2-CHANNEL CONVOLUTIONAL 3D DEEP NEURAL NETWORK (2CC3D) FOR FMRI ANALYSIS: ASD CLASSIFICATION AND FEATURE LEARNING.

In this paper, we propose a new whole brain fMRI-analysis scheme to identify autism spectrum disorder (ASD) and explore biological markers in ASD classification. To utilize both spatial and temporal information in fMRI, our method investigates the potential benefits of using a sliding window over time to measure temporal statistics (mean and standard deviation) and using 3D convolutional neural networks (CNNs) to capture spatial features. The sliding window created 2-channel images, which were used as inputs to the 3D CNN. From the outputs of the 3D CNN convolutional layers, ASD related fMRI spatial features were directly deciphered. Input formats and sliding window parameters were investigated in our study. The power of aligning 2-channel images was shown in our proposed method. Compared with traditional machine learning classification models, our proposed 2CC3D method increased mean F-scores over 8.5%.