Welcome to Nvidia's AI podcast, where we explore the expanding world of artificial intelligence with the people who are developing and applying AI to the problems and opportunities that cut across our lives.
And living healthier lives is certainly one of the areas where AI is increasingly brought to bear, which is where our guest, Dr. Bradley Erickson, comes in.
Dr. Erickson is a professor of radiology and the associate chair of research in radiology at the Mayo Clinic in freezing Minnesota.
Brad, Dr. Erickson, welcome. Thank you. You focus your research on the use of computer technologies, like AI, I guess, but to extract information from medical images. is.
I know AI is a relatively new user, correct me if I'm wrong for radiology and looking at medical images, but what's been the progression and how did you get to AI as a as a tool and an approach?
Sure, so radiology has had access to digital imaging really when CT scanning became available and the first CT scanner was actually installed here at Mayo Clinic in 1973.
But for a good 20 years, we always produced film images as the output.
Now, in the 1980s, people started to play around with research, and actually I did my PhD during that time.
We did do some AI, which was really more pattern recognition, that is really quite simplistic by today's standards, but certainly there has always been a lot of interest in applying machine learning and artificial intelligence to medical imaging.
Back in the 80s, I'm just curious, did you refer to it as machine learning and AI?
And I know that I know the discipline and the ideas have been around since the 50s, but was it working?
And what was the hope? And are you finally reaching that hope?
Yeah, so it was machine learning and AI.
And I remember part of my thesis work is we played around with these things called neural networks.
But quite honestly, they just didn't work very well.
And so it kind of died a slow death, but there was always a little bit of interest in applying machine learning Certainly we knew there was information there and it's important in medicine to detect all the information that's there and to reliably and consistently apply the information.
I think it had always been a pipe dream to use computers to help with image interpretation, but the technology just wasn't there.
And I think that's true. We've had discussions with other folks who have worked in the AI space and certainly not until relatively recently has the facility and the efficiency and the power and the usefulness really been there for lots of people working with the air, but it, but it,
But it is now. And when did you come back to it?
And what pieces needed to sort of come into place for this to work for you?
It came back to me probably five years ago.
It's a combination of the technology getting better at At that time, we still weren't into deep learning, but things like support vector machines and random forest classifiers were really starting to get good. result and libraries that made it easily accessible.
So I didn't have to focus as much on coding my own support vector machine or random forest classifier. but it let me focus on the medical imaging aspects.
At a similar time, the National Cancer Institute was starting to get more interested in quantitative imaging and reliably identifying quantitative metrics in images that reflected either cancer or response of cancer to various therapies,
And so they started to fund some laboratories to investigate this more, including my lab.
And that's really what got me back into that field again.
What you're doing today, you call radiogenomics.
And that is certainly a combination of of words, radio I understand, genomics I understand, but what is the combination and how does that describe what you're doing? and how that's different than maybe what radiology has done before and what you're able to do now.
Yeah, so when I trained in radiology, it was very much a pattern recognition approach.
We looked at thousands or millions of examples of cases and my teachers would say, this is what an adenocarcinoma of the lung looks like.
And you just looked at a bunch and you kind of figured out the pattern.
That's kind of what we started with in terms of machine learning.
Radiogenomics really reflects the almost unthinkable thought that in that appearance of the image, we can figure out genomic properties of tumors.
And I think that's still unbelievable to many people in medicine that there is that information present in an image.
But already a number of labs, including mine and other good labs, have shown that there is information in MRI and CT and other modalities that let you figure out the genomic properties of a tumor.
And that's a really important thing. And again, it was almost unimaginable five years ago.
I'm having a hard time wrapping my head around that right now.
I get how radiology In some sense, you're trained as a radiologist, as a physician in pattern recognition and really image recognition.
And that's a very... applicable task for AI image recognition.
Like I know it's a cat. I know this is what a tumor looks like.
I know it's features, uh, And therefore I have a high level of confidence that what I'm seeing as an AI system is a tumor or not, right?
But then is it because... There's so much information that we as humans can't seem that this is now a great opportunity for AI.
And like, how... Could you train a human to see all that genetic information and those nuances that it sounds like you guys are able to look at these days?
The answer is probably, and since I'm a radiologist, I'm allowed to use those sort of indefinite terms, right?
I guess, yeah. But seriously, undoubtedly what's found that reflect the genomic properties in most cases is a textural type of thing and in some cases the textures may be very subtle but if we knew what we were looking for probably a human could be taught to find it.
