On 30 Jun 2016, at 23:35, Julian Yaoan Cheng <[log in to unmask]> wrote:Hello Eduardo,Thank you for your suggestions.I will review the papers on aCompCor and see how I can integrate that with the fsl modeling.I do include the motion parameters via the rp_*.txt files from SPM, but this was admittedly an inadequate description since it would not have been obvious without familiarity with SPM.Best,JulianFrom: FSL - FMRIB's Software Library <[log in to unmask]> on behalf of Eduardo Garza-Villarreal <[log in to unmask]>
Sent: Wednesday, June 29, 2016 6:44:57 PM
To: [log in to unmask]
Subject: Re: [FSL] Generating ALFF and fALFF from resting state dataHi Julian,It's ok if you don't use Global Signal, but you have to use something. I recommend aCompCor: (http://www.sciencedirect.com/science/article/pii/S1053811911009657)Now, if you are using FSL then you need to work on the Residual data, not the actual GLM results. The residual 4D rsfmri nifti file is the one you need to use to continue. I suggest you include the motion parameters, not just WM and CSF by the way, in the GLM.I can't comment exactly on how to do it in FSL because I've only used AFNI for the GLM and fALFF.Hope this helps.Kind regardsEduardoEduardo A. Garza-Villarreal, M.D., Ph.D.
- CONACYT Research Fellow, Investigaciones Clínicas, Instituto Nacional de Psiquiatría, Mexico.
- Visiting Senior Researcher, Center of Functionally Integrative Neuroscience, University of Aarhus, Denmark.
On Wed, Jun 29, 2016 at 7:22 PM, Julian C <[log in to unmask]> wrote:Hello experts,
I've been trying to replicate the methods in this paper using FSL 5.0.9 but have been having trouble progressing: "Hoptman et al 2010. Amplitude of low-frequency oscillations in schizophrenia: A resting state fMRI study". The relevant section on the processing is reproduced below:
--
Resting state data were preprocessed as described elsewhere in detail (Margulies et al., 2007; Castellanos et al., 2008; Kelly et al., 2008). Briefly, the first 10 volumes were discarded to eliminate T1 relaxation effects. Images were then motion-corrected, time shifted, and despiked using AFNI (Cox, 1996). Next, time series were smoothed using a 6 mm FWHM Gaussian kernel and spatially normalized to MNI space (2×2×2 mm 3 resolution) using FSL (www.fmrib.ox.ac.uk/fsl). The MPRAGE
image was segmented using FSL's FAST software to obtain the masks used to extract white matter and CSF time series from the BOLD images. The white matter (WM) and cerebrospinal fluid (CSF) time series were then averaged across voxels within each mask. These time series, as well as the global signal intensity and
the time series for the six motion parameters were regressed out from each voxel's time series. Individual participant analyses were carried out with the GLM
implemented in FSL's FEAT and power spectral density toolbox. The time series for the nuisance covariates (time series regressors for global signal, WM, CSF, and six motion parameters) were entered as predictors. The residual data were then linearly detrended. Then, based upon detrended residual data, the power
spectral density of the data was calculated using FSL's tool (fslpspec). Two measures were obtained: ALFF and fALFF. ALFF represents the amplitude in the low-frequency band, whereas fALFF is the ratio of the amplitude in a low-frequency band to amplitude in the total frequency band. Here, the low-frequency
range was the slow-4 band (0.027–0.073 Hz; Buzsaki and Draguhn, 2004; Di Martino et al., 2008). The details of computation can be found in Zang et al. (2007) and Zou et al. (2008). The resulting amplitude measures were converted into Z-scores by subtracting the global mean and dividing the global standard deviation.
--
My background is in SPM so for the preprocessing portion I elected to use SPM12 to do the job since I'm more familiar with that; this includes segmentation of the whitematter/CSF masks. I will describe what I've done in the following sections, and if you have time to point out any errors I would very much appreciate it (the part where I am actually stuck on will be in the last section).
[time series extraction]
Using the segmented masks from SPM I used fslmeants to extract time series for both WM/CSF to be used later in the GLM as regressors. This part was fairly straight forward.
[GLM]
I opened the FEAT GUI interface and chose "First-level analysis" and "Statistics", then loaded in the 4D preprocessed image and confirmed both the number of volumes and TR were correct (I left the high pass filter cutoff at the default value of 100). On the "Stats" tab I changed to "Don't Add Motion Parameters" and specified the rp_*.txt file generated by SPM's realignment module using the "Select confound EVs text file(s)" button. Then, in the "Full model setup" popup window I specified 2* original EVs using the "Custom (1 entry per volume)" setting for basic shape, and selected "None" for convolution. I then left "Add temporal derivative" and "Apply temporal filtering" options enabled by default. The model was run with everything else not specified left at its default state.
*Note: I only specifed 2 because I am only supplying WM and CSF regressors and elected not to include the global signal due to concerns on co-linearity.
[ALFF and fALFF]
This is where I am stuck. First the paper noted that they linearly detrended the residual data, but it is not clear to me how they did this (or why, as most of my search seemed to suggest detrending done prior to modeling, not after). Furthermore I cannot figure out how to use fslpspec to obtain the ALFF and fALFF measures as it doesn't seem to be producing any output when fed with the residual image (res4D.nii.gz).
Any help will be appreciated as I'm currently totally stumped.
Best,
Julian