A novel dwt-based discriminant features extraction from task-based fmri: An asd diagnosis study using cnn

Task-based functional magnetic resonance imaging (TfMRI) is a brain imaging modality that reveals functional activity of the brain to study the effects of a brain disease or disorder. One of the challenging brain disorders is the autism spectrum disorder (ASD) which is associated with impairments in social and linguistic abilities. Relatively few studies have applied deep learning techniques to TfMRI for diagnosing autism. This study develops discriminant TfMRI feature extraction techniques for global diagnosis of ASD by adopting a convolutional neural network (CNN) model. To achieve this goal, we propose both temporal and spatial feature extraction and reduction pipeline that consists of three main stages. The first stage involves preprocessing and brain parcellation of TfMRI scans with the fMRIB software library (FSL). The second stage reduces spatial dimensionality by extracting informative blood oxygen level-dependent (BOLD) signals after performing K-means clustering on selected brain areas exhibiting high activation in a response to speech task. Further feature reduction is applied in the temporal domain with a compression step using discrete wavelet transform (DWT) on each extracted BOLD signal. A wavelet similar to the expected hemodynamic response is selected to highlight activation information while performing DWT compression. To increase the number of the training data, an augmentation approach based on clustered data has been introduced. The third stage classifies subjects as ASD or typically developed with the deployment of deep learning 1D CNN. Preliminary results on 66 TfMRI dataset have achieved 77.2% correct global classification with 4-fold cross validation, proving high accuracy of the proposed framework.