Overall Pipeline

Intensity normalization

  • Conventional MRI intensites (T1-w, T2-w, PD, FLAIR) are acquired in arbitrary units
  • Images are not comparable across scanners, subjects, and visits, even when the same protocol is used.
    • This affects algorithm performance, prediction, inference.
    • Even simple things like thresholding an image
  • Intensity normalization brings the intensities to a common scale across people.
  • In this tutorial we will normalize intensities within subject using two methods:
    • Whole-brain normalization
    • White Stripe normalization (Shinohara et al. 2014).

Visualizing whole-brain intensities (each line is a person)

  • We will work with the T1-w images from the training data.
  • Full brain densities are mixtures of the three tissue class distributions.

Visualizing the intensities by tissue class

And these are all the same scanner/protocol!

Whole-brain normalization

  • Let’s Z-score each voxel using mean \(\mu_{WB}\) and standard deviation \(\sigma_{WB}\) computed from all voxels in the brain mask.

\[ T1_{WB} = \frac{T1 - \mu_{WB}}{\sigma_{WB}}\]

  • zscore_img is a function in neurobase that does this.
  • It takes an image and a binary mask. The default is to use all voxels in the brain mask.
zscore_img(img = img, mask = mask)

Whole-brain normalized intensities

Whole-brain normalized intensities

  • Gray matter distributions are more comparable.

Whole-brain normalized intensities

  • White matter distributions are more comparable.

Other Normalizations: White Stripe

  • Whole-brain normalization may be sensitive to outliers.
  • Lesions in MS can have very high intensities, which lead to bad estimates of mean/variance
    • Other more robust transformations may be used, such as using the median to center, IQR to scale, etc.
  • White Stripe (Shinohara et al. 2014) is based on parameters obtained from a sample of normal appearing white matter (NAWM), which is robust to outliers.
    • The idea is to make normal appearing white matter comparable across subjects and visits.

White Stripe normalization

Procedure:

  1. Find white matter area on histogram

White Stripe normalization

Procedure:

  1. Find white matter area on histogram

  2. Estimate mean \(\mu_{WS}\) and variance \(\sigma_{WS}\) of voxel intensities in that area

  3. Normalize with those means/variances: \[ T1_{WS} = \frac{T1 - \mu_{WS}}{\sigma_{WS}}\]

White Stripe normalization

  • After normalization, NAWM will have a standard normal distribution and units will be in standard deviations of NAWM.
  • Gray matter and CSF distributions may not be comparable after White Stripe.

White Stripe normalization

Procedure:

  1. Find white matter area on histogram

  2. Estimate mean \(\mu_{WS}\) and variance \(\sigma_{WS}\) of voxel intensities in that area

  3. Normalize with those means/variances: \[ T1_{WS} = \frac{T1 - \mu_{WS}}{\sigma_{WS}}\]

  • After normalization, NAWM will have a standard normal distribution and units will be in standard deviations of NAWM.
  • Gray matter and CSF distributions may not be comparable after White Stripe.

White Stripe normalization code

library(WhiteStripe)
ind = whitestripe(img = t1, type = "T1", stripped = TRUE)$whitestripe.ind
ws_t1 = whitestripe_norm(t1, indices = ind)
  • The whitestripe function takes an image, image type (in our case T1), and a logical indicating whether the image has been skull stripped.
  • The indicies of voxels in the NAWM used for estimating the normalization parameters are located in the list element $whitestripe.ind.
  • The function whitestripe_norm takes an image and the indicies from a call to whitestripe and returns the White Stripe normalized image as a nifti.

WhiteStripe normalized intensities

WhiteStripe normalized intensities

WhiteStripe normalized intensities

Conclusions

  • Intensity normalization is an important step in any image analysis with more than one subject or time point to ensure comparability across images.
  • White Stripe normalization may work better and have better interpretation than whole-brain normalization for subsequent lesion segmentation algorithms and analysis.
  • Other intensity normalization methods that make intensites comparable across subjects for all tissues exist.
    • RAVEL, which is an extension of WhiteStripe is one example (Fortin et al. 2016).
    • Located at https://github.com/Jfortin1/RAVEL
      • This was shown to have better comparability than histogram matching

Website

References

Fortin, Jean-Philippe, Elizabeth M Sweeney, John Muschelli, Ciprian M Crainiceanu, Russell T Shinohara, Alzheimer’s Disease Neuroimaging Initiative, and others. 2016. “Removing Inter-Subject Technical Variability in Magnetic Resonance Imaging Studies.” NeuroImage 132. Elsevier:198–212.

Shinohara, Russell T, Elizabeth M Sweeney, Jeff Goldsmith, Navid Shiee, Farrah J Mateen, Peter A Calabresi, Samson Jarso, et al. 2014. “Statistical Normalization Techniques for Magnetic Resonance Imaging.” NeuroImage: Clinical 6. Elsevier:9–19.