Biomedical Devices Use MathWorks’ MATLAB and Amazon EC2 P3 Instances to Advance Research in Organ Cell Growth

For 36 years, scientists, engineers, and researchers across various fields of study have used software provider MathWorks’ integrated development platform, MATLAB, to quickly, easily, and cost effectively build environments for research. One such person is Dr. Sam Raymond, who is currently a postdoctoral researcher at Stanford University and has been using MATLAB to develop new biomedical devices for organ cell growth. To process the simulation and train the deep machine learning network needed to execute a proof of concept, Raymond, who began his project as a PhD candidate at the Massachusetts Institute of Technology (MIT), turned to the compute power of Amazon Web Services (AWS).

In his bioengineering work, Raymond has used simulators to create a 20 GB dataset and deep learning to understand simulator input and output. He has combined the flexibility of Amazon Elastic Compute Cloud (Amazon EC2) P3 Instances powered by NVIDIA V100 Tensor Core GPUs with MathWorks’ MATLAB’s deep learning capabilities to facilitate this work. In doing so, he has quickly and cost effectively built an environment that scales to accommodate a growing dataset and enables remote work, thus accelerating artificial intelligence (AI) research to develop new biomedical devices for organ cell growth. As Raymond explains, “This method enables biomedical engineers to design better devices more rapidly by leveraging the AI that has been imbued with the results of the simulations. Positioning cells has a number of uses in the field: tissue growth, diagnosis, testing, and filtering. This workflow is intended to offer a new approach for scientists and engineers to explore the combination of simulation data and AI.”

MathWorks is a software company that develops easy-to-use tools for engineers and scientists, helping them turn ideas into designs and move designs into production. MATLAB, MathWorks’ main product and one that is widely used in academic research, gives users a scripting base, desktop development environment, and deep learning capabilities. “We try to make it very accessible and easy for researchers, engineers, and scientists to get to their core work,” says Andy Thé, partner manager for AI and deep learning at MathWorks. “We don’t want people to focus on low-level things that may not add any value to what they need to get done.” MATLAB integrates with roughly 100 different add-on products to enable activities from deep learning to statistics and computer vision. Its abstraction layer makes it easy to set up and use. “MATLAB removes that level of friction for us so that we can just get down to the business of doing research,” says Raymond, who used MATLAB to combine research workflows into a MATLAB program—a project MathWorks took notice of when Raymond presented his work at an MIT conference.

As part of his PhD work, Raymond began working to develop new biomedical devices for growing cells into organ tissue. Executing this delicate process involves positioning cells through ultrasonic sound acoustics. “For example,” says Raymond, “a piece of silicone is etched away to create a space where fluid fills the space and cells are injected. This cavity is then vibrated using an ultrasonic transducer. The sound waves push the cells to specific areas, which enables the careful and deliberate placement of biological materials.”

To form the complex structures required for tissue growth, Raymond and his team needed a way to predict cell arrangement. “While we had a simulator that could give us an answer for any particular input, we wanted to know what input we needed,” says Raymond. “The idea was to use our simulators to generate a huge dataset of possible inputs and outputs, set up a deep learning platform that would understand the connection between the input and the output, and then flip it so that the output that the simulator would have given was actually the input to our deep learning platform. This enabled us to know what we needed to build.” Those inputs reflect coefficient vector geometries, or shapes, and the outputs reflect 2D images in the form of a pressure field. “We ended up with a dataset of about 50,000 different shapes that would have been input to the simulator and then the resultant output pressure fields for each of those tens of thousands of shapes,” says Raymond.

Next, the project required a deep learning platform “to try to understand the connection between the output pressure fields and input geometries,” according to Raymond. “At the end of the day, we wanted to produce a neural network that would—if you gave it a particular shape of a pressure field—spit out the geometry in the Fourier vectors that you would need to then build the device itself.”

Professor John R. Williams, an MIT professor in the Department of Civil and Environmental Engineering and a leading expert in high-performance computing (HPC) simulations and the application of machine learning in the physical sciences, served as Raymond’s PhD advisor. “This work was inspired by patterns in the sand made by ocean waves,” explains Williams. “Until now, it has been impossible to generate complex shapes. But by using AI, Raymond has shown that almost any pattern of particles or human cells in a fluid can be generated using acoustics. Using MATLAB and AWS, he developed his solution in months and not years.”

