Segmentation with Corrected
MPRAGE Scans and FSL
Jason Su
Homogeneity Correction
• Reiss’s CIBSR group noticed a significant bias
non-uniformity in the MPRAGE images,
described as a “bullseye artifact”
• They suggested a few tools correction:
homocor3d, N3, and SPM8
• I was not able to obtain a copy of N3, their
homepage is currently being transferred
Correction Comparison - Original
Correction Comparison – homocor3d
Correction Comparison – SPM8
Correction Comparison
• Visually, SPM8 seems to eliminate the bullseye
artifact better than homocor3d
• SPM8 does over brighten areas outside of the
brain, in the neck, which has made brain
extraction not as clean
• Let’s see how this affects the segmentation,
which is what we’re ultimately interested in
FAST Estimated Bias Fields
• Produces the most uniform field.
• This is clearly the better method as was
also qualitatively discerned.
N010 – Original – WM
N010 – homocor3d – WM
N010 – SPM8 – WM
• The corrected versions show an improved
ability to differentiate fine WM structures
such as in the cerebellum
• The SPM8 segmentation has slightly less
• Tissue near the ventricles still classified as GM
• Speckling from noise in the scan is a hurdle
– Potential fixes: anisotropic diffusion filter, median
Median Filtered SPM8 WM
Notes and Future Work
• I did segmentations for our 6 sample subjects
but only showed the normal since that should
be our “best case” with no lesions
• Try correction with N3
• Multispectral segmentation (Fumiko is doing
something along these lines as well)
• Segmentation with SPM8
• Segmentation with Freesurfer