September 29, 2017
At this point, most of us understand every industry will be transformed by artificial intelligence, but few sectors can expect more benefit from the AI Spring than healthcare and medicine. Smart researchers are compiling massive databases to help AIs diagnose rare diseases. Consumer products like FitBits allow regular people to track their activity, exercise regimen, and sleep patterns like never before. There are even strides being made in mental health, where “psychologists” like Ellie can actually get patients to open up in ways that can be hard when we’re talking to a person we’re afraid might judge us.
But perhaps the most promising sector of medicine for artificial intelligence is imaging. Medical images are, in some ways, ideal training data for computer vision projects. Not only are they fairly uniform–as in, one CAT scan shares a lot of attributes with another–but there are tons of available medical images to use a training and tuning data. Just as an example, thousands of radiology X-rays could teach a deep learning model to effectively diagnose everything from broken bones to arthritis.
However, not every branch of medicine enjoys the preponderance of data that radiology does. In fact, some fields suffer from a lack of actionable data. One such area? Cellular imaging. And David Van Valen, the first of our AI for Everyone Challenge winners, is hoping to change just that.
Van Valen’s project is titled DeepCell. It aims to marry deep learning techniques with live-cell cellular imaging to create datasets and algorithms that can spur innovations in disease research.
So what makes his work a bit different? Van Valen is interested expressly in mapping the behavior of individual cells. He’s applied this approach earlier in his research and saw great success but, like so many researchers (and data scientists, for that matter), he discovered that was holding him back wasn’t the models he was using but rather the data he had access to. Put simply, there just isn’t enough of it.
Like many research heads, David could get access a small army of grad students for image labeling. But that wouldn’t get him to the amount he needs to build a scalable live-cell model. That’s why he submitted to our AI for Everyone Challenge (and won!). He has the requisite expertise to build a model for live-cell research, just not the preponderance of well-labeled data. In a year’s time, we hope he will.
DeepCell will be running jobs on CrowdFlower till late 2018, with our contributors and platform supplying pixel labeled to microscopy imaging. The goal is fuel experiments in “live-cell imaging to probe the behavior of signaling networks, as well as use single-cell imaging as a readout for genomic scale screens.”
We look forward to keeping you up to date on DeepCell’s progress and sharing the breakthroughs, papers, and research that arise from the project.