2009 Christine McLeavey & Jessica Tsang
Back to Psych 204 Projects 2009
Testing the Ability of TBSS to Detect Tract Variations Between Groups
TBSS is a popular program [1] for exploring anatomical connectivity in the brain by analyzing anisotropic diffusion of water in white matter tracts. In this project, we create an artificial difference between the diffusion patterns of two otherwise similar groups, and determine TBSS's ability to find this difference. In future work, we plan to test SPM/VBM in a similar manner. In this way, we effectively know the "answer", and are in a position to examine the effectiveness of available analysis tools.
Background
Tract-Based Spatial Statistics
TBSS provides a popular alternative to Voxel-Based Morphometry (VBM) analysis. It uses the following algorithm:
1. Align multiple FA images using a voxelwise nonlinear registration. TBSS attempts to strike a middle ground between low-dimensional warping (keeping brain structure, but losing alignment) and high-dimensional warping (good alignment, but extreme warping of original images). Alignment is best when a single subject's FA image is chosen to be the target. TBSS exhaustively tests using every individual subject as a target for all the others, and then chooses the best result.
2. Create mean FA skeleton. TBSS averages the transformed FA images to create a mean FA image. Tract skeleton generation then aims to create a single line through the center of the tract. It does this by sorting through voxels perpendicular to the direction of the tract, and choosing the one with the highest FA as the center of the tract.
3. Project individual subjects' FA onto group skeleton. For each individual subject, TBSS looks in the perpendicular tract direction and chooses the voxels with the maximum FA value, assigning this to the skeleton voxel. This helps to fix any misalignment from step 1. However, it introduces the possibility that a strong neighbouring tract will mistakenly be included as the tract of interest, hence our considering the Arcuate Tract to be more difficult than the Occipital Tract.
4. Run statistics to test for statistically significant FA differences between groups. Correct for multiple-comparisons, recognizing that by pure chance, statistically significant differences will be found if many voxels are compared.
Tracts of Interest
We have available data from brain scans of 55 subjects, aged 7 to 12. From this data, we choose two tracts to examine for this project, as shown below. We consider the Arcuate tract to be more difficult than the Occipital portion of the Corpus Callosum tract for TBSS, as there are several other tracts in the neighbourhood that can potentially confuse tract assignment.
Occipital Tract
Arcuate Tract
Methods
Assignment of subjects into two groups
This study began with data from 55 subjects, ages 7 to 12. 3 subjects were not included in the analysis, as they were lacking Arcuate Tract data. We assigned the remaining 52 into two groups, prioritizing:
1. gender
2. age
3. IQ
as items to match. We performed T-tests on the age, IQ, ADHD score, and mean FA distributions of the two groups, to demonstrate no statistically significant separation between the two.
Details of the group assignment are available in: File:Groups.pdf
Cleaning Occipital Tract Data
Occipital tract data was opened using MrDiffusion, and then tracts pushed into CINCH for easier viewing.
Tracts were cleaned according to the following principles (and records kept describing the cleaning for each individual brain):
1. Thin tracts clearly not connected to a main tract were removed.
2. Tracts that did not begin and end in the right plane were removed.
3. Tracts that crossed the midline more than once were removed.
4. Tracts that looped into other quadrants of the brain were removed.
Figures show a representative case. Tracts in red represent the final "cleaned" version. Those in blue indicate tracts that would be hand-removed, following the above algorithm.
Mean FA Histograms
We next examine mean FA scores for each of the 52 subjects included in the study. Results are included in File:Groups.pdf and we see that the two groups are statistically similar. Histograms of the mean FAs for the two groups are shown below, with group 1 in beige and group 2 in purple.
Shift FA in Group 1
We determine from the mean FA scores and group standard deviations that in order to create a ~1.5 standard deviation shift in Group 1, we would need to lower the FA by 8.5%. Fortunately, both Arcuate and Occipital tracts have similar SD values, so we can apply a single shift for both. The following figures demonstrate the goal outcome of a script written to produce this FA shift.
In order to do this correctly, we need to weight the FA shifting for each voxel by the fiber density of that voxel. The following figure shows the fiber density of the occipital tract, and we see that some voxels (in yellow) have much higher fiber density than others (in red). This figure also shows edges of the Arcuate tract fibers.
We use the formula:
newFA = originalFA * [1-WeightedEffectSize]
WeightedEffectSize = .085*FiberDensity
TBSS Analysis
We now enter the two groups (Adjusted Group 1 and Unadjusted Group 2) into TBSS and then run the TBSS-recommended statistics on the outcomes to determine points of significant difference identified between the two. The results are shown in the next section.
Results
TBSS Mean Skeleton
TBSS generated the following mean FA skeleton for the 52 subjects. We highlight the Arcuate and Occipital tracts, for orientation purposes. We then use these results to determine statistical differences between the two groups, as shown in the next section. We address the missing top-section of the brain in "Trouble Shooting".
Statistics
TBSS was able to detect differences in part of the Occipital portion of the Corpus Callosum tract, at a significance of p<.05. However, changes were made to the entire tract, and most of these changes were missed. Importantly, the results shown here are uncorrected for multiple comparisons. This adjustment, accounting for the likelihood that many comparisons would yield some random points of statistical significance, would further lower the statistical significance of the findings.
TBSS was basically unable to detect any difference in the Arcuate tract. It requires a setting of p<.2, uncorrected for multiple comparisons, before any signal in the Arcuate tract is seen. However, at this setting, these results can be attributed to randomness rather than a true signal, as indeed many other false positives throughout the brain are seen.
Trouble Shooting
After some searching, we determined that the top part of the brain is missing in the TBSS results because the scan data from one subject was also missing data from the top. TBSS would then throw out a section when it found it did not have data for every subject. We might exclude this subject from the study in order to have nicer pictures, however the tracts of interest are not near the missing data, and we elected to keep the subject in the study.
Conclusions
If these results are to be trusted, they would create a significant caveat for TBSS users. However, it must be emphasized that these are preliminary studies, and at this point it is quite likely that the problem lies in the algorithm used to shift the mean FA scores, or in some other part of our methodology. The next steps of this project would focus on verifying our methodology. Once we are more confident in this, we can start examining TBSS's abilities and limitations in finer detail. Additionally, we could use a similar approach to look at a comparable program (SPM/VBM).
Future Work
1. Shift Group 1 Mean FA scores by a larger amount.
2. Examine Mean FA histograms for Group 1 post-shift, to ensure they match predicted values.
3. Repeat this experiment using other tracts, or using shifts in only part of a given tract.
4. Run similar tests using SPM/VBM and compare to results with TBSS.
References
[1] Smith, Steven M. et al. Tract-based spatial statistics: Voxelwise analysis of multi-subject diffusion data. NeuroImage 31 (2006) 1487-1505.
[2] TBSS Website
Appendix I - Code
Code
Appendix II - Work partition
Christine is new to fMRI work and did much of the brunt work - she divided the kids into two groups and showed the groups to be well matched statistically. She did the cleaning of the occipital tracts for each of the 55 kids and wrote the wiki page. Jessica provided the brains and experience and devised the way to reduce the FA in the tracts of one group, ran the results through TBSS, ran statistics on those results, and generally taught Christine the ropes.