David Van Valen Uses Deep Learning to Analyze Cells

January 30, 2017

The deep learning methods that social media platforms, such as Facebook, enlist to identify people and objects in photographs could someday be used to analyze cells, bacteria and viruses in real-time in the laboratory.

In a paper recently published by the journal PLOS Computational Biology, Hertz Fellow David Van Valen and his fellow Stanford University researchers illustrated that algorithms like those used to recognize human forms and faces can be used to solve a major problem with live-cell imaging known as image segmentation.

“A lot of what people want to look at these days are individual cells because they vary so much from cell to cell,” Van Valen explained. “[With cell imaging], you take these cells, look at them under a dish and take a bunch of pictures to make a movie. At the end of the day, there’s this problem with image segmentation that you have to solve -- in each picture you collect, you need to identify which part of that picture corresponds to which individual cell. It turns out it’s actually a really hard problem.”

While a number of tools and computer vision techniques have been combined by researchers to overcome the problem, a singular method that could be shared among different labs didn’t exist, Van Valen said. Manually curating images by hand is a long, laborious process. After mulling it over, Van Valen stumbled on an answer while on Facebook, noticing the photos he’d uploaded were being automatically annotated.

“I was browsing through some pictures of a Brazilian jiu-jitsu class, and realized that not only had Facebook drawn squares around each person’s face, but they kind of had a sense of who each person was,” Van Valen said. “So I thought, if they could do this, then why is it so hard to identify cells in our microscope finishes?”

So Van Valen went to work, researching the latest computer vision and deep learning methods, and discovering he could apply one of the breakthrough applications in deep learning – image classification. Instead of manually programming data features, Van Valen found he could start with a set of labeled images and let the algorithms do the rest.

“The algorithm will take in a picture and label what it is,” Van Valen said. “Mathematically the problem that we need to solve with image segmentation, if you think about it in the right way, is almost exactly like an image classification problem. Once you make that connection, then it’s really straightforward to use what people have done in a different context.”

Along with his team, Van Valen – who credits advisor Markus Covert and professor Euan Ashley for their instrumental work on the project – implemented the algorithms with images from a number of cell types, including bacteria and mammalian cell lines. In almost all of the cases, the approach worked as well or better than any previously published methods.

While the results are promising, Van Valen thinks the algorithms can be improved. With more development, Van Valen said the deep learning approach could help usher in a new era in high-throughput single-cell biology.

“The bottleneck now isn’t so much our ability to collect data, it’s our ability to process and analyze,” Van Valen said. “By removing this barrier of image segmentation, we’ll be able to get a lot more data and higher-quality data for a number of different biological systems and a number of different biomedical contexts.”

Van Valen in the lab
David Van Valen at his bench, preparing bacterial samples for imaging.

At 31, Van Valen is a true Renaissance man, holding an MD from UCLA, degrees in math and physics from MIT, and a PhD in applied physics from Caltech. His unique education path took him from being home schooled with his brother Joseph (who graduated from Stanford) to high school at age 10. By 14, he was already a freshman at MIT, and he completed his dual degrees when most students were just entering college.

“I was a very motivated student, and as long as I got to learn something interesting, I didn’t feel or notice the differences,” Van Valen said. “I definitely benefitted from having a number of professors who genuinely wanted to see me succeed. The environment enabled me to get by doing that without paying too much of a price.”

For his PhD, Van Valen took on his favorite project: a study of how bacterial viruses obtain their DNA. After analyzing movies of viruses infecting single cells, the findings suggested that instead of DNA pushed in from outside, it’s pulled from inside the cell, a discovery contrary to the general consensus knowledge of how host-pathogens behave.

Currently a postdoctoral fellow at Stanford, Van Valen wants to continue his career in academia and expand his study of host-virus interactions. He's working on single-cell studies of signaling and transcription, particularly looking at how the innate immune system responds to particular bacteria. The hope is, Van Valen said, that by understanding information transmission he can reach a better understanding of disease.

“My philosophy is that I want to have my cake and eat it, too,” Van Valen said. “I’m working on problems that have related clinical questions, and in the future I will steer research further in that direction… If you explore these kind of questions in the right system and right context, not only will you get scientific insight, but you will get insight into a medically-relevant problem.”

Using deep learning, scientists can now identify the location and identity of single cells in microscope images.

A phase microscope image of two different cell types - MCF10A and NIH-3T3. Both cell types express a different fluorescent protein for identification.
A deep learning based algorithm predicts the location and identity of every cell in the image.
The algorithm has a 95% success rate in determining cellular identity.