COBIDAS Easter Egg

Deep in the bowels of the COBIDAS report lives an appendix that arose from one of my own pet peeves: Why can’t people describe, concisely, the statistical model they’re using at each voxel? Mostly, authors leave out all detail and just say something like “We used FSL 5, FEAT”; or, rarely, they’ll spend multiple paragraphs detailing each step of the standard software they’re using, when a short summary would do.

This lead me to create Appendix C. Short descriptions of fMRI models in the COBIAS report (pasted below). This provides ‘one-liner’ description of commonly used software, AFNI, Freesurfer, FSL and SPM, for task fMRI and ICA. (Many thanks to Christian Beckmann, Vince Calhoun, Gang Chen, & Doug Greve who helped out on this).

Ultimately, we need to move to a place where the methods sections are grounded on machine readable provence files, so there can be no/little ambiguity about the analysis. But until then authors have a balancing act to describe with precision each analysis step, without having to re-write a paper’s worth of detail.

I’d be very intersted to know what people think of these, whether they already seem too wonky and detailed, or are leaving out too much.

Copied from COBIDAS report:

Appendix C. Short descriptions of fMRI models

While any analysis software consists of myriad modelling decisions, an author must be able to describe the key facets of an analysis in the methods section of their paper. To facilitate this, and to suggest a level of detail that is useful to readers unfamiliar with the software yet not distractingly long, we provide short descriptions for the most commonly used statistical models in widely used software packages.

C1. Task fMRI

Summaries for AFNI, Freesurfer, FSL, & SPM are based on versions AFNI_2011_12_21_1014, FreeSurfer 5.3, FSL 5.0.8 and SPM 12 revision 6470, respectively.

AFNI 1st level – 3dDeconvolve: Linear regression at each voxel, using ordinary least squares, drift fit with polynomial.
AFNI 1st level – 3dREMLfit: Linear regression at each voxel, using generalised least squares with a voxel-wise ARMA(1,1) autocorrelation model, drift fit with polynomial.
AFNI 2nd level – 3dTtest: Linear regression at each voxel, using ordinary least squares.
AFNI 2nd level – 3dMEMA: Linear mixed effects regression at each voxel, using generalized least squares with a local estimate of random effects variance.
AFNI 2nd level – 3dMVM: Multivariate ANOVA or ANCOVA at each voxel.
AFNI 2nd level – 3dLME: General linear mixed-effects modeling at each voxel, with separate specification of fixed and random variables.
Freesurfer 1st Level – selxavg3-sess: Linear regression at each surface element, using generalized least squares with a element-wise AR(1) autocorrelation model, drift fit with polynomial.
Freesurfer 2st Level – mri_glmfit: Linear regression at each surface element, using ordinary least squares.
FSL 1st level: Linear regression at each voxel, using generalized least squares with a voxel-wise, temporally and spatially regularized autocorrelation model, drift fit with Gaussian-weighted running line smoother (100s FWHM).
FSL 2nd level – “OLS”: Linear regression at each voxel, using ordinary least squares.
FSL 2nd level – “FLAME1”: Linear mixed effects regression at each voxel, using generalized least squares with a local estimate of random effects variance.
SPM 1st level: Linear regression at each voxel, using generalized least squares with a global approximate AR(1) autocorrelation model, drift fit with Discrete Cosine Transform basis (128s cut-off).
SPM 2nd level – no repeated measures: Linear regression at each voxel, using ordinary least squares.
SPM 2nd level – repeated measures: Linear regression at each voxel, using generalized least squares with a global repeated measures correlation model.

C2. Single-Modality ICA

Methods for ICA analyses are not as consolidated as mass univariate linear modelling, but we provide short summaries of some typical analyses in GIFT and MELODIC (alphabetical order), based on versions GIFTv3.0a and FSL 5.0.8, respectively. [Optional aspects, depending on particular variants used, indicated in brackets.]

