Skip to content

Latest commit

 

History

History
114 lines (108 loc) · 4.6 KB

Preprocessing.md

File metadata and controls

114 lines (108 loc) · 4.6 KB
Raw

Preprocessing

Date: 1/14/15

Stream

Basic pipeline (check data between steps)

  • Reconstructed image data

  • Sometimes do fieldmaps (be careful if it helps or hurts)

  • Motion correction

  • Slice timing correction (optional)

  • Spatial smoothing (optional; on volume or surface)

  • Stats

  • not including normalization as preproc step

Quality control

  • Use eyes to look at data!
  • Possible problems:
    • Missing data from half the brain?
    • Spikes in fMRI data
      • signal at very specific frequency
      • problem in k-space
    • Ghosting
      • offset in phase between different lines in k-space in EPI acq
      • could be worse in MUX
  • How to:
    • Create movie in fslview
    • Model free analysis
      • ICA, locate artifacts in the data
        • e.g., melodic in FSL
        • find independent non-Gaussian things
          • physio artifact (periodic)
          • movement (component response spike)
            • good for non-rigid body movement, esp that where movement happens between 2 sequential TRs
        • Remove from data
          • Using melodic just pull out the components
          • Add component time series to GLM
            • not ideal if a lot of components
            • Check version of FSL, might do unnecessary matrix inversion and use a ton of RAM

Distortion Correction

  • Air+tissue boundary causes inhomogeneity
    • e.g., ears, sinuses
    • miss parts of brain, and stretches some areas
      • due to phase encoding, can combine a bunch of voxels in real space into one in image space
  • When?
    • at beginning/end, or interleaved between runs
    • the more you average together the better off you are
  • How?
    • Collect field maps, perhaps multiple
      • register field maps
    • map out inhomogeneity in magnetic field, and then fix it
    • map of magnitude and relative phase between 2 echoes
    • unwrapping

Slice Timing Correction

  • Slices are collected in sequence, top to bottom, interleaved, etc.
    • interleaved
      • can't make perfect square, so collecting slice B also includes some info from A and C.
  • worst for ER designs, not block, since each second matters more
  • Resample dataset back to one point in time
  • Caveats (esp with shorter TRs (<2s) & interleaved acq)
    • If have spike, could smear across data with sinc interpolation
    • Should tweak techniques for MUX!
  • Could add temporal derivatives to the model
    • help with slice timing problems to soak up that variance

Motion correction/realignment

  • Retrospective (after the fact)
  • Prospective: Siemens can do this during the scan
    • with cameras to track head motion, and then update the slices
  • Doesn't do a good job with:
    • physio noise
      • to get rid of that, can do cardiac gating, e.g., if imaging brain stem
      • record heart beat and respiration and use those to remove that artifact
      • can get aliased signal since stuff like respiration is relatively fast, can get low frequency in images
  • Bulk motion
    • head movement, where whole head moves
    • artifact near edges of brain, and near ventricles
  • Spin history effects
    • spin history of voxel will be in new area after movement
    • might be able to correct with SPM
  • Might be hard to break apart signal from motion
    • But timing matters! motion should happen immediately, and then brain activation later
  • How to fix: Realignment
    • Rigid-body, 6 param registration
      • x,y,z translation, rotation
    • How to visualize:
      • Try plotting the derivatives, so not looking at slow drifts that are plotting in absolute diff
      • Or plot framewise displacement, to collapse across 6 params
      • Also plot intensity changes over time
    • Use middle image from each run as the target image
      • first image could be messed up
      • could use mean image, but is more blurry, and extra calculation
    • Cost function:
      • FSL: normalized correlation error
      • caveat! if signal is strong, might look like motion
        • could use mutual information instead
    • Interpolation
      • linear: more spatial smoothing, but fast
      • Higher order methods:
        • sinc, spline, fourier-based
  • Other
    • relationships between motion and susceptibility artifacts
      • rigid-body motion correction won't fix
    • motion or slice-timing correction first?
      • if slice first, could move voxels into different slices
      • if motion first, could be errors in motion intensity spread out over time
      • nipype tool 4drealign does both at the same time! but takes a long time. should work in multiband
    • When to throw out data
      • maybe not pick a given threshold, unless too many corrupted timepoints
      • 1/2 voxel