A Sparse Multi-Scale Convolutional Neural Network to Classify Functional Connectivity for Brain Disease Diagnosis

Adopting deep learning methods to classify functional connectivity (FC) patterns for brain disease diagnosis has become a new research hotspot. Recently, deep neural network (DNN) with fully connected structure has been widely applied. However, the dense connections between adjacent layers bring abundant trainable parameters. Moreover, the current DNN models only extract FC features along one dimension, which lack potential multi-channel descriptions for the FC features. With the rapid development of convolutional neural network (CNN), various CNN-based deep learning models have been applied in pattern recognition. In this study, we proposed a sparse multi-scale CNN (SMS-CNN) model to classify FC patterns for brain disease diagnosis. In our model, multi-scale convolution operations were conducted to aggregate the FC features from different scales of spatial neighborhoods. In addition, to further alleviate over-fitting, we designed a lightweight feature sparse layer to scale the redundant FC features and increase the sparsity of the model. We conducted systematic experiments on the large-scale Autism Brain Imaging Data Exchange (ABIDE) dataset to validate the classification performance of the proposed model. Experimental results showed that the proposed SMS-CNN can better learning and classifying FC patterns and achieved high classification performance in distinguishing autism patients from healthy controls. Our study indicates the potential of CNN-based models in FC pattern analysis, and provides a promising method to further improve the classification performance for FC-based brain disease diagnosis.