Brain Extraction/Segmentation

Let’s reset and read in the T1 image from a MS lesion data set:

library(neurobase)
t1_fname = "training01_01_t1.nii.gz"
t1 = neurobase::readnii(t1_fname)
rt1 = robust_window(t1, probs = c(0, 0.975)); 
red0.5 = scales::alpha("red", 0.5) # for plotting later

ortho2(rt1)

Here we will use FSL’s Brain Extraction Tool (BET) to extract the brain tissue from the rest of the image (general overview):

• 2nd and 98th percentiles are calculated. (98th - 2nd) * 10% + 2nd percentile used to threshold out background
• From non-thresholded voxels - calculate center of gravity (COG)
• Calculate radius of brain and median intensity of all points within “spherical brain” (used in last step)
• Perform region growing and iterating to get brain surface
• Smooth surface
• Use median intensity to shrink surface to the “real” surface

fslr - wraps FSL commands to use in R - registration, image manipulation

fslr::fslbet - takes in a filename/nifti and calls FSL bet function - additional options can be passed to FSL command in using opts

library(fslr)
ss = fslbet(infile = t1_fname)

ortho2(robust_window(ss))

ortho2(rt1, ss > 0, col.y = red0.5)

Before doing skull-stripping/brain extraction, we would do bias correction:

library(extrantsr)
bc_img = bias_correct(file = t1, correction = "N4")

bc_bet = fslbet(bc_img); ortho2(bc_img, bc_bet > 0, col.y = red0.5)

• Function requires arguments: template.images (T1-weighted images in this case) and template.structs (labels/structures/masks, brain masks here)
• We will use the extrantsr::malf function
  • Performs non-linear registration using Symmetric Normalization (SyN) (Avants et al. 2008), a form of diffeomorphic registration:
  • combines the template structures

library(malf.templates) # load the data package
library(extrantsr)
timgs = mass_images(n_templates = 5) # let's register 5 templates
ss = extrantsr::malf(
  infile = bc_img, 
  template.images = timgs$images, 
  template.structs = timgs$masks,
  keep_images = FALSE # don't keep the registered images
)
mask = ss > 0

mask = readnii("training01_01_t1_mask.nii.gz") # already computed

ortho2(bc_img, mask, col.y = red0.5)

• FSL BET can perform brain extraction (additional ex shows when it works)
  • It did not work sufficiently here
  • There are options you can change for performance
• Bias-correction before brain extraction is a good idea
  • Especially if the method depends on intensities
  • Didn’t change the results here
• MALF/MASS is a good option, but needs templates and is computationally expensive
  • weighted templates or local weighting is done in other software (not discussed)

http://johnmuschelli.com/imaging_in_r

FSL BET did not work in the previous data set. It works in many cases, though. We will look at a subject from the kirby21 dataset (Landman et al. 2011).

library(kirby21.t1)
t1_fname = get_t1_filenames()[1]
t1 = readnii(t1_fname)

ortho2(robust_window(t1))

ss = fslbet(infile = t1_fname); ortho2(robust_window(ss))

bc_img = bias_correct(t1, correction = "N4"); 
bc_bet = fslbet(bc_img)
ortho2(robust_window(t1), bc_bet > 0, col.y = red0.5)

We use the modification of BET in extrantsr, which is called through fslbet_robust. fslbet_robust:

• bias correct image
• remove neck (double_remove_neck performs 2 registration steps, more robust than one (which is the default).)
• run BET
• estimate center of gravity (COG)
• run BET again with new COG

ss = extrantsr::fslbet_robust(
  t1, 
  remover = "double_remove_neck",
  correct = TRUE,
  correction = "N4",
  recog = TRUE)

ortho2(ss)

• Brain extraction allows you to analyze the brain only
  • Important for tissue segmentation/registration
• BET may work (look at your data!)
  • Should bias correct first
  • May need to remove neck
  • High values may affect results - may need to remove/Winsorize them

http://johnmuschelli.com/imaging_in_r