The Early Detection of Autism Within Children Through Facial Recognition; A Deep Transfer Learning Approach

Abstract

Over the past years, autism rates have increased alarmingly, with 1 in 59 children, aged between 1 to 6 years, being affected globally. While treatment is available, if detected at a later stage or not detected at all, children must face lifelong consequences and even a reduced life expectancy. Therefore, an early diagnosis has the potential to enhance the children’s probability of having near-to-normal development. However, current methods of diagnosis are not accessible to everyone due to the high costs involved in clinical assessments and the time taken to reach a conclusive diagnosis thus leading majority of children being under-diagnosed. Deep learning has transformed multiple sectors thanks to its "high perform a nee" feature as opposed to traditional machine learning models and could have been long used for the early detection of autism as an attempt to reduce the affliction rates. Although autistic children have unique facial features which could be exploited using Deep Learning, not much effort has been put in that area. As such, this work takes on a Deep Transfer Learning approach for the detection of autism within children based on facial images by applying CNN-based models of ResNet50, VGG-16 and MobileNet with the latter being the most performant. After tuning, an overall accuracy of 89.5% and AUC of 0.97 were reached. Furthermore, on an endnote, the practical & ethical implications are looked at while also proposing that, as this work shows promising results, future works could look at a more real-time approach for the same.