The other thing is that typically in imaging, particularly some imaging types like MRI, it isn't just one type of image, but Kind of like how a camera acquires red, green and blue images and we integrate those together in one picture.
An MRI acquires many different types of contrast images.
We call them T1, T2. We can get images that reflect the diffusion properties of water. the perfusion of tissues.
We sometimes administer intravenous contrast, and that leaks into tissues, reflecting their profusion as well as their leakiness.
The computer can probably integrate this multidimensional data set more effectively than we can with our eyeballs and brain.
So that is at least theoretically possible for a human to do.
And certainly part of what we do as radiologists is we try to integrate that.
We know some of those rules. But a computer probably can do it better than what a human can do.
And there's probably a lot of information that we don't even know to reconstruct.
One of the really interesting areas is we actually manipulate the signals that come out of the imaging device to produce something that's recognizable as an image, but we'd probably throw away information as part of that process.
And some people are discussing maybe we should feed that raw data into the computer so that there's none of that loss.
Right, so it's sort of machines talking to other machines.
They can understand each other in different ways than maybe we can.
Right. That's fascinating. And clearly we get better quality images that the capabilities of CT, MR, ultrasound imaging devices is clearly better.
But also we just have more data, which is a critical part of machine learning and particularly deep learning.
It really shouldn't be surprising to people in the crowd that medical images actually follow something like Moore's Law that the same doubling of dataset sizes or computing capacity every two years or 18 months happens with medical images.
And that's because medical images are created by computing devices and they're transmitted and stored by computing devices and the vendors make imaging devices that push the limits of computing technology.
So when a computer chip is twice as fast and can transmit the data twice as quickly, They will make images that are twice as large.
And that actually has been shown to be the case that our medical imaging data sets double every 18 to 24 months.
I guess that's one of those things we see it on our smartphones and also our storage on our computers, whether it's solid state or spinning disk.
As much storage, as much speed as we can get, we will always use.
And I suppose it's no different in the imaging realm.
Yep, exactly true. So I have to ask then, I mean, I know genomics, we have genetic data, but what are you able to tell them?
Like, it's fascinating to me that you're saying that you can... kind of see genetic predisposition or genetic traits in an image.
Yeah, so the particular area that I do research on is brain tumors.
And for them, there are a few genomic properties of tumors that are important both for identifying prognosis, but also for selecting the most appropriate type of therapy.
One example of this is 1p19q chromosomal deletion.
We've known for a while that there are some general visual patterns that tend to correlate with that, but with the machine learning methods, we're much more accurate than what humans are able to accomplish.
Another thing that is important in tumors is methylation of MGMT.
It's got a long name and I won't repeat it here, but that's another thing that we can identify with reasonable accuracy from imaging.
Typically, these are done by tests by getting pieces of tissue, but in some cases, we're not able to collect enough tissue that's suitable for testing, and so that the It's not known whether this tumor has MGMT methylation or 1P19Q.
And so that's where imaging can be a useful adjunct to tests of tissue.
So less invasive, all the precision, if not more, and then you can act upon that information in terms of treatment.
What are the outcomes or not necessarily the health outcomes, but how does this change treatment and how you approach In this case, brain tumors or any other kind of case where this can be applied?
There are a few reasons to think that this may be very important for healthcare.
The first is that when you do a genomic test on a tumor, you say, okay, it's got MGMT methylation or 1P19Q chromosomal deletion, for instance.
But clearly, it is not a 100 percent predictor that all the 1p19q patients don't have either 100% response or 0% response.
It's some number in the middle. And so that means that it's much more complex than just one gene.
And of course, that's not a surprise to anybody.
But I think the interesting thing is that imaging probably reflects the more of how that tumor actually is getting by in the host.
In other words, what is that person's normal genome doing to that tumor?
How effectively is it fighting it, trying to kill it? what other genes that maybe we haven't found out about yet that imaging is saying, well, this one is acting more aggressively than what the normal tumor with a certain genotype is, and so it may help us to elucidate subpopulations of patients that we need to look for other genomic markers in.
So I think that imaging by reflecting how that tumor is actually living out or trying to reproduce within the host, probably is a better reflection of biology and may give us a lot of insights into how tumors operate.
Right, and then ultimately that insight doesn't just help patients today, but helps us approach patients tomorrow or even prevention, et cetera.
Yeah, I think that's very true. And you mentioned prevention.
You know, my work today is on tumors, but there's no reason to think that In the same way that we can identify tumor genomics, you know, There are other things we all know about, for instance, the BRCA1 gene that indicates a high probability of breast cancer. it may well be that with imaging we can figure out, well, there is also this other genomic marker or something about their metabolism that moderates the effects of BRCA1 because again, BRCA1 is not a 100% yes, no, you will get cancer, you don't get cancer.
