A Convolutional Neural Network for gaze preference detection: A potential tool for diagnostics of autism spectrum disorder in children
Early diagnosis of autism spectrum disorder (ASD) is known to improve the quality of life of affected individuals. However, diagnosis is often delayed even in wealthier countries including the US, largely due to the fact that gold standard diagnostic tools such as the Autism Diagnostic Observation Schedule (ADOS) and the Autism Diagnostic Interview-Revised (ADI-R) are time consuming and require expertise to administer. This trend is even more pronounced lower resources settings due to a lack of trained experts. As a result, alternative, less technical methods that leverage the unique ways in which children with ASD react to visual stimulation in a controlled environment have been developed to help facilitate early diagnosis. Previous studies have shown that, when exposed to a video that presents both social and abstract scenes side by side, a child with ASD will focus their attention towards the abstract images on the screen to a greater extent than a child without ASD. Such differential responses make it possible to implement an algorithm for the rapid diagnosis of ASD based on eye tracking against different visual stimuli. Here we propose a convolutional neural network (CNN) algorithm for gaze prediction using images extracted from a one-minute stimulus video. Our model achieved a high accuracy rate and robustness for prediction of gaze direction with independent persons and employing a different camera than the one used during testing. In addition to this, the proposed algorithm achieves a fast response time, providing a near real-time evaluation of ASD. Thereby, by applying the proposed method, we could significantly reduce the diagnosis time and facilitate the diagnosis of ASD in low resource regions.