Hi Germaine,

Did you first denoise your data (e.g. with MRtrix3)? That can make a difference to the results.

Which version of FSL did you use?

Did you use eddy_openmp or the old “eddy” (which had problems)?

Did you check the mask for each subject to make sure that no brain voxels were excluded?

Did you check the outputs of tbss_1_preproc (to make sure that no brains were flipped or upside down)?

Check out my handbook for further information on this (http://jeromemallershandbookofstructuralbrainmrianalysis.yolasite.com/ )

I have often found that there is no significant difference between groups for FA and MD. Instead, the differences may be in their constituents i.e. axial (L1) or radial (average of L2 and L3) diffusivity.

Did you control/covary for ICV or brain volume and/or age?

For your FA and MD findings, were there group differences at any level of significance at all, even p=0.5?

Regards,

Jerome

Hi Jerome,

Yes I did denoise the data by conducting eddy. I was not able to do FUGUE or TOPUP as the scans were done only in one direction and FUGUE distorts the images instead of correcting them.

The FSL version that I’m currently using is FSL 5.0.10 although I ran the old “eddy” command in terminal as eddy_openmp command in my case generated an empty eddy_outlier_report file and did not tell me where the outliers are in the eddy_outlier_map file.

Also, I did check the mask for each subject and the outputs of tbss via the overview webpage generated after tbss_1_preproc.

And for ICV, I placed age and gender as covariates.


And even at p = 0.5 (In fslview, min = 0.5, max = 1), there were no significant group differences. Some slight differences appeared only at p = 0.7 (In fslview, min = 0.3, max = 1).

I will try to run AD and RD to see if there are differences. If there are no differences, what would you suggest may be the problem? Should I recruit more participants or do more corrections?

And thank you for creating and sharing the handbook! Appreciate it much!

Thank you,

Germaine

Hi Germaine,

Eddy does not denoise data. Denoising is done with MRtrix3 (although other packages can denoise as well, I like MRtrix3) using the ‘dwidenoise’ command.

The old ‘eddy’ is now known to make data worse in some cases. So, I strongly suggest you use eddy_openmp (after denoising the data). Even though the eddy_outlier_report file was empty, were eddy_corrected output files generated? That is the most important thing.

FSL 5.0.10 - excellent. Checking the mask for each subject – excellent. Covarying for potential confounds – excellent.

Yes, please run AD and RD. If there are no differences, then it is possible that either a) there really are no diffusivity differences between the groups, or b) there are outliers that are skewing the results.

However, if TBSS shows no FA, MD, AD, or RD differences, it is possible that the differences (if they exist) are not voxel-based but rather WM bundle based. That is, there may not be voxel-by-voxel differences, but rather, differences when entire bundles are considered. So, in that case, I would suggest you calculate the FA, MD, AD, and RD for each of the major WM bundles (there are scripts in my handbook that you can copy and paste into your terminal), export those numbers to your preferred stats package (SPSS, Excel, etc) and see whether there are group differences.

Keep me informed!

Jerome

Hi Jerome,

Thanks so much for your prompt reply and guidance.

I just ran AD and no differences were found between both groups at p=0.05.

By your suggestion in analysing the FA, MD, AD and RD for each major WM bundles, are you suggesting tractorgraphy (and which can be found in your handbook chapter 2?)

Thanks and have a good day ahead!

Germaine Chew

Hi Germaine,

Instead of tractography, you can extract mean FA etc by using the predefined bundles as per the JHU atlas included in FSL. Have a look at the my handbook, Part 1, pages 58 to 63.

Regards,

Jerome

Hi Jerome, 

I need your expertise again for the following question I have! 

Instead of calculating FA, MD, AD and RD for each of the major WM bundles, due to the low sample size, I ended up trying to see if there are any significant regions at p = 0.05 in the files *_tfce_p_tstat1.nii.gz and *tfce_p_tstat1.nii.gz.

There were some significant regions but I’m only interested in the arcuate fasciculus, cerebellum and inferior frontal gyrus (which the atlas doesn’t seem to have). 

I was thus wondering if you’ll be able to know how I can find out whether these regions are significantly different between the two groups and how I can extract values out so that I can quantitatively analyse the data? I tried the cluster command which gave me the Voxels, max, max x, y z and cog x, y z but I’m not sure how I’m able to analyse the data.

My main aim is to find out if the ROI differs in both groups and how much significantly different they are. For this, do I need to do a t test? And if I were to do a t test for correlation, do I need a behavioural measure to do so?

Thanks so much for your help!

