DTI Analysis using rcamino for HCP data

John Muschelli


All code for this document is located at here.

Resources and Goals

Much of this work has been adapted by the Tutorial for DTI analysis from ISMRM 2015: http://camino.cs.ucl.ac.uk/index.php?n=Tutorials.ISMRM2015. Also, some of the model fitting (such as using gradient information) has been taken from http://camino.cs.ucl.ac.uk/index.php?n=Tutorials.HCP.

We will show you a few steps that have been implemented in rcamino: camino_pointset2scheme, camino_modelfit, camino_fa, camino_md, and camino_dteig.

Data Location

Reading in the Data

First, we download the data from HCP. You must have your access keys set (see Getting Data from the Human Connectome Project (HCP)).

We will use the neurohcp package to download one subject data.

hcp_id = "100307"
r = download_hcp_dir(
  paste0("HCP/", hcp_id, "/T1w/Diffusion"), 
  verbose = FALSE)
[1] "bvals"                   "bvecs"                  
[3] "data.nii.gz"             "grad_dev.nii.gz"        
[5] "nodif_brain_mask.nii.gz"

It contains 4 files:

  1. data.nii.gz - a 4D image of the DWI data.
  2. nodif_brain_mask.nii.gz - A brain mask of the DTI data
  3. bvals - a text file with the b-values
  4. bvecs - a text file with the b-vectors as the first 3 columns.


As dtifit requires the b-values and b-vectors to be separated, and this data has b-values of \(1000\) when the b-vectors is not zero. This is very important and you must know where your b-values and b-vectors are when doing your analyses and what units they are in.

camino_set_heap(heap_size = 10000)
outfiles = r$output_files
names(outfiles) = nii.stub(outfiles, bn = TRUE)
scheme_file = camino_fsl2scheme(
  bvecs = outfiles["bvecs"], bvals = outfiles["bvals"],
  bscale = 1)
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/fsl2scheme -bvecfile /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b5fd527a4//bvecs -bvalfile /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b5fd527a4//bvals -bscale 1   >  /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b517d29e01.scheme

The imaging scheme contains measurements at b=5, b=1000, b=2000, and b=3000 s / mm^2.

Subsetting data

By selecting a subset of the measurements, we can reduce processing time and memory requirements. Also, the high b-value shells aren’t optimal for estimating the diffusion tensor. So we’ll select data from the b=5 and b=1000 shells, which is still higher angular resolution than most DTI (90 directions).

If you have sufficient RAM, you can load the whole data set and extract a subset:

camino_ver = packageVersion("rcamino")
if (camino_ver < "0.5.2") {
sub_data_list = camino_subset_max_bval(
  infile = outfiles["data"],
  schemefile = scheme_file,
  max_bval = 1500,
  verbose = TRUE) 
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/split4dnii -inputfile /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b5fd527a4/data.nii.gz -outputroot /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b5439819d9_
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/image2voxel -imagelist '/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b536c9183a.txt' -outputfile '/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b54c55a3ef.Bfloat' -outputdatatype float
sub_data = sub_data_list$image
sub_scheme = sub_data_list$scheme

This process may take a while, and using the RNifti package may be quicker.

nim = RNifti::readNifti(outfiles["data"])
sub_data = tempfile(fileext = ".nii.gz")
sub_scheme_res = camino_subset_max_bval_scheme(
  schemefile = scheme_file, max_bval = 1500,
  verbose = TRUE)
nim = nim[,,, sub_scheme$keep_files]
RNifti::writeNifti(image = nim, file = sub_data)
sub_scheme = sub_scheme_res$scheme
rm(list = "nim");
for (i in 1:10) gc(); 

