ArtRepair Software

The ArtRepair algorithms and software are designed to improve the accuracy of effect size estimation for fMRI data from high motion clinical subjects. While not every dataset can be recovered, tests have shown good improvements in cases with sporadic motion up to 5 mm, and more than 30% bad scans.

The software is a package of SPM-compatible tools to visually review fMRI data for artifacts, detect and repair volume, slice, and spike noise artifacts, improve estimates by deweighting selected scans, and review the quality of estimates produced after the repairs are made. Version 2.2 was released on Aug. 31, 2007.

A new feature is the “Repair and Compare” capability that can start with an existing SPM analysis, produce a new analysis with repaired data, and compare the accuracy of results before and after repairs.

This neuroimaging research and method development is supported by the National Institute of Mental Health (NIMH).

Summary

     Large artifacts in clinical fMRI data sets can occur when there are large motions by the subject. The artifacts can be larger than a 10% signal change in a voxel time series, which far dominates any cognitive effect size. These outliers can sharply degrade the accuracy of the General Linear Model, and produce activations that are falsely strong, or falsely absent. Hence, the size of activations is not a trustworthy method to validate results when there was substantial subject motion in the scanner.

     The methods here implement an alternative approach to make repairs and validate performance accuracy.

    The repair method in the ArtRepair package tries to cut all outlier scans, plus additional scans that might have outliers, in order to try to preclude any artifacts from entering from the GLM estimation process. The algorithm takes the approach that it is better to discard some good data rather than risk leaving in some massively bad data. Some test cases have shown double the accuracy (i.e., cutting the RMS error in half) when 25% of the scans were repaired.

    The results evaluation method estimates the average of contrasts (or effect sizes), where the average is taken over the ensemble of voxels in the head. It assumes that motion induced noise is much larger than cognitive effect size (i.e., the true contrast on each voxel is far smaller than the estimation error), so that the ensemble of estimates over all the voxels reflects the statistical efficiency of the GLM estimator. This accuracy metric is named Global Quality.

     Global Quality can be measured on a subject by subject basis to compare estimation accuracies between subjects, and to monitor the quality of repairs. Test cases with motions > 1 mm have shown very tight correlation (r² > 0.95) between Global Quality and actual estimation accuracy of the size of the test pulses. Since Global Quality is observable in any experiment (whereas actual accuracy is unknown), the tight correlation implies that Global Quality can be used as a measure of estimation accuracy in cognitive experiments.

   Global Quality can be used to validate the estimation performance after repairs. We have found that repairs work best when there are intermittent patches of bad data. However, for continually moving subjects, repairs may make the results worse. Thus, we recommend repairing the data, compare the Global Quality before and after repairs, and use the repaired data if the score is better. (The “Repair and Compare” button in the software will run this procedure.) Adding motion regressors and visually checking for noisy slices can help in difficult cases. The repair method does not affect normal subjects because they have few, if any, outliers at the default threshold settings.

    Research is continuing to make the repairs more reliable over a wider variety of experimental conditions. In the meantime, this software is released because it works well for at least some fMRI data sets that may not have been usable using standard GLM methodologies.

Documentation

ArtRepairOverview.pdf

Overview presentation of the software features.

ArtRepairInstructions.txt

Step-by-step instructions for running ArtRepair.

ArtRepairInstallation.txt

Instructions for installing ArtRepair into SPM toolbox.

ArtRepairHBM2005.pdf

Poster from Human Brain Mapping conference in 2005.

ArtRepairHBM2007.pdf

Poster from HBM 2007, including 3D large motion
correction.

Software Download and Installation

The software is free, but we request that users register to download the software in order to help us track its usage.

Download Software

Install the software by putting the ArtRepair2 software into the SPM2 Toolbox folder, or ArtRepair5 into the SPM5 toolbox folder. Select ArtRepair2 or ArtRepair5 within the SPM toolbox menu to start it.

Version 2.2 adds a GUI compatible with Matlab 6.5, automatically scales Global Quality, and fixes bugs in the ArtifactRepair and Repair and Compare functions.

Requirements: These programs assume that MATLAB and SPM2 or SPM5 are installed. The ArtRepair programs use the SPM read/write capabilities, and thus use AnalyzeFormat images with SPM2 and Nifti images with SPM5. The programs have been tested in Matlab Versions 6.5 - 7.1, with SPM2 and SPM5, on RedHat Linux and Windows XP.

Disclaimer:  This software is made available to promote better understanding and quality review of fMRI data. This software is supplied as is. No formal quality assurance checks were made on the software, and no formal support or maintenance is provided or implied.

Credits

    This software was written by Paul Mazaika, at the Center for Interdisciplinary Brain Science Research at Stanford. It is derived from software developed by Susan Whitfield-Gabrieli, Paul Mazaika, and Jeffrey C. Cooper in the Gabrieli Cognitive NeuroSciences Lab. Please send any bug reports, questions, or comments to mazaika_ AT_ stanford.edu.

The best citations for this software are:

“Artifact Repair for fMRI Data from High Motion Clinical Subjects”, by Paul Mazaika, Susan Whitfield-Gabrieli, and Allan Reiss, presentation at Human Brain Mapping conference, 2007.

“Detection and Repair of Transient Artifacts in fMRI Data”, by Paul Mazaika, Susan Whitfield, and Jeffrey C. Cooper, Human Brain Mapping conference, 2005.