COMPARISON OF NON-UNIFORMITY-CORRECTION TECHNIQUES
USING BIASED BRAIN PHANTOM DATA
We evaluated the performance of six non-uniformity-correction algorithms
(n3, bfc, eq, hum, cma, and spm)
on a set of six brain images created by adding known biases to the Montreal
Brain Phantom (MBP) with no added noise (
http://www.bic.mni.mcgill.ca/brainweb/selection_normal.html). Bias
volumes with differing spatial distributions were created as the product
of (i) three orthogonal parabolic functions and (ii) three orthogonal sinusoidal
functions. Each bias volume was computed at three different magnitudes
of relative bias (±2%, ±4% and ±8%) centered about
a value of 1.0; the periods of the sinusoids varied from 0.8 to 1.2 times
the orthogonal dimensions of the MBP. [For example, for ±2% the
bias ranged from 0.98 to 1.02.] The six resulting bias volumes were applied
multiplicatively to the MBP followed by the addition of Gaussian random
noise (mean = 0, SD = 4.0) to produce the test set of biased phantom brain
volumes. After the six non-uniformity-correction algorithms had been run
on the biased phantom test set, the corrected volumes were normalized to
have the same mean as the original, unbiased volume. To account for differences
between the correction algorithms with regard to their representation of
the estimated bias and, thereby, to allow for meaningful comparisons, all
calculated bias volumes were scaled to the same data range and offset.
- Multiple Repeat T1-weighted
MRI Scans of a Single Subject (Figs. 5-10)
- High-Resolution T1-weighted
MRI Scans from Two Different Centers (Figs. 11-12)
- Return to Introduction
Page
Parabolic Function (±8% Bias)
(Figs. 1-2)
Sinusoidal Function (±8% Bias)
FIGURE 3
Comparison of six non-uniformity-correction algorithms using the Montreal
Brain Phantom with ±8% sinusoidal bias. Columns (left to right)
represent measurements derived from consecutive axial, coronal and sagittal
slices. Top Row. Average voxel values for the applied (black line)
and extracted (colored lines) bias volumes. Second Row. RMS error
between bias-corrected and unbiased volumes. The black line represents
the RMS "error" between the biased and unbiased volumes. Gaussian random
noise was added after biasing the phantom volume, so that all measures
include an RMS error of ~4.0 based on noise alone. Third Row. RMS
error between the extracted and applied bias volumes. Fourth Row.
Voxel-wise correlation coefficient for the extracted vs. applied bias volumes.
FIGURE 4
Voxel-wise scatter plots of the extracted vs. applied (±8% sinusoidal)
bias volumes for six bias-correction algorithms.