Fit the diffusion tensor

# wdtfit caminoProc/hcp_b5_b1000.Bfloat caminoProc/hcp_b5_b1000.scheme \
# -brainmask 100307/T1w/Diffusion/nodif_brain_mask.nii.gz -outputfile caminoProc/wdt.Bdouble
mod_file = camino_modelfit(
  infile = sub_data, scheme = sub_scheme, 
  mask = outfiles["nodif_brain_mask"], 
  gradadj = outfiles["grad_dev"],
  model = "ldt_wtd")
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/modelfit -inputfile '/private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b54c55a3ef.Bfloat' -outputfile '/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b548434e0f.Bdouble' -inputdatatype float -schemefile /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b51867c0a3.scheme -bgmask /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b5fd527a4/nodif_brain_mask.nii.gz -maskdatatype float -model ldt_wtd -gradadj /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b5fd527a4/grad_dev.nii.gz

Getting FA vlaues

# fa -inputfile caminoProc/wdt_dt.nii.gz -outputfile caminoProc/wdt_fa.nii.gz
fa_img = camino_fa_img(
  infile = mod_file,
  header = outfiles["nodif_brain_mask"],
  retimg = FALSE)
cat '/private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b548434e0f.Bdouble' |  /Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/fa -inputmodel dt -outputdatatype double > '/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b51173dae.Bdouble'
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/voxel2image -inputfile /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b51173dae.Bdouble -header /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b5fd527a4/nodif_brain_mask.nii.gz -outputroot /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b51cd26b11_ -components 1 -gzip 

Visualizing FA images

We want to read the FA image into R:

fa_nii = readnii(fa_img)

In order to visualize the values, we are going to read in the mask so that we don’t visualize non-brain values:

mask = readnii(outfiles["nodif_brain_mask"])
hist(mask_vals(fa_nii, mask = mask), breaks = 1000)

Using ortho2, we can visualize these FA maps:


Getting MD vlaues

# md -inputfile caminoProc/wdt_dt.nii.gz -outputfile caminoProc/wdt_md.nii.gz
md_img = camino_md_img(
  infile = mod_file,
  header = outfiles["nodif_brain_mask"],
  retimg = FALSE)
cat '/private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b548434e0f.Bdouble' |  /Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/md -inputmodel dt -outputdatatype double > '/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b51845b1d4.Bdouble'
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/rcamino/camino/bin/voxel2image -inputfile /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b51845b1d4.Bdouble -header /private/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T/RtmpuyYwcz/file94b5fd527a4/nodif_brain_mask.nii.gz -outputroot /var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b579de1bf_ -components 1 -gzip 

Visualizing MD images

We want to read the MD image into R:

md_nii = readnii(md_img)
hist(mask_vals(md_nii, mask = mask), breaks = 1000)

md2 = md_nii
md2[ md2 < 0] = 0
hist(mask_vals(md2, mask = mask), breaks = 1000)

Using ortho2, we can visualize these MD maps:



rb_md = robust_window(md2, probs = c(0, 0.9999))

hist(mask_vals(rb_md, mask = mask), breaks = 1000)

Export DTs to NIfTI

Using camino_dt2nii, we can export the diffusion tensors into NIfTI files. We see the result is the filenames of the NIfTI files, and that they all exist (otherwise there’d be an errors.)

# dt2nii -inputfile caminoProc/wdt.Bdouble -header 100307/T1w/Diffusion/nodif_brain_mask.nii.gz \
# -outputroot caminoProc/wdt_
mod_nii = camino_dt2nii(
  infile = mod_file,
  header = outfiles["nodif_brain_mask"])
# dteig -inputfile caminoProc/wdt.Bdouble -outputfile caminoProc/wdt_eig.Bdouble
eigen_image = camino_dteig(infile = mod_file)

We can read these DT images into R again using readnii, but we must set drop_dim = FALSE for diffusion tensor images because the pixel dimensions are zero and readnii assumes you want to drop “empty” dimensions

dt_imgs = lapply(mod_nii, readnii, drop_dim = FALSE)

Downloading T1 image

For image registration to a template, we will use the subject-level

r_t1_mask = download_hcp_file(
    "HCP", hcp_id, "T1w", 
  verbose = FALSE
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/brainmask_fs.nii.gz"
t1_mask = readnii(r_t1_mask)
r_t1 = download_hcp_file(
    "HCP", hcp_id, "T1w", 
  verbose = FALSE
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/T1w_acpc_dc_restore.nii.gz"
t1 = readnii(r_t1)
brain = mask_img(t1, t1_mask)
hist(mask_vals(brain, t1_mask), breaks = 2000)

rob = robust_window(brain, probs = c(0, 0.9999), mask = t1_mask)
hist(mask_vals(rob, t1_mask), breaks = 2000)

Rigid-body Registration of DTI to T1

Here we can register the FA image to the T1-weighted image using a rigid-body transformation. We could have also used the MD image or the diffusion data directly, such as the mean over the tensors.

