Yaron Gurovich’s Quest to Spot Genetic Disorders with Facial Recognition Data
Facial recognition technology — computer systems that can identify individuals from a photo or video — have been a polarizing topic of conversation to date.
While many people are already fairly accustomed to some of its applications, like tagging photos on social media, mass surveillance and other potential uses, much of the discussion about facial recognition is about how its capabilities can give us meaningful insights.
In facial recognition, Yaron Gurovich, Chief Technology Officer at the biotechnology firm FDNA, sees the potential for good. The Israel-based researcher is lead author of a new paper1 about test findings of DeepGestalt, FDNA’s facial analysis framework built to help identify rare genetic disorders.
Gurovich wasn’t always set on marrying machine learning technology with medicine. In fact, he began his career using the technology to help identify people’s activities in videos — say, whether they were walking, or running, or jumping. He then worked at a company developing machine learning tools to detect abnormalities in the production of silicon chips.
“I was fascinated about the problem of identifying patterns and being able to construct useful information,” Gurovich says. “I find it very satisfying to teach a computer to see and to reason.”
Inspired to use his training to solve more meaningful problems, Gurovich joined FDNA in 2012. The company was founded in 2011 to aid in the diagnosis of complex genetic disorders whose symptoms include distinct facial features, and has offices in Boston, MA and Herzliya, Israel. Gurovich has been part of this effort for more than six years.
Genetic diseases such as Noonan syndrome or Cornelia de Lange syndrome have serious effects including developmental delays, heart defects, or intellectual disabilities. These, along with thousands of others, are also marked by different combinations of unique facial characteristics, such as wide-set eyes or a small, upturned nose. However, there are many different types of genetic disorders, and many are extremely rare. They can be difficult and expensive to diagnose, which can result in significant delays to patients getting the care they need. Gurovich understood the potential of using machine learning to recognize these facial patterns and make faster diagnoses.
“The idea of applying computer vision technology in a field where [recognizing] facial appearance is a real clinical practice fascinated me,” Gurovich says.
DeepGestalt is the product of that fascination. Using deep learning algorithms and cloud computing, Gurovich and other researchers trained the facial analysis framework to quantify the similarities between over 17,000 photos of patients, each of which had previously been diagnosed with 1 of about 200 genetic syndromes. Then, they ran experiments to see how well DeepGestalt could match a photo with the particular disorder represented within it. On the final experiment, the facial analysis framework correctly identified the syndrome represented in 502 different patient photos with 91 percent accuracy.
A Growing Cycle
The images used in the study were collected using Face2Gene, a suite of applications FDNA developed that allows clinicians to snap a photo of their patient, which is then uploaded to FDNA’s server and then analyzed by DeepGestalt against a large database of syndromes. Within seconds, the app provides clinicians with a list of genetic disorders that manifest in phenotypes similar to the patient’s facial characteristics, as well as information about those syndromes.
Each time a clinician uses Face2Gene to take a photo or add notes on a patient, the database gets bigger, which helps to improve the technology’s functionality.
“As clinicians use our products more, they make them better and better,” Gurovich explains. “And this cycle is growing.”
Gurovich says that the platform needs so much data because identifying these facial patterns is so complicated. All facial analysis frameworks must learn how to read the subtleties of a face in order to identify different people, but DeepGestalt has the added layer of first needing to distinguish between different people, and then needing to distinguish between different disorders. This becomes even more complex given the individual variables in how those disorders might manifest, especially considering the patients may be different ages, genders, or ethnicities.
These specificities are accommodated through a machine learning technique called “transfer learning,” in which the knowledge a computer gains from one task is then transferred to improve the performance of a second task. In this case, the computer first learns how to analyze faces, and then it learns how to identify the variety of genetic disorders.
Making an Impact
According to the study, DeepGestalt is so effective that it even outperformed clinicians in making accurate diagnoses. Which could raise some concern: Gurovich says some geneticists worry that technology like this will take away from the education of their successors, because they will rely on it too heavily instead of learning to make diagnoses themselves. However, Gurovich says the computer system is not intended to replace humans.
“It is a reference tool, not a diagnostic tool,” Gurovich says. And from his perspective, the fact that technology will bring change is inevitable: “I used to know all my phone numbers by heart, but I don’t do that anymore.”
At this point, the facial recognition technology still has limitations. Because the study only included photos of patients who had been diagnosed with a genetic disorder, it says nothing about whether DeepGestalt can identify patients with genetic disorders from within a pool of patients who don’t all have genetic disorders. In other words, the study doesn’t say that DeepGestalt can determine whether a patient has a genetic disorder — rather, if it is known that a patient has a genetic disorder, DeepGestalt can help determine which disorder he or she has.
Gurovich and the FDNA team plan to continue improving the facial analysis framework. Gurovich says he loves the work; not only does he get to continue training computers to see and reason, but he also says he frequently hears about the impact that FDNA’s technology has had on people’s lives — whether that’s clinicians who credit their ability to diagnose a patient with the Face2Gene application, or parents who were finally able to understand what was going on with their child.
“It’s amazing to think about the impact you can make on the world,” Gurovich says.
This content is produced by WIRED Brand Lab in collaboration with Western Digital Corporation.
- Identifying facial phenotypes of genetic disorders using deep learning
- Genetic disorders in children and young adults: a population study
- 2018 Year in Review