After an image is preprocessed and cropped into regions, the artificial intelligence system, DeepGestalt, is programmed to associate facial features with genetic disorders.
Rare genetic disorders can take years or even decades to properly diagnose. This prevents timely access to appropriate treatment and can have severe, irreversible, and life-threatening consequences. Because many genetic disorders are associated with unique facial features, facial analysis technologies have been developed to help with diagnosis. Until recently, however, this technology was limited to the identification of only a few diseases. In a new study, a team of scientists demonstrate the power of a facial image analysis framework to quickly and affordably recognize the features of multiple disorders.
Yaron Gurovich, Chief Technology Officer at the artificial intelligence-based company FDNA, Inc. in Boston, MA, and his team created an artificial intelligence system called DeepGestalt to power their diagnostic aid application, Face2Gene. The app allows healthcare workers to take photographs of patients (which are de-identified before they are uploaded for analysis) and label them with their diagnoses, thereby creating a large, crowdsourced dataset linking faces and genetic disorders. This enabled researchers to train DeepGestalt on over 17,000 images to recognize facial features associated with the genetic markers of 216 different conditions.
Gurovich and his collaborators ran four experiments to demonstrate the ability of DeepGestalt to associate facial features with genetic disorders. In two experiments, the program’s performance surpassed human experts in diagnosing specific syndromes (Cornelia de Lange syndrome and Angelman syndrome) by roughly 20 percent. In the third experiment, DeepGestalt correctly differentiated between five genetic variations of Noonan syndrome—a condition associated with wide set eyes and droopy lids—64 percent of the time, far better than random chance (20 percent). Finally, when the program was tested on 502 new images comprised of patients with ninety-two different syndromes, it put the correct diagnosis in a top ten list of possibilities in over 90 percent of the cases.
Because these experiments assumed that patients in the test images have a genetic syndrome, the authors point out that the results cannot be applied to test sets that include images of people without genetic disorders. They also suggest that the use of the app should be monitored for those outside of the health care field who might use the data to discriminate against patients—for example, to deny insurance claims or employment. “Photos of our faces are everywhere,” says Gurovich, “Facebook, Google, and others—and no one thinks about regulating access to these photos in any way.” (Nature Medicine)