Warning: multiple methods tables found for 'pixdim'

Warning: multiple methods tables found for 'pixdim'

Attaching package: 'extrantsr'
The following objects are masked from 'package:oro.nifti':

    origin, origin<-
rigid = registration(
  filename = fa_img,
  template.file = rob,
  correct = FALSE,
  verbose = FALSE,
  typeofTransform = "Rigid")
rigid_trans = rigid$fwdtransforms
aff = R.matlab::readMat(rigid$fwdtransforms)
aff = aff$AffineTransform.float.3.3

double_ortho(rob, rigid$outfile)

Alternatively, we can also do a brain mask to brain mask transformation, which we can estimate using a last squares metric. This should be sufficient for what we need and more robust to artifacts in the T1 or the FA map, so we’ll use this transformation.

rigid_mask = registration(
  filename = outfiles["nodif_brain_mask"],
  template.file = r_t1_mask,
  correct = FALSE,
  typeofTransform = "Rigid",
  affMetric = "meansquares")
# Running Registration of file to template
# Applying Registration output is
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b54f457f480GenericAffine.mat"

[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b54f457f480GenericAffine.mat"
# Applying Transformations to file
# Writing out file
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b5242842fb.nii.gz"
# Reading data back into R
rigid_mask_trans = rigid_mask$fwdtransforms

aff_mask = R.matlab::readMat(rigid_mask$fwdtransforms)
aff_mask = aff_mask$AffineTransform.float.3.3

double_ortho(t1_mask, rigid_mask$outfile, NA.x = FALSE)

Non-linear Registration of T1 to template

Here we will use symmetric normalization (SyN) to register the Winsorized skull-stripped brain image of the HCP subject to the Eve template.

eve_brain_fname = EveTemplate::getEvePath(what = "Brain")
eve_brain = readnii(eve_brain_fname)
nonlin = registration(
  filename = rob,
  template.file = eve_brain_fname,
  correct = FALSE,
  typeofTransform = "SyN")
# Running Registration of file to template
# Applying Registration output is
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b535a1df4b1Warp.nii.gz"      
[2] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b535a1df4b0GenericAffine.mat"

[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b535a1df4b0GenericAffine.mat" 
[2] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b535a1df4b1InverseWarp.nii.gz"
# Applying Transformations to file
# Writing out file
[1] "/var/folders/1s/wrtqcpxn685_zk570bnx9_rr0000gr/T//RtmpuyYwcz/file94b51bcb23e6.nii.gz"
# Reading data back into R
double_ortho(eve_brain, nonlin$outfile)

nonlin_trans = nonlin$fwdtransforms

Registering DTI to Eve

Now, we can use the transformed images from the rigid-body transformations above and apply this non-linear transformation to those FA and MD registered images. The one problem is that the rigid-body registration interpolates the data and the non-linear registration interpolates the data.

We can compose the transforms so that the data is only interpolated once.

In ants_apply_transforms, which calls antsApplyTransforms, the transform list must be specified in reverse order to which they are done. We want to perform the rigid body transformation then the non-linear registration, but need the composed list of transforms to first have the non-linear transformation then the rigid-body transformation.

Here we apply this composed transformation to the FA and MD values.

composed = c(nonlin_trans, rigid_mask_trans)
fa_eve = ants_apply_transforms(
  fixed = eve_brain_fname,
  moving = fa_img,
  transformlist = composed)
double_ortho(eve_brain, fa_eve)

md_eve = ants_apply_transforms(
  fixed = eve_brain_fname,
  moving = rb_md,
  transformlist = composed)
double_ortho(eve_brain, md_eve)