Face Time: How AI Can Diagnose Rare Genetic Diseases FasterMarch 14, 2017
From tagging people on social media to identifying travelers at ports of entry, facial recognition technology has become commonplace. Now, a young company is looking to save lives with it.
More than 30 million Americans — 80 percent of them children — suffer from one of 7,000 rare genetic diseases. Many of these people bear abnormal growth patterns of the face or skull specific to their disease.
When it comes to identifying such diseases, however, the diagnosis process is surprisingly archaic. Doctors rely on a variety of antiquated approaches, from manually measuring the distances between facial features to drawing on decades of experience to detect patterns. On average, getting an accurate diagnosis takes seven years, after visiting multiple doctors.
A Huge Unmet Need
A few years ago, the situation caught the attention of a team of entrepreneurs who had just sold their facial recognition company, Face.com, to Facebook. They decided to apply the underlying machine learning technology they’d developed to help geneticists overcome this diagnostic challenge.
So they founded FDNA, a Boston-based startup devoted to applying facial recognition, deep learning and artificial intelligence technologies to advance the diagnostics and therapeutics for rare diseases.
“We understood immediately that this practice could benefit from advanced technology,” said Dekel Gelbman, CEO of FDNA.
It’s not just caregivers that could put FDNA’s technology to use. Pharmaceutical companies are eager to capitalize on a market for orphan drugs (which are used to treat rare diseases) that McKinsey & Co. estimates will be worth $176 billion by 2020, representing nearly one-fifth of the global prescription drug market.
Putting a Face on Genetic Syndromes
The team went to work on building a database of images and a network to interpret them. They spent three years forming collaborations with geneticists and clinics around the world to start crowdsourcing images and data. Simultaneously they began building and training their system.
By 2014, FDNA had introduced a product, Face2Gene, that could help identify about 50 known genetic syndromes. The company’s network was trained on NVIDIA GPUs, with cloud-based GPUs running on an Amazon instance used when extra horsepower was needed to accelerate performance.
“The key to support this system is to be able to build algorithms that can train fast, and to train in a way that is feasible,” said Yaron Gurovich, vice president of research and development at FDNA. “Without GPUs, that’s not possible.”
On to Deep Learning
In 2015, FDNA decided to up the ante and adopt deep learning methods, building a new architecture and expanding what had been a reference tool into a full suite of apps.
Using NVIDIA’s CUDA parallel processing platform and cuDNN library, FDNA built a powerful network that supports apps for clinical evaluation and clinical consulting forums, as well as a medical library and an API. Together, these allow labs to use anonymous patient facial characteristics and phenotypes to increase the estimated 25 percent likelihood of finding diagnoses to 40 percent, Gelbman said.
In the process, it also reduced what had become a week-long training process to just a few hours.
The new suite of Face2Gene apps, introduced last October at the annual meeting of the American Society of Human Genetics in Vancouver, B.C., now boasts more than 2,000 documented syndromes.
Gelbman said 70 percent of geneticists worldwide are now using Face2Gene, constantly expanding FDNA’s image database as they upload patient photos to more quickly generate diagnoses.
With Face2Gene established as a standard for clinical evaluations, Gelbman expects it to soon become standard in lab analysis. By 2018, the company hopes to be helping pharmaceutical companies accelerate their patient and drug discovery.
“What we theoretically can do could absolutely save the lives of millions of people and increase and improve the quality of life for these patients and their families,” said Gelbman.
To learn more about how AI computing is changing healthcare and other industries, join us for the GPU Technology Conference, May 8-11 in Silicon Valley. Register early and save.