DTI Tutorial 2 - Normalization and Statistics

Tuesday, October 13, 2015 Do Tromp 0 Comments

In our first tutorial we processed our raw images and calculated tensor images. In this second tutorial we will address how to normalize a set of diffusion tensor images (DTI) and then run statistics on the normalized brain images. First it is important to understand why you have to normalize brain images; lets see what Wikipedia has to say about spatial normalization:

"In neuroimaging, spatial normalization is an image processing step, more specifically an image registration method. Human brains differ in size and shape, and one goal of spatial normalization is to deform human brain scans so one location in one subject's brain scan corresponds to the same location in another subject's brain scan." - Wikipedia

As you see, inter-individual statistics will only be effective when brain images are in the same space, and spatial normalization can take care of that. An important distinction between diffusion tensor imaging and other imaging modality (like T1 anatomicals or fMRI scans), is that DTI allows for a more sophisticated normalization. Where T1 and fMRI scans are normalized based on intensity scales, diffusion tensor images instead offer the ability to register tensors, that have both length and direction. The property thus allows for a better capability to register white matter structures that have highly anisotropic tensors. In this video the developer of a publicly available tensor based normalization toolbox, Gary Zhang, discusses why tensor-based normalization can be better than other normalization methods, what statistical analyses you can run after normalization, and the difference between tract based spatial statistics (TBSS) and tract specific analysis (TSA). The below image details some of the advantages. For example in the area's with the yellow circles, you see areas of the brain where T1 anatomicals would have a hard time effectively registering inter-subjects anatomy for different white matter tracts, because the contrast in that region is equal despite it including multiple tracts with different orientation. Using the tensor information as show on the right will help discriminate between those regions, and allow for a more advanced registration between subjects.

Advantage of tensor specific normalization. Image by DTI-TK

In this tutorial there will be a number of steps that we have to run through in order to test statistical models, these steps are detailed below:

   1. Make sure the tensor files are in the right format for tensor normalization.

   2. Make population template by normalizing tensor images

   3. Register population template to MNI space
   4. Produce scalar images (FA, MD, AD, RD) and check quality
   5. Run whole brain voxel-wise statistics
   6. Share your data.

You can look through slides that go into some of the background of this tutorial here. Throughout the tutorial there will be questions to help with a better understanding of each of the steps and introduce you to the software that is being used. This might require installing said software and downloading example data from my website. At the end of the document you will find a pdf with the answers to the questions in this tutorial, although try to figure things out yourself as much as possible. If you run into issues going through this tutorial please direct your questions to the diffusion imaging forum, where there will likely be someone able to help you out. Please make sure to describe your question as clearly as possible. Remember to please direct software specific questions to their own respective forums or mail lists.

To get the tensor images ready for normalization we will have to make sure that tensors:
   - have the right diffusivity unit,
   - remove extreme outliers,
   - and are symmetric and positive-definite.

These steps are described in detail on the website of DTI-TK; which is the website of the software package that we will use for the tensor normalization. 

Example code:

i=052212_s09_dti_eddy_fm_strip_dt.nii
prefix=052212_s09_dti_eddy_fm_strip
echo ~~~Adjusting Diffusivity Units~~~
TVtool -in $i -scale 1000000000 -out ${prefix}_sdt.nii.gz
echo ~~Checking for and removing Outliers~~
TVtool -in ${prefix}_sdt.nii.gz -norm
SVtool -in ${prefix}_sdt_norm.nii.gz -stats
BinaryThresholdImageFilter ${prefix}_sdt_norm.nii.gz ${prefix}_non_outliers.nii.gz 0 100 1 0
TVtool -in ${prefix}_sdt.nii.gz -mask ${prefix}_non_outliers.nii.gz -out ${prefix}_tmp.nii.gz
mv -f ${prefix}_tmp.nii.gz ${prefix}_sdt.nii.gz
TVtool -in ${prefix}_sdt.nii.gz -norm
echo ~Stats for${prefix} - max should be below 100~
SVtool -in ${prefix}_sdt_norm.nii.gz -stats
echo ~~~Enforcing positive semi-definiteness~~~
TVtool -in ${prefix}_sdt.nii.gz -spd -out ${prefix}_tmp.nii.gz
mv -f ${prefix}_tmp.nii.gz ${prefix}_sdt.nii.gz
echo ~~~Standardizing Voxel Space~~~
TVAdjustVoxelspace -in ${prefix}_sdt.nii.gz -origin 0 0 0 -out ${prefix}_sdt.nii.gz
echo ~~~Resample to isotropic voxels~~~
TVResample -in ${prefix}_sdt.nii.gz -align center -size 256 256 128 -vsize 1 1 1
rm -f ${prefix}_tmp.nii.gz
rm -f ${prefix}_non_outliers.nii.gz
rm -f ${prefix}_sdt_nonSPD.nii.gz
rm -f ${prefix}_sdt_norm.nii.gz

The below figure shows how the steps are applied and combined.

Iterative normalization process. Image by DTI-TK

Example of un-normalized average FA of population of images

"In geometry, an affine transformation, affine map or an affinity (from the Latin, affinis, "connected with") is a function between affine spaces which preserves points, straight lines and planes. Also, sets of parallel lines remain parallel after an affine transformation. An affine transformation does not necessarily preserve angles between lines or distances between points, though it does preserve ratios of distances between points lying on a straight line." - Wikipedia

Example script:

echo ~~~Create an initial mean~~~;
for spd in `ls *sdt.nii.gz`;
do
subj_list_file=subj_list;
echo $spd >> subj_list_file.txt;
done
TVMean -in $subj_list_file.txt -out mean_initial.nii.gz;
TVResample -in mean_initial.nii.gz -vsize 1 1 1 -size 256 256 128;
echo ~~~Rigid alignment with initial mean~~~;
sh dti_rigid_population mean_initial.nii.gz ${subj_list_file}.txt EDS 3;
echo ~~~Affine alignment with output from rigid alignment~~~;
sh dti_affine_population mean_rigid3.nii.gz ${subj_list_file}.txt EDS 3;
echo ~~~Diffeomorphic alignment with output from affine alignment~~~;
TVtool -tr -in mean_affine3.nii.gz;
BinaryThresholdImageFilter mean_affine3_tr.nii.gz mask.nii.gz 0 .01 100 1 0;
sh dti_diffeomorphic_population mean_affine3.nii.gz ${subj_list_file}_aff.txt mask.nii.gz 0.002

Additional online resources:

- Courses on fMRI spatial normalization by Wellcome Trust centre here and here
- Courses on fMRI spatial normalization by SPM here
- Advanced normalization tools (ANTS) for normalization of scalar images
- Background documentation on normalization by ANTS
- Further documentation from DTI-TK on tensor normalization