It's a more complex thing and imaging may help to elucidate that more complex reality that we see.
Yeah, the interplay between genes, those that we know that are directly associated with, in this case, cancer mutations and But there's all these ones that we don't know what that interplay means and how it works or doesn't work and why some people show up with cancer and why others don't.
Yep. but I can see how all this information helps us unravel that complex and subtle picture.
If you back up and think about this general radiogenomics approach, how else can it be applied?
I mean, we've talked about cancer if you think about just imaging in general in healthcare and this greater precision and perhaps ability to integrate other images in a more complete picture, And let's say that that data doesn't come just from imaging, but from other things.
It could be wearables, it could be, I don't know, 10 years of history that my doctor has when I go and take my blood.
How does this picture start coming together for you?
And how does AI... apply to that. Yeah, I think that's a great point.
And I personally think that Everybody or nearly everybody nowadays has a cell phone, right?
And most of us talking to the cell phone, I guess people are increasingly texting, but...
Yeah, talking. What do you mean? But, you know, the fact is they collect a lot of information about us.
There's research on the ability of computers to identify stress levels and veracity just based on voice.
Furthermore, if you had that complete record through time, you could probably identify changes in that person.
Maybe it would reflect throat cancer or other problems with the throat.
Maybe it can identify, hey, your stress level is up.
Gee, your stress level is up when you do this.
If it also has a camera, which most do, It can look at pictures and maybe if we start to put infrared detectors into it, your temperature and it can say, gosh, you're not having your normal temperature cycle or it's exaggerated.
And we have now learned that that means you are going to develop influenza virus in the next 24 hours.
So I'm going to send a message to your local pharmacy and they are going to issue an anti-influenza medication.
Yeah, and or even more practically or more sort of prosaic, I'm going to change your meetings that you had scheduled for tomorrow or you're going to not be able to go play in your soccer game or whatever it is.
I love that. I think there's just so much information out there and there's so much interplay with health and healthy living. even just its ability to recognize, you know, what you are eating and how active you are.
I mean, again, that's being collected on everybody.
And I think as we start to associate that with health and healthcare, I think the ability to do more active coaching of people to live healthier lives is very feasible.
You describe something that I think of as this somewhat lower resolution data, right?
All this information that we can probably gather today and we will gather more of 24 seven, it's persistent sensors and persistent gathering of data, but it's lower resolution.
The medical profession has had this approach where I go get my blood tested once a year or four times a day.
There's this very accurate testing. How does that relate to this world of perhaps lower resolution data, but these persistent feeds of data?
You know, a lot of us who went into medical school, you know, spent a lot of time in science and we like the precision associated with that.
And that that probably is reflected in the characteristics that you described.
But I think Increasingly we recognize that having more pervasive information, even if it's of lower resolution, can be extremely valuable.
Again, I think the more that we can at least get alert saying, hey, there's something not quite right here, even if it's not the precise answer, The ability to intervene or at least contemplate earlier intervention is really valuable.
How do you bring the medical community and the world of practitioners along in this?
And what, from a cultural and or I don't know, practical approach.
What are you doing and what has to be done to educate and also to get people comfortable with these kinds of approaches?
So the sad fact is money is probably one of the biggest drivers in healthcare, no matter which political party you're you're from, you know that healthcare and healthcare costs is a big challenge.
And I think that the more you can develop an approach that can improve quality and quality of life and at the same time reduce costs just by leveraging these sorts of technologies You're going to have both sides of the political spectrum on your side.
And physicians also want to have the most cost-efficient care delivered.
For the most part, physicians don't benefit when money is wasted.
So I think that if we can demonstrate cost-effectiveness and improve quality that will be something that everybody is going to buy into.
I think the challenge is that particularly those my age and older have the sense that talking to a doctor and that, you know, pat on the back and the stethoscope on your chest kind of thing is really important for healthcare.
My sense is that the younger generation, in the same way that they don't necessarily talk on the phone but they text, they may be a lot more receptive to... their phone interceding for their healthcare needs, and may not be quite as demanding of having a personal relationship with a physician.
If they can see the big picture and that They're going to get better care by having this sort of technology work on their behalf.
I think that that's going to be something that they will be receptive to.
I think that what you describe is absolutely where we're headed because we've certainly gotten there in many other parts of our lives and there's no reason that that can't happen in healthcare.