To fuel this model, Raymond needed agile HPC. At the proof of concept stage, Raymond realized that deploying a local CPU-based solution to run his workflows would take days and bog down the machine. To solve this problem, he would need GPUs. Continuing to use the local CPU solution for postprocessing, Raymond switched to GPU-based Amazon EC2 P3 Instances to handle the project’s HPC needs.

MATLAB’s Cloud Center enables users to integrate their environments with offerings from cloud service providers like AWS. As he built in MATLAB, Raymond turned to Amazon EC2 P3 Instances, which deliver high-performance compute power in the cloud with up to eight NVIDIA V100 Tensor Core GPUs and up to 100 Gbps of networking throughput for machine learning and HPC applications. The GPU-based instances are 100 times faster than the CPUs Raymond previously employed. “Amazon EC2 P3 Instances provided compute that we didn’t have to go out and buy when we made the decision to scale up,” says Raymond. At the proof of concept stage, the dataset was only 50,000 simulation results (approximately 2 GB of data), and he ran it on only one Amazon EC2 P3 Instance. To store the data produced by the simulator and deep learning processes, Raymond used Amazon Simple Storage Service (Amazon S3), an object storage service that offers industry-leading scalability, data availability, security, and performance. “We used Amazon S3 buckets to store, process, read, and load all the data. Then Amazon EC2 P3 Instances processes the data, with MATLAB as the nexus point that provides the programming logic to put on Amazon EC2 instances and to read from and write to the Amazon S3 data,” explains Raymond.

Now the project is stepping up in order of magnitude as the datasets grow and expand from 2D to 3D images, requiring the use of multiple instances. And working in the cloud gives Raymond scalability and flexibility to adjust the environment to suit his ever-changing research needs without requiring him to weigh hefty cost factors at the initial investment. “You have to try one configuration to try another configuration and figure out how your workflow is working,” he says. “If we scaled up the on-premises components that we have, we’d need at least five or more of those at about $20,000 each, which is just not cost effective.” Using cloud computing also gave Raymond the flexibility to access his work anywhere at any time, reducing research friction. “Without sacrificing compute power, I could be anywhere to do this—on an underpowered laptop nearby or halfway around the world,” says Raymond. Not suffering interruptions in his research while visiting his home country of Australia or due to the COVID-19 pandemic ultimately shaved roughly 6 months off the project’s timeline compared to if he had used only local hardware.

In the future, Raymond predicts he will only use on-premises devices for data processing and visualization—but the compute power will come from Amazon EC2 instances: “We’ll be looking at some of the larger instances for our HPC. And these things are obviously going to be far better than anything we could reasonably buy to build up on-premises devices.”

Since publishing his findings, which showed—through both the predicted AI results and the building of the devices in the lab—that the cells were positioned to the locations as desired, Raymond plans to grow the dataset even more and apply new AWS tools to optimize his deep learning capabilities. For example, he notes that Amazon SageMaker, a fully managed service that provides every developer and data scientist with the ability to build, train, and deploy machine learning models quickly, “looks like it has a huge potential for disrupting the way we normally do a lot of our deep learning.”

Raymond’s project is just the tip of the iceberg in how the physical sciences can use the flexible, scalable, and easily used and built environments offered through MathWorks and AWS. “This new combination of deep learning and simulation data is something that we’re starting to see pop up in pretty much every field of science and engineering,” says Raymond. MathWorks has already seen AI grow in popularity among customers, according to Thé: “It’s really kind of exploding. Even some of our customers have said that AI is going to be the undercurrent across all of their big initiatives. AI and deep learning have much better results than traditional methods.”

Using the deep learning capabilities of MATLAB and the power of AWS, Raymond brought a complex idea, born of his own PhD work, to fruition. By enabling the development of new biomedical devices for growing cells into organ tissue, he has built an environment that can adapt and scale with his project as it continues to advance.