Classification of Schizophrenia using Intrinsic Connectivity Networks and Incremental Boosting Convolution Neural...

One of the key challenges in the use of resting brain functional magnetic resonance imaging (fMRI) network analysis for predicting mental illnesses such as schizophrenia (SZ) is the high noise levels variability among individuals including age, sex, and different protocols used in labs. To deal with these challenging problems, we designed a recognition method for using brain functional networks to classify SZs and healthy controls (HCs). Our method includes two stages of training. In the first stage, we use a deep convolutional neural network (DCNN) to extract valuable deep features from functional network connectivity (FNC) images. In the next stage, these deep features are used as inputs to a gradient-boosting trees classifier. After the training process, the boosting trees classifier gains a remarkable performance compared to the DCNN classifier. We evaluate this approach using a large dataset of schizophrenia and healthy controls divided into separate validation and training sets. Experimental results showed that the recognition accuracy is over 98 %, compared to a support vector machine baseline of 77% demonstrating the ability of our system to distinguish differences between the two groups. We also estimate heatmaps for each FNC image, representing a 2D FNC matrix indicated which pairs of networks are most predictive of SZ. Our method thus provides both high accuracy, and provides insights into the relevant brain regions for SZ.