A Method of Feature Selection via Deep Convolution Neural Networks For Encoding Nonlinear Functional Network Connectivity and Its Application To The Classification of Mental Disorders

Abstract

In functional magnetic resonance imaging (fMRI) studies, it is common to evaluate the brains functional network connectivity (FNC) which captures the temporal coupling between hemodynamic signals derived from whole brain networks. FNC has been linked to various psychological phenomena. However, analysis of FNCs mainly focuses on linear statistical relationships, which may not capture the full complexity of the interactions among brain intrinsic connectivity networks (ICNs). Therefore, it is important to explore approaches that can better account for possible intricate nonlinear interactions involved in cognitive operations and the changes observed in psychiatric conditions such as schizophrenia. This exploration can lead to a better understanding of brain function and provide new insights into neural links to various psychological and psychiatric conditions. In this paper, we present an innovative approach which utilizes a deep convolutional neural network (DCNN) to extract nonlinear heatmaps from FNC matrices. By analyzing the heatmaps, multi-level nonlinear interactions can be derived from the corresponding input FNC data. Our results show these networks represent a significant improvement over previous approaches and offer a robust framework for understanding the complex interactions between brain regions. By incorporating two stages in the training process, our method ensures optimal efficiency and effectiveness. In the initial stage, a deep convolutional neural network is trained to create heatmaps from various convolution layers of the network. In the next stage, by utilizing a t-test-based feature selection method, we can effectively analyze heatmaps from different convolution layers. This approach ensures that we are able to functional connectivity with varying degrees of nonlinearity with a focus on the heatmaps that play an important role in distinguishing different groups. We used a large dataset consisting of both schizophrenia patients and healthy controls, which were divided into separate training and validation sets to evaluate this approach. Results showed patients with increases in default mode networks connections to itself and cognitive control regions and controls with increases within visual and between visual, motor, and auditory domains. We also find significantly increased cross-validated classification accuracy (at 92.8%) compared to several competing approaches. Our approach shows the potential to accurately distinguish differences between the schizophrenia and healthy control groups with high accuracy.