Skip to content

Latest commit

 

History

History
133 lines (119 loc) · 4.77 KB

Normalization.md

File metadata and controls

133 lines (119 loc) · 4.77 KB
Raw

Normalization

Date: 2/4/15

Initial notes

  • Spatial normalization: putting people into some type of standard space

Talairach space

  • 3d Cartesian coordinate space
  • Based on anatomical landmarks
    • AC, PC, midline sag, bounding box, coronal plane to create multiple boxes
    • shove brain into box aligned across brains
    • modify each box independently, piecewise linear transform
  • Talairach atlas is based on 1 person's brain, so not great to fit many people

Atlas vs. Template

Atlas

  • Show anatomical locations in a coordinate space
  • Examples
    • Harvard Oxford, AAL Atlas, etc.
  • Talairach atlas
    • postmortem drawings?

Template

  • Example: MNI152 (152 people registered together)
    • Can align images to this target template
    • Supplies a coordinate system
  • MNI305
    • 9 dof affine registration
    • align the 305 images to Talairach atlas
  • ICBM 152/MNI152
    • Brains from around the world
    • Register high res images to MNI 305 template
    • different from MNI305!
    • nonlinear registration?

Preprocessing anatomical images

Bias field correction

  • gradient in brightness across image
  • HP filter and/or bias field correction
  • Can impact anatomical images & influence registration

Brain extraction

  • Freesurfer (make_mask, or fs.ApplyMask for nipype) works better than BET (FSL)
  • BET (FSL)
    • starts with a ball in center of the brain, iteratively expands

Tissue segmentation

  • Separate gray, white matter & CSF
  • Could save as regressors for CSF, etc.
  • Make sure do bias field correction
  • Can't threshold images to separate
    • distributions of image intensities overlap
    • voxels might be on the boundaries
  • Technique:
    • Unified segmentation (Ashburner & Friston, 2005)
    • Put brain into MNI space, and have priors on each voxel for whether gray, white or CSF

Normalization

When?

  • prior to stats (SPM) -- normalize data
  • post stats (FSL) -- normalize stats
  • Pros/Cons:
    • File sizes are smaller if normalize post-stats
    • Can do different types of normalization once you have your stats quickly

Which images are used?

  • Can register EPI to EPI template, but not great since not much contrast in EPI
  • Or, 2-step procedure, register functional to anat, and anat to T1-weighted, and then concatenate both
    • this limits the accumulation of interpolation error
  • 3 step procedure
    • collect coplanar to cover same slice as func, and register to that, then to anat, etc.
    • bbregister is better

How do you realign?

  • Landmark based
    • need anatomy expert
  • Volume-based registration
    • MI (SPM) or normalized correlation (FSL)
  • Diffeomorphic transformation
    • Treat brain like viscous fluid registration
    • Penalizes major warps
  • Warpfield (nonlinear registration), ANTS
    • Take brain on square grid, show how voxels get pushed around, done in 3D space or on the surface in 2D space
  • Surface based methods
    • Rely on sulci and gyri
    • Automated, but check for handles and donut holes!
    • Register to surface atlas
    • Map fMRI data to surface space
    • Appears to be more accurate
    • Fine, but only good for cortical surface
    • CIFTI files: stores surface and subcortical gray matter (grayordinates)
      • could do analysis on all of the data
  • Which to use?
    • Maybe nonlinear is better, some examples where FNIRT shows more activation than FLIRT
    • Klein 2009 says nonlinear
    • ANTS might be the best, but annoying to install/use
    • Try taking 2 T1s, and average together to try to increase Freesurfer's reliability
      • might be useful for cortical thickness
      • but motion might trump that?

QA

  • T1-normalization
    • View individual subjects normalization on top of template
  • Average normalized brains (e.g., all T1s)
    • Should look like a blurry version of the breal brain
    • Good for large datasets where time/brain is limited
  • Is orientation correct?
  • Double check that brain extraction/skull stripping got everything

Different ages

  • Children
    • Pediatric templates
      • only good if have a single age group
      • common methods are robust to age differences, at least for 7+
  • Elderly
    • Decrease in gray matter, increase in CSF
    • Create custom templates
  • Lesions
    • Account for lesions in cost function
      • leave the lesion spots out of warp
      • might impact surrounding structure too

Structure-specific alignment

  • pick that one structure and align to that (Craig Stark?)
  • Add weight during the alignment to parts that you've traced
    • 20-50% weighting on your tracing, and let ANTS figure out the rest of the brain
    • Maybe try inverse warp on template, and see how well it matches up