stuff. So properties of the system become the way to visually understand the system. On the other hand, those aspects make it even more critical to understand why the visual representation of something, something is so important. So for me, what's interesting in that intersection, that understanding that, yeah, we can just extract some data and the neural nets or whatever AI system will be able to see something that we wouldn't see with our eyes. Where do you then start making the difference between are we creating an image for us or for the AI system? Like, how do you handle that?
- [15:17]
Daniel Sudiksen
Yeah, yeah. Well, I think that leads me to an area I've been thinking a lot about when it comes to AI in imaging, but also in general, I Think. Most often when we envision AI, no pun intended, we think of it as sort of this downstream process. We generate our data, we make our image, we have our sensor data, whatever, we put it all together and then we let AI loose instead of our brains. And it's sort of a competition between the human and the machine, right? I mean, so much of the discussion about AI is what jobs is it going to replace? Is it doing better than this specialist or that specialist? To me, that's a little bit limited. Why are we limiting ourselves to downstream? Why are we limiting ourselves to tasks that humans can already do that we want AI to do better? Why aren't we thinking of tasks that the AI can do like that, as you were saying, no human could ever do? So, for example, if we know we have a machine learning system on the other end, why don't we change the data we gather? Why don't we rebuild the machines we use to gather the data? And that's some of the things I've been starting to think about more and more. For imaging, rather than designing machines that generate a nice, pristine, clear image that is perfect input for our eyes and our brains, why don't we generate machines that might be much cheaper, much more accessible things that you could build into a seat or a bed rather than a huge phone or a mobile phone. Right. Rather than a huge expensive tube. And could we maybe not get a perfect pristine image, but could we measure whether you have changed since the last time we saw you in one of our big tubes that would allow us to have a continuous change monitor that traveled with us rather than having to come back into a big expensive MRI at regular intervals. And maybe we could have a health monitor that was sort of a safety net for you, an early warning system that no human would be able to interpret, but a machine putting it in context of all your other data would interpret perfectly. So that's an example of what I call upstream AI changing the data we gather based on the capabilities of our machines?
- [17:50]
Vasco
That sounds very interesting. It sounds also that you're kind of always from the perspective of the field of medicine, but it sounds like you're going to to this experimentation side more like going from a traditional scientific discovery approach to more like a modern product development rapid iteration. Is that what you are doing at Function Health? Yes.
- [18:13]
Daniel Sudiksen
So that, in a nutshell, is actually why I joined function health. After 29 years in academia. I've had all kinds of lovely opportunities in my career for sort of free discovery to explore what the images are showing us to Follow the leads of these sort of strange bits of noise that turn into signal. But more and more I've started to wonder over the years as I've worked with these amazing MRI machines why we use these amazing devices only once we already know you're sick. Why do we use them retroactively, reactively rather than proactively? And once I started thinking about that, I started wondering how we could change both our machines and our AI tools to enable proactive health. Eventually, function came along and called my bluff. I had written about this at the end of my book. This notion of proactive imaging and proactive health and function health does just that. It measures large panels of blood tests repeatedly along with MRI scans and other imaging in otherwise healthy people to provide that early warning and guidance. And my, my mission now is to figure out how we build the AI systems that take all of that proactive data and turn it into kind of a GPS for your health.
- [19:43]
Vasco
So this sounds very familiar for all of us who work with large distributed systems because we've building what we would call observability on the technology side, observability subsystems that allow us to detect problems before they become catastrophic, or even to create. These days we call it anti fragile architecture. So like self healing architectures and so on. But I would like to return to that idea of the data that you are now collecting, which may be the same or different from what you were collecting before, and how the use of AI with these data sources or signals is actually helping us to perceive things differently. Like you were talking about proactively using data to detect potential issues before they happen.
- [20:32]
Daniel Sudiksen
Yes.
- [20:32]
Vasco
So how are you planning to do this? Because I can only imagine that, I mean, I'm sure there will be studies and so on, but you're almost kind of looking at an MRI and saying, okay, MRI maybe a great way to image the human body, but because we can't use it, we need to find other user, other ways to image it. And they need to be at least as good as the MRIs. So like, is it, is it only a question of finding different signals? Is it different signals with different technology? Is it completely different models of thinking about how we perceive the human body? What's going on?
