The Effect Analysis of Atlas and Global Signal Regression in Classification based on Brain Network for Major Depression Disorders

Abstract

Automatic classification of major depression disorders (MDD) is an arduous task. When constructing the brain network automatic classification model based on functional magnetic resonance imaging (fMRI), the selection of global signal regression (GSR) and brain atlas are two key factors. However, their impact on the classification has not reached a consensus so far. The main reasons include the following two points: first, the sample size of previous studies is small, and different studies lead to inconsistent results; second, there are too many parameters in their models, which could not clearly reveal the effects of the above two factors. Therefore, we believe that only by using the data of multi-center and large samples, it is possible to find out the influence of these two factors on the classification results. To test our hypothesis, data sets (The REST-meta-MDD project) from 17 centers were used in this study. The set was divided into two parts, training set and independent validation set. The training set used 10-fold cross-validation to evaluate the classification performance, and the independent validation set used the features of the first part to classify directly. Feature selection adopted two sample t-test plus least absolute shrinkage and selection operator (LASSO), and the classifier was linear support vector machine (SVM). Finally, the classification effect of factors was confirmed by statistical analysis. The results showed that the impact of GSR on the classification results was related to the selection of brain atlas. In anatomical automatic labeling (AAL)-based networks, GSR would reduce the classification accuracy. But for Dosenbach networks, GSR would improve the classification performance. The classification ability of networks constructed by different brain templates was different, and the AAL was the best. In conclusion, the choice of brain atlas was a key factor affecting classification performance in MDD classification.