%%% FSLNets - simple network matrix estimation and applications %%% FMRIB Analysis Group %%% Copyright (C) 2012-2014 University of Oxford %%% See documentation at www.fmrib.ox.ac.uk/fsl %%% change the following paths according to your local setup addpath /home/fs0/steve/NETWORKS/FSLNets % wherever you've put this package addpath /home/fs0/steve/matlab/L1precision % L1precision toolbox addpath /home/fs0/steve/matlab/pwling % pairwise causality toolbox addpath(sprintf('%s/etc/matlab',getenv('FSLDIR'))) % you don't need to edit this if FSL is setup already %%% setup the names of the directories containing your group-ICA and dualreg outputs group_maps='your-group-ICA.ica/melodic_IC'; % spatial maps 4D NIFTI file, e.g. from group-ICA %%% you must have already run the following (outside MATLAB), to create summary pictures of the maps in the NIFTI file: %%% slices_summary 4 $FSLDIR/data/standard/MNI152_T1_2mm .sum ts_dir='groupICA.dr'; % dual regression output directory, containing all subjects' timeseries %%% load timeseries data from the dual regression output directory ts=nets_load(ts_dir,3,1); %%% arg2 is the TR (in seconds) %%% arg3 controls variance normalisation: 0=none, 1=normalise whole subject stddev, 2=normalise each separate timeseries from each subject ts_spectra=nets_spectra(ts); % have a look at mean timeseries spectra %%% cleanup and remove bad nodes' timeseries (whichever is not listed in ts.DD is *BAD*). ts.DD=[1 2 3 4 5 6 7 9 11 14 15 17 18]; % list the good nodes in your group-ICA output (counting starts at 1, not 0) % ts.UNK=[10]; optionally setup a list of unknown components (where you're unsure of good vs bad) ts=nets_tsclean(ts,1); % regress the bad nodes out of the good, and then remove the bad nodes' timeseries (1=aggressive, 0=unaggressive (just delete bad)). % For partial-correlation netmats, if you are going to do nets_tsclean, then it *probably* makes sense to: % a) do the cleanup aggressively, % b) denote any "unknown" nodes as bad nodes - i.e. list them in ts.DD and not in ts.UNK % (for discussion on this, see Griffanti NeuroImage 2014.) nets_nodepics(ts,group_maps); % quick views of the good and bad components ts_spectra=nets_spectra(ts); % have a look at mean spectra after this cleanup %%% create various kinds of network matrices and optionally convert correlations to z-stats. %%% here's various examples - you might only generate/use one of these. %%% the output has one row per subject; within each row, the net matrix is unwrapped into 1D. %%% the r2z transformation estimates an empirical correction for autocorrelation in the data. netmats0= nets_netmats(ts,0,'cov'); % covariance (with variances on diagonal) netmats0a= nets_netmats(ts,0,'amp'); % amplitudes only - no correlations (just the diagonal) netmats1= nets_netmats(ts,1,'corr'); % full correlation (normalised covariances) netmats2= nets_netmats(ts,1,'icov'); % partial correlation netmats3= nets_netmats(ts,1,'icov',10); % L1-regularised partial, with lambda=10 netmats5= nets_netmats(ts,1,'ridgep'); % Ridge Regression partial, with rho=0.1 netmats11= nets_netmats(ts,0,'pwling'); % Hyvarinen's pairwise causality measure %%% view of consistency of netmats across subjects; returns t-test Z values as a network matrix %%% second argument (0 or 1) determines whether to display the Z matrix and a consistency scatter plot %%% third argument (optional) groups runs together; e.g. setting this to 4 means each group of 4 runs were from the same subject [Znet1,Mnet1]=nets_groupmean(netmats1,1); % test whichever netmat you're interested in; returns Z values from one-group t-test and group-mean netmat [Znet5,Mnet5]=nets_groupmean(netmats5,1); % test whichever netmat you're interested in; returns Z values from one-group t-test and group-mean netmat %%% view hierarchical clustering of nodes %%% arg1 is shown below the diagonal (and drives the clustering/hierarchy); arg2 is shown above diagonal nets_hierarchy(Znet1,Znet5,ts.DD,group_maps); %%% view interactive netmat web-based display nets_netweb(Znet1,Znet5,ts.DD,group_maps,'netweb'); %%% cross-subject GLM, with inference in randomise (assuming you already have the GLM design.mat and design.con files). %%% arg4 determines whether to view the corrected-p-values, with non-significant entries removed above the diagonal. [p_uncorrected,p_corrected]=nets_glm(netmats1,'design.mat','design.con',1); % returns matrices of 1-p %%% OR - GLM, but with pre-masking that tests only the connections that are strong on average across all subjects. %%% change the "8" to a different tstat threshold to make this sparser or less sparse. %netmats=netmats3; [grotH,grotP,grotCI,grotSTATS]=ttest(netmats); netmats(:,abs(grotSTATS.tstat)<8)=0; %[p_uncorrected,p_corrected]=nets_glm(netmats,'design.mat','design.con',1); %%% view 6 most significant edges from this GLM nets_edgepics(ts,group_maps,Znet1,reshape(p_corrected(1,:),ts.Nnodes,ts.Nnodes),6); %%% simple cross-subject multivariate discriminant analyses, for just two-group cases. %%% arg1 is whichever netmats you want to test. %%% arg2 is the size of first group of subjects; set to 0 if you have two groups with paired subjects. %%% arg3 determines which LDA method to use (help nets_lda to see list of options) [lda_percentages]=nets_lda(netmats3,36,1) %%% create boxplots for the two groups for a network-matrix-element of interest (e.g., selected from GLM output) %%% arg3 = matrix row number, i.e. the first component of interest (from the DD list) %%% arg4 = matrix column number, i.e. the second component of interest (from the DD list) %%% arg5 = size of the first group (set to -1 for paired groups) nets_boxplots(ts,netmats3,1,7,36); %print('-depsc',sprintf('boxplot-%d-%d.eps',IC1,IC2)); % example syntax for printing to file