meeting date: 2019-02-06
days until DIPY: 32

last run: 2019-02-06
last edited: 2019-02-05
link: http://rpubs.com/navona/thesis_week01

Summary

This week, I have been working with SPINS DWI data from the Toronto site, that is accompanied by uncorrupted fieldmaps. The following analyses all include data from 73 patients (SSD) and 34 controls (HC), for a total of N=107.

I tracked how I spent my time here.

Participant characteristics

HC, n=34 (32%) SSD, n=73 (68%) p
Sex (F:M) 12:22 25:48 1.000
Age 27.94 (9.30) 27.99 (7.57) .980
Education 15.24 (1.69) 13.59 (2.03) .000
Parental education 16.18 (2.65) 15.39 (3.25) .188

As expected, we have differences in sex (M>F), but sex is roughly equal between HC and SSD groups. We also have differences in education between the HC and SSD groups, but no differences in parental education (which is defined as highest level of mother or father’s education). The groups are equal in age.

Preprocessing

Fieldmaps from CAMH contain two magnitude images with different TEs, and one phase image. I split these files, and used the script that Natalie developed for GE scanners to process them and register them to the DWI data. Then, using the same methods in general use in the lab, the DWI images were denoised, unsampled, and skullstriped, and eddy currents were corrected using FSL’s eddy (cf. eddy correct). A tensor was fit using dtifit. Data were QC’d at each step by creating html pages with slicedir.

Next step: Need to figure out how to process other sites’ fieldmaps, so can run analysis with complete sample.

TBSS

Next, I ran TBSS as per the simple instructions outlined by FSL, here. As recommended, I used the FMRIB58_FA atlas as the non-linear registration/alignment target, rather than the “most typical” participant in the dataset, though mean FA and the FA skeleton were derived from study participants (not FMRIB58_FA).

Below is an image of the derived FA skeleton (threshold of .2) on top of the average FA image.

Next step: I don’t think there’s anything left to do / improve here.

Randomise

Next, I ran voxel-wise statistics using FSL’s randomise, as per the instructions outlined here. In all cases, I ran 50 permutations, and used TFCE (threshold-free cluster enhancement) as the test statistic. Non-binary values (i.e., negative symptom score, age, parental education) were de-meaned, as is required. I filled in the design.mat and design.con text files with guidance from John, after generating templates with FSL’s design_ttest2.

HC vs. SSD

First, we ran a group comparison of HC vs SSD, without controlling for any variables. No FA differences survived FWE correction for multiple comparisons. We do see some minor clusters emerge when we substantially lower the 1-p value. The image shows the HC > SSD contrast, with a min/max display range of 0/1 (cf. .95/1), at coordinates x=92, y=106, z=75.

HC vs. SSD, controlling for age and parental education

We chose to control for age and parental education (i.e., the highest value of education from mother and father) as this is what Saba did in her 2019 Neuropsychopharmacology paper. Both age and parental education were demeaned. Again, no FA differences survived FWE correction for multiple comparisons, but did see some minor clusters emerge when we substantially lower the 1-p value. The image shows the HC > SSD contrast, with a min/max display range of 0/1, at coodinates x=90, y=147, z=76.

Negative symptoms

Lastly, we looked at negative symptoms, as represented by SANS total score (demeaned). This analysis only included the 73 SSD participants, as the SANS is not administered to HCs. I used fslsplit to unbind the mean FA image and FA skeleton volumes, and recombined those of the 73 SSD participants using fslmerge. In this analysis, we examined only one contrast, i.e., the voxels in which FA has a negative correlation with negative symptom score. We did not find an significant voxels at 1-p, though did see clusters emerge at higher p values. The image shows a min/max display range of 0/.66, at coordinates x=90, y=106, z=111.

Next steps: We were a bit perplexed at not finding differences (especially group differences), but the voxel-based statistics can be improved in several ways:

  • re-run the negative symptom analysis, controlling for age and parental education
  • re-run the negative symptom analysis, examine possible positive correlation between FA and negative symptom score
  • re-run the negative symptom analysis, using negative symptom factors (not total SANS score)
  • think of other variables that it may be important to control for (or perhaps decide not to control for age and/or parental education)
  • re-run all analyses with 500 permutations, as suggested (I ran 50 to save time; need to figure out how to run in parallel)
  • figure out how to automatically derived cluster and peak information, instead of manually/visually reviewing the images

ENIGMA DTI

After running TBSS, I extracted ROI measures from the FA skeletons, in accordance with the ENIGMA DTI protocol, which uses the JHU atlas. The plot below shows mean FA by group (HC and SSD), without controlling for any other variables.

Next step: No obvious next steps for now; will have to re-run when have all data. Also need to learn relevant neuroanatomy, and how JHU relates to MNI (used by FSL)… why do we use JHU given FSL use of MNI?

PLS

As we didn’t find any signficant differences with voxel-wise statistics (surprising, though our analysis should be improved), John suggested we try PLS, as PLS is a more sensitive measure, as it does not correct at the voxel level, and therefore maximises any differences that exist between groups.

First, we created a ‘residualized’ file by substracting the FA maps of the SSD and HC groups. Then, because PLS in MatLab requires 3D files (as opposed to 4D files required by FSL), we split the new residualized file with fslsplit, and because MatLab doesn’t like zipped files, we unzipped the data. We used the PLS GUI to generate our matrices, specify mean-centering, and indicate number of permutations (we chose 50) and bootstraps.

As with voxel-wise statistics, PLS found no group difference in FA between HC and SSD, when controlling for age and parental education. In the image below, we set the thresholds at 2/-2 (these threshold values are akin to z-scores, so 2/-2 is roughly equivalent to p=.05), and the max and min ratio, which determines the colour scale, at 3/-3 (blue indicates HCs have higher FA, and red indicates SSDs have higher FA).


Next steps: Need to review the PLS analysis; didn’t understand a few steps, and don’t understand the underlying stats. Also, John advised that the fact no differences were found with PLS might be suggestive of the possibility that something went wrong at some earlier point in the analysis, or processing.