At the same time, do these technologies help doctors kind of do what they're best at or focus on the things that really matter in a different way?
So I think you'll probably get different answers to that from different physicians.
Certainly in many cases, the implementation of electronic medical records has been a distraction, but I think to some degree that's because Many electronic medical records have largely been computers trying to recreate paper processes with computer screens. rather than rethinking the whole picture and say, okay, how do we best use technology?
In the optimal healthcare situation, you may still have paper and you still may have people walking around doing the various things that they need to do.
So I think that that receptiveness to the bigger picture of how do we get to the end game of people staying healthy rather than how do we efficiently care for them when sick, From that bigger picture perspective, I think that that's where you can start to change some of the attitudes that exist today.
And certainly the economics, if we move more toward prevention than treatment because something's gone wrong.
Right. Is radiogenomics specifically, is it something that exists only at the Mayo Clinic?
Is it becoming more widespread? And are there any regulatory or... other kind of more bureaucratic hurdles to cross.
Yeah, there are several sites that are investigating and are publishing good work on radiogenomics.
They're kind of the typical leaders in healthcare that you might expect.
There are important regulatory hurdles And that probably is the thing that's slowing things down the most and And I don't say that in a disparaging way.
You know, one has to be careful to make sure that this sort of a technology is rolled out in a safe way.
One wouldn't want an algorithm to say, you've got cancer and it's incurable.
We're not going to do anything with that patient. if the computer happened to be wrong.
So we have to make sure that this is done right.
And that's the role of the FDA and government is to make sure that when People think they have a new medical device that does great things, that they really prove it, and that it's really correct to say that.
So there are regulatory issues. And because the technology is so new, Compared to the old style science of, OK, here is the mechanism, you know, the antibiotic works by disrupting the membrane here and that's how it kills it. especially deep learning, it's much harder to understand how the computer came to its conclusion.
Yes, we have large datasets that say we get it right this many times, but we don't know why.
Understanding why has always been an important part of the approval process.
And I think the FDA has certainly contemplated how they will approve this sort of a device.
But I think that that gives a lot of people pause to approve a device that we don't really understand.
But I'm optimistic. There are good people out at the FDA, and I think they will do the right thing And, you know, hopefully that will be a good partnership between the government, the regulatory agencies, industry and academia to figure out the right trade-off of how much do we need to really understand?
How can we get technology that helps us to extract the knowledge or understanding that's present in those deep learning networks and how do we then translate that efficiently to better patient care.
We started this conversation talking a little bit about how back in the eighties you were working on these kinds of problems and trying to apply machine learning and artificial intelligence. to better analyzing images and imagery.
I can imagine you working on your dissertation back then.
It was a very exciting time. And others, you sort of were banging your head against AI and it wasn't working.
How do you describe this period that you're in now?
Like now that you have the tools and now that you can revisit that period, promise that you were looking at in the 80s.
How do you describe what's going on in healthcare and for you as a person who's been interested in this for a long time?
Well, obviously, it's really exciting. You know, when I think back, you know, this is going to date me, but I was My PhD thesis, I wrote machine language code or assembly language code on this really hot six megahertz 8286 processor.
Yeah. And, you know, when you think about the processing power of that versus what I have on my iPhone versus what we have now in... you know, the GPUs and things like the DGX, it's almost unthinkable.
And I think that's important to also project forward is that we tend to be linear thinkers, right?
We made this advance this year compared to what we were two years ago.
And you project that line forward five years and say that's where we'll be.
But these sorts of technologies are exponentials.
I think especially there's a lot of articles out there about machine learning algorithms, programming machine learning algorithms and I think that's another exponential on top of an exponential.
So to me, it's really exciting and it's one of those things of you better keep your eyes open and your seatbelt tight because there's so much stuff that's gonna happen and figuring out where the best opportunities are to participate and to contribute your piece to advancing the field is critical.
Well, and as somebody on the other side and speaking for us patients, I can't wait.
I mean, the advances, the insight, all of it.
It's happening quickly. And even from my side, I can see that.
And I'm looking forward to the progress and the fast progress.
I am too. I am also a patient. There are certain things that I have to go through as a patient that I don't enjoy, and boy, the faster computers can figure this stuff out and avoid the need for some of those test procedures or treatment procedures, the better off I will be and hopefully all of society will be.
I'm with you there. Brad, just don't pull out your six hertz computer from the old days, the 286 monster there.
Brad Erickson, Professor of Radiology and Associate Chair of Research in Radiology at the Mayo Clinic.
Thank you so much. Thank you.
