Apply to Graduate School | Library | Jobs & Volunteers | Visitor Information | Department Index | News      
Printer Friendly
========================================================================
     NON-STATIONARITY CORRECTION TOOLBOX FOR CLUSTER SIZE P-VALUES
========================================================================



0. Disclaimer
------------------------------------------------------------------------
This toolbox is stil in a beta version. The toolbox is distributed
as is, without any warranty. 




1. Introduction
------------------------------------------------------------------------
Non-stationarity, or non-uniform smoothness, in brain imaging data
is problematic in statistical inference based on cluster extent, 
or a cluster size test. If the smoothness is not uniform, then there 
are smooth regions and rough regions within the image data. In rough 
regions, clusters tend to be small and activations/effects/differences 
may not be easily detected. On the other hand, in smooth regions, 
clusters tend to be large regardless of signals, resulting in increased 
false positives. Non-stationarity is particularly a problem in VBM 
(voxel-based morphometry) data where images are heavily warped to match
a template. Use of cluster p-values in VBM analysis has been 
discouraged [1].

To overcome this problem, Worsley et al [2] suggested a correction method.
In this method, cluster sizes are adjusted according to their local 
smoothness. In brief, clusters in smooth areas are shrunk, while clusters 
in rough areas are expanded to account for differences in smoothness. 
This non-stationarity correction can be implemented parametrically using 
random field theory (RFT) or non-parametrically by a permutation 
test [3]. The parametric version of the non-stationarity correction has 
been implemented in the FMRISTAT package as a part of STAT_THRESHOLD 
function. Accounting for non-stationarity is important as it could 
lead to an erroneous outcome of a VBM analysis, as demonstrated
in Moorhead et al [4]. 

This toolbox is an SPM version of non-stationarity correction. 
The STAT_THRESHOLD function has been ported to SPM so that it can
interface with SPM analysis results.




2. Installation
------------------------------------------------------------------------
2.0 SPM version
This package has been successfully tested on SPM2 and SPM5 on a PC,
with MATLAB version 7.1. 

2.1 Downloading
This package comes in a tar-ziped file, containing the following five
programs:

-spm_ns.m
-spm_list_nS.m
-spm_max_nS.m
-stat_thresh.m
-spm_results_ui_nS.m

2.2 Installation
To install, you save the downloaded archive file under the toolbox folder 
in your SPM. Then untar / unzip the file. A folder named ns will be 
created. 





3. Usage
------------------------------------------------------------------------
This toolbox functions much like the "RESULTS" button on SPM. After 
defining the model and estimating, instead of pressing the RESULTS
button, select "ns" under the "toolbox" pull down menu. 

Then you will go through the process quite similar to the RESULTS
portion of SPM. At the end, this toolbox produces a table of p-values
just like a typical SPM analysis. In the table, cluster sizes are
expressed in terms of voxels as well as RESELs. The corrected p-values
are corrected for non-stationarity. The uncorrected p-values are based on 
cluster sizes in RESELs rather than voxels as well. The voxel-level 
p-values are identical to that of the SPM results. The voxel-level
p-values are listed in the same order as the SPM results (for T-test
only).  

You can obtain the cluster sizes in terms of voxels, without 
non-stationarity correction by clicking the "Results" button. 





4. For Interested Users
------------------------------------------------------------------------
Interested users can learn how the non-stationarity correction works
from [3] below. Also, [2] gives an excellent overview of the concept
of non-stationarity correction.




5. Bugs / Updates
------------------------------------------------------------------------
Please note that this release is still a beta version. Thus it is quite
possible that there are some bugs, typos, and other errors in the scripts.
If you find any bugs, or if you have any questions / complaints / suggestions,
you can contact me (Satoru Hayasaka) at shayasak_at_wfubmc_dot_edu. 

DATE:    Feb 19, 2007
BUG:     Variable nvar not integer.
DETAILS: Variable nvar in stat_thresh function is treated as a double
         rather than an integer. This causes some warnings and errors
         in the program.
FIX:     In stat_thresh function, nvar is rounded so that it will be
         treated as an integer.
UPDATE:  stat_thresh has been updated. The latest distribution includes
         the latest version (ver 0.76b) of stat_thresh function.