- [21:09]
Daniel Sudiksen
That's a great question. And I'll, I'll give it to you in one word. It's all about context. So it's not necessarily entirely new signals, but it's about taking those signals, images, blood tests, wearables signals, wearable sensor signals, other things, and putting them in context over Time and in context of each other. And, and I'm going to go back to a biological analogy here. So it turns out that we think of our biologically evolved vision as just a sort of plug and play thing, right? We open our eyes, we see the world, we get a picture of the world. Not so it turns out the world we see is a lie. Everybody talks about depth perception. Oh, I have two views of the world and I, you know, I compare them and that way I can judge depth from an imager's point of view. That's nonsense. Two views is not nearly enough to figure out exactly where everything is in space. We have to gather all kinds of views in a CT scanner, for example, to sort out what's where. It turns out what the brain is doing is using everything it's learned about the world, about how the world works, about how big people are compared to objects, how light falls on surfaces. And it's filling in what's missing from those two views to give us our picture of the world. In other words, our brain is putting today's signal in a very rich, learned context. We can do the same thing with AI. So we might have a signal today that's a sort of not so great MRI of you. But if we have previous images of you and we have lots of blood tests and other signals, what AI systems can do and what we're training to do is to detect from today's signal whether you look like yourself or whether you've changed in any worrisome way from everything we know about you. That's a much easier problem than creating a pure image just off of today's little sliver of data.
- [23:24]
Vasco
So it's like using the AI as a way to codify potential contexts and then analyze signal data to kind of create an image based on that limited signal data plus the codified context. Is that what you're trying?
- [23:39]
Daniel Sudiksen
Exactly right. Exactly right. There are some transferable concepts about how brains look or how hearts look, or how certain types of signals evolve over time. If we use those principles that we've managed to train our network to learn in combination with even an incomplete set of data from our measurement today, we can do a whole lot more.
- [24:13]
Vasco
We could just, based on this question alone, we could do a whole episode on the philosophy of seeing. Because what you're describing is something we've been struggling with as humans. Even deciding what is real has been a problem within the realm of philosophy. Right. And a hobby of mine is psychology, and it's still a big problem in psychology. What we see is not really what is there.
- [24:38]
Daniel Sudiksen
That's right.
- [24:40]
Vasco
But one thing coming back to the technology, like when you describe this idea of codified context and expectation versus observed data signals, and then what that could mean, it sounds to me that this has applications that go far beyond medicine, like satellites, phones, cameras, security cameras, infrared cameras, like, I don't know, nuclear power plants being monitored by cameras that that signal has a lot of information there. What do you think is going to be the unintended consequence of this new hybrid human machine perception system that you're building?
- [25:21]
Daniel Sudiksen
I think it's going to allow us to do a whole lot more with a whole lot less. I think it's going to in some ways. Well, here's one way of framing it. It's a way of giving all of our imaging machines memory. And I don't mean just like computer storage memory. I mean memory of the past. If you think about it, how much of our lives is dictated by our memories? We carry them with us as we go. They sort of accumulate, they build up, they inform who we are and the decisions we make. Because we've seen certain things before, and therefore we act that way. And we can make a snap decision off a tiny little bit of data because of the wisdom we've built up. Now imagine we have effectively machines with wisdom, machines that essentially have identified temporal patterns that know how things have gone in the past and can predict what's happening next. And I think in the book, I talk about a number of kind of crazy examples, like could we combine all of the cell phones on Earth into a massive radio telescope that could essentially see farther than the Event Horizon Telescope that has given us our first view of black holes. The Event Horizon Telescope is actually a bunch of distinct observatories, all exquisitely coordinated in time to synthesize something with a dish size, effectively the size of the planet. But cell phones are also electromagnetic detectors, and there are how many billions of them on the planet? If we can figure out how to interlace all of the signals from those phones, could we be essentially, as a planet, collectively looking out at the universe with all of that individual power marshaled in coordination? So there are crazy things you can think of that you could do with this kind of contextual reconstruction.
- [27:40]
Vasco
One wonders if seeing is now a too reductionist word once we start understanding this.
- [27:49]
Daniel Sudiksen
I think so. I think I've spent a fair amount of time, as you might imagine, trying to figure out what imaging really is. There are various definitions. I think, in one very practical way, I think of it as spatially organized information. So really what we're doing when we're imaging is we are taking in organized information. Now, what I've just been talking about is we don't now need to just organize it in space. We also need to organize it in time. So really what it is is a new way of gathering information about ourselves and the world and universe around us.
- [28:31]
Vasco
Information is a very loaded word for us computer scientists. And I would like to explore that, but I think. I fear it might be too deep a rabbit hole. One question, a personal question. Daniel, if I may. You've been working on this aspect of seeing, imaging, understanding, perceiving for so long. How has your own way of looking at the world changed over the years? Like when you think about how you notice things around you, whether it is in a patient or just naturally, as you are tourist thinking about Helsinki like you were some time ago, what do you do now that you didn't do before? And how has that changed your perception of the world?
