You are here
March 30, 2023
Today, researchers use a variety of imaging techniques to visualize and analyze biological systems, but there are limits to how much—and how well—these tools can see. But Duke BME’s Roarke Horstmeyer and his team are creating new microscopes and imaging algorithms to capture biomedical images at never-before-seen scales.
Michaela Kane 00:20
This is rate of change a podcast from Duke dedicated to the ingenious ways that engineers are solving some of society's toughest problems. I'm Michaela Martinez. In the world of Biomedical Imaging, most people are concerned with what an image tells us is a bone broken our tumor cells shrinking our neurons firing. Today, researchers use a variety of imaging techniques to visualize and analyze biological systems. But there are limits to how much and how well these tools can see. Many optical tools can only capture lower resolution images with limited fields of view and slow video frame rates, all of which impact image clarity, but records Meyer and his lab have made it their mission to address these shortcomings.
Roarke Horstmeyer 01:01
So my name is Roarke Horstmeyer. I'm an assistant professor in the biomedical engineering department here. And I make optical devices and algorithms, mostly for biomedical applications.
Michaela Kane 01:15
Despite his current role in Dukes biomedical engineering department, Horace Meyers first experience with optical tools wasn't actually in a biomedical lab.
Roarke Horstmeyer 01:23
Yeah, so I actually was an undergraduate at Duke. And I studied physics. And from probably like that, when I was 18 years old onwards, I always worked on projects involving optics. The first project I worked on was taking data from the Hubble telescope and processing it very simply, I was learning programming at the time, but to help an astronomer at NASA through an internship. And so I learned about the telescope and how it worked. And I got interested in that. So when I was here at Duke, I worked with some neutrino data, which is not really optics, but it was detecting light, and then joined for about a year, David Brady's lab, here, he was obviously making all sorts of different cameras and optical devices. And so through that experience, I learned a ton. And then when I graduated, I went on to continue to work with cameras and algorithms for processing images. And that then led into microscopes. So I really came to biomedical engineering, kind of late, I'd say I, and I came from the perspective of someone making optical instruments and then learned a tiny bit, and still not that much about biology. It's a huge, crazy, new world for me, and I'm slowly learning different pieces of it as I can,
Michaela Kane 02:56
Even though he had a late entry into the Biological and Biomedical world. Horstmeyer has already found success by collaborating with researchers who can use his tools to better study biomedical phenomena. One of his most popular tools is the multi camera over a microscope, which packs several cameras together into a grid.
Roarke Horstmeyer 03:14
So with a regular microscope, you zoom in to see something at high resolution, but then you can't see that much area, you're limited to a small little, you know, something the size of a dime, maybe. And so if something's moving around, it can't move around very much, or at all, you would constrain it to be able to study it. And that is problematic if you want to study its natural behavior, or social interactions or kind of behavioral aspects of, it's more, it's more morphology and how that connects to its behavior. So if you can see a much bigger area, then you can watch them move around. And so with a microscope, you zoom out, but then you can't see it at such fine detail, you lose the resolution. And so by having many many microscopes and an array, you can simultaneously see over this big area. And at high resolution
Roarke Horstmeyer 04:12
Horstmeyer first began toying around with the idea for this microscope in graduate school, when a friend who worked at a smartphone company would give him rolls of the sensors that were used to create smartphone cameras.
Roarke Horstmeyer 04:22
And they would give me suddenly sensors, which are really amazing cameras. And they almost were like, just giving them to me as if they didn't cost anything. You think, Oh, it's this amazing piece of technology, which it is. But because they make so many of them, right, billions of them a year. The cost is like a fraction of $1 for each sensor. And so the way that they deliver these sensors to me, was in a roll. So each kind of sensor was like it was as if it was a roll of film, but each physical sensor was just in the little patch, and I'd get like several 100 and a roll and Here's just here's, here's just three. And so I thought that I was just amazed. I was like, these are like, commodity, you know, they're it's like, so cheap for them to make, but they're so powerful. And so a friend of mine, Mark, and I, at the time, we're thinking of ways to try to improve how microscopes work, and this array idea came up. And it was just so clear to me that if you wanted to put 100 of these microscopes together, you should use smartphone camera, right? They're cheap, and they're just they're really high performing. And so all of our engineering kind of efforts were built around using smartphone cameras and our microscopes. And so yeah, what we do is we get somewhere between 50 and 100 of those sensors. We pack them all together, almost touching, and then we put lenses on top. And then we have the biggest hardest part is having electronics behind it to be able to read in 50 or 100 videos all at once. And we along with a friend of mine before coming to Duke I we made a prototype of that, I guess it's been about five years now five years ago. And it's just slowly kind of progressed and gotten more. more nuanced in its applications and also better performing. We've, I think, got a pretty good design working. And we're using that to mostly image small model organisms as they move around.
Michaela Kane 06:42
One of the organisms Horstmeyer's microscope helps neuroscientists study at Duke is zebrafish. In traditional studies, researchers would need to anesthetize the young fish to keep them still while measuring their development. But there were concerns that this would harm the fish's growth and behavior, which would then skew the experiments results. In two papers that appear in nature photonics and Optica, Horstmeyer and his collaborators show that the multi camera at a microscope, or the MCAM, enables researchers to measure and visualize the behavior of the fish and other organisms non-invasively. These results were informed by early experiments that were completed in close collaboration with Dr. Eva Naumann's lab at Duke, which led to a recent publication in eLife
By stitching together 54 lenses, the team was able to collect 3d images, large scale video and take photos at a near cellular level.
