Data Representations and Ensemble Deep Learning Networks for Functional Neuroimaging Datasets

This project was designed to test the predictive accuracy of combining two separate data representations of resting state functional magnetic resonance imaging (rs-fMRI) data into an ensemble deep learning architecture. Three main data representations of the same neuroimaging dataset were tested by building associated deep learning architectures and testing their accuracy in predicting if the neuroimaging data originated from healthy controls or from individuals diagnosed with autism spectrum disorder (ASD). The three data representations were 2D correlation matrices derived from time courses extracted from the blood-oxygen-level-dependent (BOLD) signal within the brain, a graph tensor representation of the same connectivity data, and a 3D profile of the posterior cingulate cortex’s (PCC) connectivity across the brain. These data representations were fed into a 2D Convolutional Neural Network (2D-CNN), a Graph Convolutional Neural Network (GCN), and a 3D Convolutional Neural Network (3D-CNN) respectively. Finally, the 2D-CNN and the 3D-CNN were chosen to combine into a single ensemble model to test the hypothesis that the combination of two different representations of the same data can improve upon the individual models. This ensemble model performed better than both the 2D-CNN and 3D-CNN models individually when validated using 5-fold cross-validation and 5 × 2-fold cross validation. However, this improvement was only statistically significant for the comparison with the 3D-CNN model (p = 0.0224). This result suggests that using combinations of multiple data representations may improve model accuracy when using functional neuroimaging data in deep learning applications.