- [29:12]
Daniel Sudiksen
Oh, I really love that question. I think partly what it's done is it. It's made me more aware of the way I'm seeing rather than just what I'm seeing. So sometimes, actually, when I'm walking along, this weird thing happens. It could be walking through a familiar street in New York City or by my home up in Westchester County, a little bit north of the city. And all of a sudden, everything I'm seeing just sort of fades away. And what I see instead is how I'm seeing. I sort of imagine light bouncing off of things and landing in my eye. I'm almost tracing the sight lines as the light's bouncing. And I imagine almost this hum, this buzz of light that's zipping around as fast as anything in the universe can go, right? Light is the very fastest thing we know, and it's zipping around and probing the environment and coming back and giving me news of it. And sometimes I get lost in this fugue state for just a few seconds. It's almost like I'm deconstructing my own vision. But in a lot of ways, that's what's going on. That's what's happening in our brains. And I find it kind of miraculous that we've evolved this ability to take essentially any probe, electromagnetic, vibrational, tactile, you name it. Essentially any probe of the world around us. We've one way or another figured out how to turn into spatially organized information to make sense of this kind of chaos of sensation. And to me, that's really kind of a miracle.
- [31:11]
Vasco
And it's so poetic, even though it's all about understanding the world and in our case, also technology. So the book is the Future of Seeing by Daniel Sudiksen. And Daniel, besides your book, and the link will be in the show notes for our listeners to easily find it. Besides your book, what do you think could be another resource on the future of perception or AI and imaging that you think we should explore if we want to go deeper into this topic?
- [31:43]
Daniel Sudiksen
Very interesting. Well, I will mention one book, not so much on AI or even on vision alone, but on animal senses that I think a lot of people are familiar with. An Immense World by Ed Yong. I found it utterly fascinating because it delves into how other creatures perceive the world. And what you learn is they in fact, have learned to use almost any sort of informative probe, as I was saying. So that's one. Otherwise, I mean, this is going to sound a little bit boring maybe, but I'm a great fan of the work of Yann Lecun and some of his like there's an early paper with Yann Lecun, Yoshua Bengio and Geoff Hinton that lays out sort of the rudiments of deep learning. And I still find, I think it's a nature paper and I still find that paper to be utterly fascinating. Okay, I'll give you one more now that I think of it. Speaking of Yann Lecun, so he has an article called A Path Towards Autonomous Machine Intelligence which talks about his view a little bit in opposition to what sort of the standard view in Silicon Valley right now with large language models and sort of scaling them. His view that in some ways we need to build something configured a little bit more the way the brain is configured in order to really generate artificial general intelligence. And in some ways, I'm actually using that as a model for some of the predictive imaging approaches I'm trying to build right now. So A Path to Autonomous Machine intelligence by Yann LeCun.
- [33:37]
Vasco
All right. And we'll put the link to all of those in the show notes so that our readers can, our listeners can easily go and read. And how about you, Daniel? If people want to know more about you and the work that you're doing, where should they go?
- [33:52]
Daniel Sudiksen
Well, certainly, I mean, I poured a lot of myself into the book. This has been a story I don't know, about 29 years in the making. So that's certainly a good place. But otherwise, I'd say if anyone sort of sorts through the papers that I and my lab have come out with over the years, you'll find this trend from sort of making imaging better faster to sort of figuring out, hey, how can we use imaging differently? So yeah, but, but but a good starting place is the future of seeing
- [34:41]
Vasco
certainly absolutely an inspiring conversation. Thank you for being with us, Daniel, and thank you very much for your generosity with your time and your knowledge.
- [34:50]
Daniel Sudiksen
Thank you so much, Vasco. I've really enjoyed the conversation and sometime I'd love to double click on those points of philosophy with you more.
- [35:02]
Vasco
All right, I hope you liked this episode, but before you hit next episode, here's the deal.
- [35:07]
This podcast is powered by people like
- [35:10]
you, the members who wanted more than just inspiration. They wanted real tools and real connection to people who are practicing Agile. Every day we're talking access to over 700 hours of agile Gold, CTO level Strategy talks, Summit keynotes, live workshops, E courses, Deep Dive interviews, books. And if you're into no estimates, we got the pioneers of no Estimates in those Deep Dive interviews as well. Agile Business Intelligence, creating product visions, coaching your product owner courses, you name it. You'll get invites to monthly live Q&As with agile pioneers and practitioners, plus a private Slack community which is free of all of that AI slop you see everywhere. And of course, without the flame wars, It's a community of practitioners that want to learn and thrive together. It's the best place to connect with community and learn together.
- [36:18]
So if this podcast has helped you
- [36:20]
before, imagine what you will get from this podcast membership. So head on over to scrummastertoolbox.org membership and join the community that's shaping the future of Agile. We have so much for you, so check out all the details@scrummastertoolbox.org membership because listening is great, it's important. But doing it together, that's next level. I'll see you in the community Slack
- [36:54]
we really hope you liked our show. And if you did, why not rate this podcast on Stitcher or itunes? Share this podcast and let other Scrum masters know about this valuable resource for their work. Remember that sharing is caring.