3952
Global Signal Regression in Resting-state fMRI Pre-processing Improves Classification Accuracy1Brainnetome Center, Institute of Automation, Chinese Academy of Sciences, Beijing, China, 2National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China, 3University of Chinese Academy of Sciences, Beijing, China, 4Department of Radiology, Renmin Hospital of Wuhan University, Wuhan, China, 5Department of Psychiatry, Xijing Hospital, The Fourth Military Medical University, Xi'an, China, 6Zhumadian Psychiatric Hospital, Zhumadian, China, 7Peking University Sixth Hospital / Institute of Mental Health, Beijing, China, 8Key Laboratory of Mental Health, Ministry of Health (Peking University), Beijing, China, 9Department of Psychiatry, Henan Mental Hospital, The Second Affiliated Hospital of Xinxiang Medical University, Xinxiang, China, 10Henan Key Lab of Biological Psychiatry, Xinxiang Medical University, Xinxiang, China, 11Department of Psychology, Xinxiang Medical University, Xinxiang, China, 12Center for Life Sciences / PKU-IDG / McGovern Institute for Brain Research, Peking University, Beijing, China, 13Key Laboratory for NeuroInformation of Ministry of Education, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China, 14Queensland Brain Institute, University of Queensland, Brisbane, Australia, 15CAS Center for Excellence in Brain Science and Intelligence Technology, Institute of Automation, Beijing, China
Synopsis
Global signal regression (GSR) is under debate whether or not influences the interpretation of functional connectivity (FC). However, few studies have compared and discussed the classification performance of GSR on a large dataset. We used a large dataset of resting-state fMRI data with 1082 subjects to test whether GSR influences the FC-based classification performance. We reached 81.35%-84.36% test accuracy using nested cross-validation. We tested the contribution of GSR, feature whitening and classifiers to the classification accuracy variance using three-way ANOVA and found significant main effects only for the GSR factor (F=7.14, P=0.0089). The results suggest GSR improves the classification accuracy.
Introduction
Global signal regression (GSR) was introduced to suppress psychological noises in time series of fMRI images. However, studies of whether GSR suppresses the psychological noises [1] or distorts signal have not reached a convention [2-4]. Meanwhile, few studies have estimated the impact of GSR on classification performance using a large dataset. In the current study, we compared classification performance using functional connectivity (FC) calculated from a large resting-state fMRI dataset, of 552 schizophrenia patients and 530 normal controls, pre-processed with and without GSR. We tested whether will GSR influence the classification performance by a three-way analysis of variance (ANOVA) full model with GSR, feature whitening and classifier as factors.Methods
FMRI pre-processing and functional connectivity (FC).
The pre-processing pipeline includes removing first 10 timepoints of EPI images, slice timing correction, within subject head motion correction, linear registration of T1 image to EPI images, nonlinear transformation to MNI standard space (voxel size=3*3*3 mm3) using T1 image, noise regression, 0.01-0.08 Hz band-pass filtering and smoothing with 6x6x6 mm3 Gaussian kernel. Noise regressors include linear trends, mean signal of white matter (WM), cerebrospinal fluid (CSF) and global signal (GS), the derivative of WM, CSF and GS, and Friston’s 24 parameters’ head motion model. The pipeline after spatial normalization was run for a second time without GS components in the regression model. We calculated Fisher-z transformed functional connectivity (FC) using 246 brain regions defined in the Brainnetome atlas [5] for pre-processed images with and without GSR. We used BRANT [6] to run the pre-processing and FC calculation pipeline.
Classification
We used 3-hidden layer multilayer perceptron (MLP), logistic regression (LR), support vector machine with linear kernel (linear-SVM) and support vector machine with radical basis function kernel (RBF-SVM). We used nested cross-validation and leave-one-site-out strategy to select model hyperparameters with highest validation accuracy in the 7-fold inner loop and tested the classification accuracy on each site in the outer loop. Before training classifiers in each inner loop, FCs as feature were first whitened by regressing out the average feature component in training set. Both whitened and not whitened FCs were then scaled to zero-mean and unit-variance using the mean and variance estimated from training set.
Three-way ANOVA
We used the full model of three-way ANOVA to test whether different factors (GSR, whitening, classifier) or their interactions influence on the classification accuracy.
Results
The mean test accuracy ranging from 81.35% to 84.36% across all methods (Figure 1). Significant main effects were found for the GSR factor (F=7.14, P=0.0089, Table 1), with higher classification accuracy observed for the GSR groups. No other significant main effects or interactions were found.Discussion
The improvement of classification accuracy implies the GSR has increased the separability of two groups in both the original feature space (suggested by the LR and the linear-SVM) and the nonlinearly transformed feature space (suggested by the MLP and the RBF-SVM). The results offered a suggestion for neural network classifiers to improve classification performance, considering the nonlinear classifiers work as ‘black boxes’ and the classification performance is more concerned than the interpretation. The results have some limitation on the use of sample and whitening methods. The results could be different in other binary classification cases since the sample we used only includes schizophrenia patients and normal controls. We used the regression methods for whitening to keep the input dimension same as not whitened feature, which could be different using feature dimension reduction methods.Conclusion
The GSR improves the FC-based classification accuracy.Acknowledgements
No acknowledgement found.References
1. Ciric, R., et al., Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. Neuroimage, 2017. 154: p. 174-187.
2. Saad, Z.S., et al., Trouble at rest: how correlation patterns and group differences become distorted after global signal regression. Brain Connect, 2012. 2(1): p. 25-32.
3. Fox, M.D., et al., The global signal and observed anticorrelated resting state brain networks. J Neurophysiol, 2009. 101(6): p. 3270-83.
4. Murphy, K., et al., The impact of global signal regression on resting state correlations: are anti-correlated networks introduced? Neuroimage, 2009. 44(3): p. 893-905.
5. Fan, L., et al., The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture. Cereb Cortex, 2016. 26(8): p. 3508-26.
6. Xu, K., et al., BRANT: A Versatile and Extendable Resting-State fMRI Toolkit. Front Neuroinform, 2018. 12: p. 52.
Figures