Germaine  

Hi Germaine,

You could use a different atlas to extract the values for the arcuate, cerebellum and IFG. For example, you could use the Talairach atlas, Probabilistic cerebellar atlas, AAL, or you could manually make masks for the cerebellum, arcuate and IFG using the MNI template and then use fslmeants to extract the values.

Yes, a t-test would be appropriate (or an ANOVA) to look for differences between groups. If performing a Pearson correlation then (of course) you will need something to correlate FA against, so a behavioral variable would be appropriate. You could run correlation analyses against almost anything, including time since diagnosis, time since treatment, age etc.

Regards,

Jerome

Hi Jerome, 

Thanks a lot for the prompt response and for helping me understand DTI better. I really appreciate it greatly. 

As I’m still trying to understand how to extract the FA values out (which a lot of lab uses a script to do so)..thus can I check with you that the steps below are how one goes about extracting FA values from a significant ROI? 

fslmaths /usr/local/fsl/data/atlases/JHU/JHU-ICBM-labels-1mm.nii.gz -thr 41 -uthr 41 -bin Rsuperiorlongifasci.nii.gz

• This command just selects the ROI

fslmaths  Rsuperiorlongifasci.nii.gz -mas mean_FA_skeleton_mask.nii.gz -bin Rsuperiorlongifascimask.nii.gz  

• The ROI is overlaid on top of the mean_FA_skeleton_mask and the resulting Rsuperiorlongifascimask consists of voxels where the ROI overlaps with the mean_FA_skeleton_mask. )

fslmeants -i *_tfce_p_tstat1_fill.nii.gz -m Rsuperiorlongifascimask.nii.gz -o RsuperiorlongifasciFA1.txt

• *_tfce_p_tstat1_fill.nii.gz has been thr at p = 0.05
• Only one value will be given and the value given is the mean FA of the ROI, provided if there is a significant difference between the two groups at p=0.05.
• If that particular ROI is not significant between both groups, the mean FA value will be 0.

From the text files provided, I can compare the FA values between two groups. E.g Group B (tstat2) has higher FA values than Group A (tstat1)?But, this then brings up another question, how can I calculate how many voxels are activated in that ROI? Cluster command doesn't seem to work for this as based on the atlases, some regions reported to be significant in the cluster file are e.g. 80% frontal pole, 2% superior longitudinal fascicules. etc.

Thank you very much for your help.

Germaine

Germaine,

I think that the second fslmaths step is redundant (you don't have to mask it twice).

Regarding the fslmeants, if you're just interested in extracting an atlas-based ROI FA value, you need to use as an input volume the all_FA.nii.gz, not the *_tfce_p_stat1 (the last one is the result from a contrast in a skeletonized map, the atlas is not skeletonized)

Second, fslmeants won't tell you if there are significant differences between groups. You should test this either in randomise or in SPSS/R/matlab with the extracted values.

Third. I followed this question and I believe you have results only with the uncorrected contrast (not FWE-corrected). I advise you against performing a second analysis of already existing results (double dipping). Cluster/fslmeants commands are normally used to extract values in order to plot results or fill tables.

Forth, the cluster/fslmeants commands will only give you the mean of the "intensity", which will depend on the input file. In your case, you will extract the mean FA value of that ROI (not the activation).

If you're interested in extracting the FA of a *_tfce_tstat1, you may use instead the following:

This is going to return you a *.nii.gz and a text file with your thresholded results, the cluster number, its size, the p-value and its coordinates in MNI

cluster -i *_tfce_tstat1.nii.gz -t 0.95 --oindex=output_index > output_results.txt

This other will binarize your significant clusters, ordered in the same numbers that in your output_index.txt (put your cursor in top of the cluster and see the "intensity" box in FSLeyes/view). Assuming your text file depicts 2 clusters.

fslmaths –dt int output_index -thr 1 -uthr 1 -bin cluster1_mask

fslmaths –dt int output_index -thr 2 -uthr 2 -bin cluster2_mask

Finally, this will extract in a text file all the FA values of the skeletonized results significant results (do this for each cluster).

fslmeants –i all_FA_skeleton.nii.gz –m cluster_mask1 > results_cluster1.txt

...

If you're also interested in reporting the label names according to the atlases in FSL for white-matter. You can feed the masks into this command.

bash atlasquery -a “JHU White-Matter Tractography Atlas” -m cluster_mask1

or

bash atlasquery -a “JHU ICBM-DTI-81 White-Matter Labels” -m cluster_mask1

Hope this helps!

J

Hi Jonatan,

thanks for the detailed steps and guidance.

They were really helpful given that I'm still learning about DTI and how to analyse it.

I will try the steps you mentioned and see how it goes.

Thanks so much! :)

Germaine