Detection of Rare Genetic Diseases using Facial 2D Images with Transfer Learning

Around 6-8% of people are affected by rare genetic disorders in the world. Due to unavailability of proper identification of gene variants causing these rare disorders, the genetic tests for many of them do not exist with an exception to some common conditions like Down’s Syndrome. A limited number of individuals have the proper training and ability to recognize these disorders and have to depend on the pronounced facial features which occur in about 30-40% of these diseases. Fine-tuning of the pre-trained model VGGFace(with ResNet50 network architecture) on our dataset comprising of 12 classes of rare genetic syndromes leads to a top accuracy of 97.66% and an F1 score of 0.86. On using a smaller dataset with 8 classes of syndromes same as Clinical Face Phenotype Space, the state-of-the-art, our model produces a top accuracy of 98.1% and an F1 score of 0.92. This could be used to screen for rare genetic diseases in newborns.