Deep Neural Network Diagnosis of Autism Spectrum Disorder Through Visual Image Eye Movements
Student
Connor Theisen
Advisor
Catherine Qi Zhao
Abstract
Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder and is defined by repetitive and characteristic patterns of behavior and difficulties with social communication and interaction. Examining ASD traits is time-consuming along with advanced diagnostic instruments. A deep learning based ASD classification model would help reduce time spent on ASD diagnosis activity and recognize ASD earlier to get necessary medication or therapy. Previous studies approach quantitative and objective ASD classification based on eye tracking and deep neural networks. Based on where subjects look in an image, they use a convolutional neural network (CNN) to extract image features from the subjects’ fixations and use a recurrent neural network (RNN) to integrate features across multiple fixations for classification. In this work, we aim to address two main challenges with regards to the previous ASD classification models by 1) enhancing the accuracy, and 2) aggregating image-wise results for subject-wise predictions. Specifically, we propose a new CNN backbone for better feature extraction of eye fixations, and a weighting-voting strategy for improved subject-wise predictions. The new CNN backbone addresses the issues around overfitting from small data and increasing the confidence in the classifier’s predictions by removing redundant feature maps, reducing the number of parameters, and layer’s having more access to preceding feature maps. The weighting-voting strategy gives proper credit to certain images that a subject has more distinguishing activity than in other images. Our experiments demonstrated a higher accuracy compared to the previous methods. Within that enhanced accuracy, there was reduced overfitting in our training and testing, and there were higher confidences in our classifier’s predictions. The conversions of image-wise to subject-wise predictions also saw enhanced accuracy prediction values due to our weighted-voted strategy.
Video
Deep Neural Network Diagnosis of Autism Spectrum Disorder Through Visual Image Eye Movements