GIFT, single-subject fMRI with ICASSO stability: Spatial ICA estimated with infomax where scaling of original data, spatial components and time courses constrained to unit norm, resulting best-run selected from 10 runs; post-ICA Z statistics produced for maps, between temporal component correlation (Functional Network Correlation), time courses, spectra, tested within a GLM framework.
GIFT, multi-subject PCA-based back-reconstruction with ICASSO stability: Single-subject PCA followed by temporal concatenation, group-level PCA and then spatial ICA with infomax; calculation of single subject maps using PCA-based back-reconstruction, resulting best-run selected from 10 runs; post-ICA Z statistics produced for maps, time courses, spectra, and between temporal component correlation (Functional Network Correlation) tested within a GLM framework. [Time-varying states computed using moving window between temporal components (Dynamic Functional Network Correlation).]
GIFT, spatio-temporal (dual) regression of new data: Using provided component maps calculates per-subject components from new data using regression-based back-reconstruction; produces component maps, time courses and spectra and between temporal component correlation (Functional Network Correlation) tested within a GLM framework.
GIFT, spatial ICA with reference: Spatial ICA using one or more provided seed or component maps. Components found by joint maximization of non-Gaussianity and similarity to spatial maps resulting in subject specific component maps and timecourses for each subject, scaled to Z-scores, following by testing voxelwise (within network connectivity), between temporal component correlation (Functional Network Correlation), spectra, tested within a GLM framework.
GIFT, source based morphometry of gray matter maps: Spatial ICA of multi-subject gray matter segmentation maps (from SPM, FSL, etc) resulting in spatial components and subject-loading parameters tested within a GLM framework.
MELODIC, single-subject ICA: Spatial ICA estimated by maximising non-Gaussian sources, using robust voxel-wise variance-normalisation of data, automatic model-order selection and Gaussian/Gamma mixture-model based inference on component maps.
MELODIC, group level (concat ICA): Temporally concatenation of fMRI data, followed by spatial ICA estimated by maximising non-Gaussian sources, using using robust voxel-wise variance-normalisation of data, automatic model order selection and Gaussian/Gamma mixture-model based inference on component maps
MELODIC, group-level (tensor-ICA): Higher-dimensional decomposition of all fMRI data sets into spatial, temporal and subject modes; automatic model order selection and Gaussian/Gamma mixture-model based inference on component maps
MELODIC dual regression: Estimation of subject-specific temporal and spatial modes from group-level ICA maps or template maps using spatial followed by temporal regression.

C3. Multi-Modalitiy ICA

Available multi-modality ICA methods include FIT and FSL-FLICA (alphabetical order), based on versions FITv2.0c and flica_2013-01-15, respectively.

FIT, joint ICA, two-group, fMRI + EEG fusion: Joint spatial ICA of GLM contrast maps and temporal ICA of single or multi-electrode event-related potential time course data (can be non-concurrent) with infomax ICA; produces joint component maps (each with an fMRI component map and ERP component timecourse(s)) and subject loading parameters which are then tested for group differences with a GLM framework.
FIT, N-way fusion using multiset CCA+joint ICA: Multiset canonical correlation analysis applied to several spatial maps to extract components, then submitted to spatial ICA with infomax ICA; produces multi-modal component maps and subject-specific loading parameters which are tested within a GLM framework.
FIT, parallel ICA, fusion of gray matter maps and genetic polymorphism array data: Joint spatial ICA of gray matter segmentation maps and genetic ICA of single nucleotide polymorphism data performed through a maximization of independence among gray matter components, genetic components, and subject-wise correlation among one or more gray matter and genetic components. Produces linked and unlinked gray matter and genetic components and subject loading parameters which are then tested within a GLM framework.
FSL-FLICA multi-subject/multi-modality (Linked-ICA): ICA-based estimation of common components across multiple image modalities, linked through a shared subject-courses.

Advertisements

COBIDAS report approved, posted!

The voting has closed with modest but incredibly positive turn out (153 yes, 6 no). We have added the result to the report, tacked on a linked table-of-contents, and posted to bioRxiv: http://biorxiv.org/content/early/2016/05/20/054262

The report can be referenced in different ways, but from bioRxiv:

Nichols, T. E., Das, S., Eickhoff, S. B., Evans, A. C., Glatard, T., Hanke, M., Kriegeskorte, N., Milham, M. P., Poldrack, R. A., Poline, J.-B., Proal, E., Thirion, B., Van Essen, D. C., White, T., Yeo, B. T. T. (2016). Best Practices in Data Analysis and Sharing in Neuroimaging using MRI. bioRxiv doi: 10.1101/054262.

(The DOI isn’t yet live; it can be accessed directly through the link above).

Thank you so much to those who fed us comments throughout the process. Please continue to feed any comments here, as this can influence updates to this report and other initiatives advanced by the OHBM Council.

COBIDAS: Revised, final report out! Now please vote!

You will find the final draft and information on voting at the OHBM COBIDAS page.