Thesis (PhD) - Cyprus International University. Institute of Graduate Studies and Research Management Information Systems Department

Includes References (sheets 96-110)

ABSTRACT
Autism spectrum disorder has been a condition which affects people for a very long period of time. One key point, which still poses a serious challenge is the early diagnosis of the disorder. Early diagnosis of the disease is crucial since the early intervention significantly improves the outcome. The diagnosis of autism spectrum disorder poses a challenge, since it appears in wide variety of ways in subjects. In the literature it is shown that the children with ASD show worse quality of forming letters. Thus, in this thesis, machine learning based, automated ASD diagnosis method has been developed using handwriting as input. In this approach, several different tasks are given to pupils such as drawing rectangles and these drawings are fed into CNN to diagnose ASD. Best to our knowledge, there is no dataset available for this task. Thus, first, a dataset has been formed, where some drawing tasks are given to pupils with ASD and without ASD. Since it is difficult to collect data from pupils with ASD, transfer learning has been employed to increase accuracy. Also,
for each task a different network trained and classification results for the same pupil are combined by taking median of estimated classes (majority voting) by the networks. This has been done since in practice a pupil with ASD might not be willing to finish the task and thus, the input size may vary. Consequently, a dataset
with 104 pupils (split as 80% for training and 20% for testing) is formed and it is shown that the proposed approach can correctly classify ASD with an accuracy of 90.48%, where sensitivity, and specificity are calculated as 80%, and 100% respectively.
Keywords: Autism Detection, Autism Spectrum Disorder, Computer Vision, Convolutional Neural Network, Googlenet, Transfer Learning.