|dc.description.abstract||Autism Spectrum Disorder (ASD) is a neurological disorder. It refers to a wide range of behavioral and social abnormality and causes problems with social skills, repetitive behaviors, speech, and nonverbal communication. Even though there is no exact cure to ASD, an early diagnosis can help the patient take precautionary steps. Diagnosis of ASD has been of great interest recently, as researchers are yet to find a specific biomarker to detect the disease successfully. For the diagnosis of ASD, subjects need to go through behavioral observation and interview, which are not accurate sometimes. Also, there is a lack of dissimilarity between neuroimages of ASD subjects and healthy control (HC) subjects which make the use of neuroimages difficult for the diagnosis. So, machine learning-based approaches to diagnose ASD are becoming popular day by day. In the machine learning-based approach, features are extracted either from the functional MRI images or the structural MRI images to build the models.
In this study at first, I created brain networks from the resting-state functional MRI (rs-fMRI) images, by using the 264 regions based parcellation scheme. These 264 regions capture the functional activity of the brain more accurately compared to regions defined in other parcellation schemes. Next, I extracted spectrum as a raw feature and combined it with other network based topological centralities: assortativity, clustering coefficient, the average degree. By applying a feature selection algorithm on the extracted features, I reduced the dimension of the features to cope up with overfitting. Then I used the selected features in support vector machine (SVM), K-nearest neighbor (KNN), linear discriminant analysis (LDA), and logistic regression (LR) for the diagnosis of ASD. Using the proposed method on Autism Brain Imaging Data Exchange (ABIDE) I achieved the classification accuracy of 78.4% for LDA, 77.0% for LR, 73.5% for SVM, and 73.8% for KNN.
Next, I built a deep neural network model for the classification and feature selection using the autoencoder. In this approach, I used the previously defined features to build the DNN classifier. The DNN classifier is pre-trained using the autoencoder. Due to the pre-training, there has been a significant increase in the performance of the DNN classifier. I also proposed an autoencoder based feature selector. The latent space representation of the autoencoder is used to create a discriminate and compressed representation of the features. To make a more discriminate representation, the autoencoder is pre-trained with the DNN classifier. The classification accuracy of the DNN classifier and the autoencoder based feature selector is 79.2% and 74.6%, respectively.
Finally, I studied the structural MRI images and proposed a convolutional autoencoder (CAE) based classification model. The T1-weighted MRI images without any pre-processing are used in this study. As the effect of age is very important when studying the structural images for the diagnosis of ASD, I used the ABIDE 1 dataset, which covers subjects with a wide range of ages. Using the proposed CAE based diagnosis method, I achieved a classification accuracy of 96.6%, which is better than any other study for the diagnosis of ASD using the ABIDE 1 dataset.
The results of this thesis demonstrate that the spectrum of the brain networks is an essential feature for the diagnosis of ASD and rather than extracting features from the structural MRI image a more efficient way is to use the images directly into deep learning models. The proposed studies in this thesis can help to build an early diagnosis model for ASD.||