DeepGestalt facial recognition will change the way we diagnose disease

Eight percent of the population is affected by genetic disorders that can be detected through distinct facial characteristics. However, because these disorders are so rare, many physicians have trouble recognizing them. This creates significant stress for the patients and families who are seeking help in understanding what may be wrong.

DeepGestalt: Facial Recognition Meets Genetics

Recently, Nature Medicine published the results of a study in which a facial-image-analysis framework, known as DeepGestalt, was able to outperform physicians at diagnosing genetic conditions using only images of faces. The framework was created by FDNA, a genomics/AI company that creates applications which derive genomic insights from physiological data. Founded just eight years ago, the company is currently used by 70% of the world’s geneticists.

DeepGestalt works by segmenting facial images into multiple regions and assessing each unique region individually. The application compares facial features within each region to its bank of genetic disorder knowledge and then aggregates results to make a final judgment. As with other AI technology, DeepGestalt’s only major shortcoming is its inability to explain how it arrives at its conclusions.

Advertisement

Testing and Training DeepGestalt

Training DeepGestalt required a data set that was once the second largest in the world behind Facebook’s private data set. The framework was trained using 500,000 facial images belonging to 10,000 individuals. Developers relied on the application’s deep learning capability which uses multiple computational layers to process large volumes of complex data.

Initially, DeepGestalt’s developers tested the application with two different genetic disorders, Cornelia de Lange syndrome and Angelman Syndrome. In both cases, DeepGestalt was able to identify the correct disorder from facial images greater than 90% of the time, around 20 percentage points higher than expert physicians.

The FDNA team then tested the framework using images of individuals with Noonan syndrome. They were interested in seeing if DeepGestalt could distinguish different genotypes from a small group of people who all had the same genetic disorder. Although less successful (64% accurate), the framework performed over 40% higher than projected. DeepGestalt’s final test was a diagnostic evaluation in which the framework was able to identify 216 different genetic disorders with 90% accuracy.

AI Enabling Personalized Medicine

DeepGestalt is one of many modern innovations that is enhancing personalized medicine. AI applications built with facial and image recognition capabilities are continuing to evolve in the healthcare space and augment physician skill sets in tremendously valuable ways. Because of companies like FDNA, we are likely to see many more AI-enabled technologies that will lead to better patient outcomes and experiences.

Facebook Comments