And videos posted to their website, users can explore images collected with the MCM. And one example which shows a mosaic of butterfly wings, it's possible to zoom in enough to see the individual colored scales on each wing. The technology has enabled Horstmeyer to build numerous collaborations both at Duke and at universities across the US. It has also led to the launch of Ramona Optics.
Roarke Horstmeyer 07:47
Yeah. And then we have a startup here in Durham. That's, that's developing the technology for not just labs at Duke, but other labs that are the universities. And so yeah, the microscopes definitely getting out there. It's probably at 10 different universities right now, primarily universities. And it's really starting at other research labs, because its capabilities are so new, that it's it's, you know, facilitating new discoveries. And so, we, yeah, are excited to try to make sure that the microscope can lead to these new findings. I think that's what motivates me, yeah, to push that technology for.
Roarke Horstmeyer 08:35
But if you're getting a ton of detail over a large space, not to mention collecting data from 54 synced sources, your file size is going to be big, really, really big. And when you're working with an image or video file jam packed with detail, it can be easy to miss small near cellular changes. How big is that video file?
Roarke Horstmeyer 08:53
Huge. Yeah. So currently, each frame that we create is almost gigabyte. So we call it a gigapixel. Okay. So your smartphone is, let's say 10 megapixels, or five megapixels? I don't know. It depends on the phone you buy. Some of them go bigger. And they say, Oh, you can get like a 30 megapixel phone. We're taking those sensors and putting 50 of them together. And so that quickly gets you up to 1000 megabytes, or megapixels or gigapixel. Okay, and then we're doing video. Yeah. And so if we're taking a video right now, let's say it's around 10 frames per second. You can see how much data can quickly accrue. When you just press play and wait for a minute, you fill up a terabyte hard drive. So yeah, that's fine, right? You can do that. And then say here, here's all your data to whoever is doing the experiment. The harder part is, what do you do with all that data? Yeah, you have to process it and extract useful information. And so that's why we're also really excited about developing algorithms to help sift through and make use of all of that image data. So the second project we work on is algorithms to process all these images. So the amount of data we're capturing is really large. If we're looking at a big area at high resolution, it's almost too large for, you know, someone just to sit there and look at and scroll through and try to make sense of easily in a efficient way. So we have to develop tools to kind of find organisms as they're moving around, track them, and try to understand what they're doing automatically. And that kind of algorithmic development not only applies to looking at things moving around these little organisms, but also looking at things that aren't moving, which is like specimen material from patients at the Duke clinic. So we have a number of collaborations there, looking at thin tissue sections, blood smear, cytology, smears, and trying to help the hematologists and cytologists and pathologists more rapidly and efficiently and effectively make clinical decisions from from lots and lots of image data that they have to sift through as well. So the algorithms that we're developing, are using machine learning to infer diagnostic or prognostic information that you would otherwise not expect, or perhaps otherwise not be able to directly achieve just from looking at the image data.
Michaela Kane 11:35
An example of this work involved Horstmeyer and his team imaging blood smears from patients during the COVID 19 pandemic
Roarke Horstmeyer 11:42
What we did is we took lots of images of different white blood cells from each patient, pass them into a machine learning network and had the network predict if the patient had COVID-19 or not, that's a start the simplest thing to start. And we did that for hundreds of patients. And then we also had a cohort that was sick, but not with COVID-19. Okay, and challenged the algorithm with that data, just from looking at the images of the blood cells, it was able to get pretty high accuracy, like over 90% accuracy. And so that was just interesting to see, and curious to us. But what we're really after is, you know, you can go ahead and just do that and say here's, here's, here's this result. My lab is really interested in why these machine learning algorithms are doing what they're doing. And so we took a few steps to try to figure that out. And one of them was kind of interesting, we first categorized all the white blood cells into their type of cell, and then looked at which cell type was most important, okay, for the diagnosis, and that link to some really recent findings that were saying these particular cell types were impacted by COVID-19. And not these other ones. Okay, so our algorithm was learning science that other people were doing, you know, right, concurrently at the same time, okay. And then we additionally showed a lot of the images that the algorithm of specific cells, the cells that were really important for its diagnosis to pathologists, so we kind of closed the loop from, you know, having the machine try to make a diagnosis to then trying to teach or show a clinician what was important. And yeah, some of the feedback we got was really interesting. And so what we're hoping is like, kind of like creating this loop between algorithms trying to make decisions and then having clinicians or others look at what the algorithm is doing. There's kind of this virtuous cycle of learning improvement, making the algorithms better, but then also teaching the clinicians something new that they may not have otherwise seen or suspected, whether it's
Michaela Kane 13:57
developing new algorithms to examine and identify cells, or creating new microscopes to view never before seen details and zebrafish development. Horstmeyer is excited about how his tools can be of use to all researchers.
Roarke Horstmeyer 14:09
Yeah, that's what's great about and I mean, what I love about making microscopes or just Yeah, imaging devices is that everyone needs images, or everyone wants to take pictures, either super high resolution or really far away of the stars or whatever. And so you get to learn about all sorts of applications. If you know you, you can help make better images for what people are trying to capture.
Michaela Kane 14:36
Thanks for tuning in to this episode of Rate of Change. Remember to follow us on social media for updates and be sure to subscribe to get alerts about future episodes. Thanks for listening.