DATE:    Mar 28, 2007
BUG:     Cluster p-values not available for F-images
DETAILS: This is not a bug, but a shortcoming in SPM. SPM cannot calculate
         cluster p-values for F-tests. 
FIX:     Cluster p-values from from stat_thresh function [2,3] are passed 
         on and displayed in the output. The p-values are corrected for
         non-sationarity. 
UPDATE:  spm_list_nS function has been updated to implement the change
         described above. The latest distribution includes the latest
         version (ver 0.8b) of spm_list_nS. 


DATE:    Mar 28, 2007
BUG:     Toolbox pull-down menu
DETAILS: Now this package can be invoked from the 'toolbox' pull-down menu
         if installed appropriately.
UPDATE:  The latest distribution includes spm_ns.m, which drives this whole
         package.


DATE:    Apr 17, 2007
BUG:     Variables hReg, xSPM, SPM are not available
DETAILS: After running the non-stationarity extension, the variables hReg,
         xSPM, and SPM are not available.
UPDATE:  The latest distribution outputs these variables in the workspace,
         so users can click on other buttons ('volume', 'cluster', etc) after
         running the non-stationarity extension.


DATE:    Apr 17, 2007
BUG:     Set-level results taking too much space
DETAILS: In the SPM output, the set-level inference results are taking up two
         columns, although very few people use the set-level p-value.
UPDATE:  The updated version of spm_list_nS (ver 0.81b) eliminated the first two
         columns of the set-level results. The set-level results are now displayed
         in the footer for interested users.


DATE:    Feb 11, 2008
BUG:     Cluster p-values unavailable for F-contrasts under stationarity
DETAILS: The NS toolbox assumes non-stationarity. Cluster sizes are adjusted based
         on RPV no matter what.
UPDATE:  The updated version of spm_list_nS (ver 0.82b) can calculate cluster 
         p-values, both for T- and F-contrasts under stationarity, provided that
         a global variable ASSUME_STATIONARY=1. Otherwise cluster p-values are
         calculated assuming non-stationarity. In the results, the table header 
         for clusters displays whether cluster p-values are calculated under 
         stationarity or non-stationarity.
 

6. Acknowledgements
------------------------------------------------------------------------
The theoretical work, as well as the main statistical programing was done by
Dr. Keith Worsley of McGill University. I would like to thank him for the
permission to port his STAT_THRESHOLD function into SPM.

Also I would like to thank Dr. Tom Nichols of GlaxoSmithKline for useful 
discussions during the development of this toolbox.

This Human Brain Project / Neuroinformatics research (R01-MH069326) 
is supported by the National Institute of Mental Health, the National
Institute on Aging, and the National Institute of Biomedical Imaging and
Bioengineering. 





7. References
------------------------------------------------------------------------
[1]. Ashburner J & Friston K J
     Voxel-based morphometry --- the methods
     NeuroImage 11: 805-821 (2000)
[2]. Worsley K J, Andermann M, Koulis T, MacDonald D & Evans A C
     Detecting changes in nonisotropic images
     Human Brain Mapping 8: 98-101 (1999)
[3]. Hayasaka S, Phan K L, Liberzon I, Worsley K J & Nichols T E
     Nonstationary cluster-size inference with random field and permutation
     methods
     NeuroImage 22: 676-687 (2004)
[4]. Moorhead T W J, Job D E, Spencer M D, Whalley H C, Johnstone E C & Lawrie S M
     Empirical comparison of maximal voxel and non-isotropic adjusted
     cluster extent results in a voxel-based morphometry study of comorbid
     learning disability with schzophrenia
     NeuroImage 28: 544-552 (2005)


------------------------------------------------------------------------
Last Updated: February 11, 2008
By Satoru Hayasaka, 
   ANSIR Laboratory, Wake Forest University
   Winston-Salem, North Carolina, USA
   http://www.fmri.wfubmc.edu/