Retrospective Evaluation and Correction of Intensity Inhomogeneity - Advance Image Processing in Magnetic Resonance Imaging - page 159

(N3).* This signal processing motivated approach takes advantage of the usually
simple form of the tissue-class distributions. It searches for a smooth inhomoge-
neity field that maximizes the frequency content of the image intensity distribu-
tion. In order to achieve this, the method iteratively deconvolves narrow Gaussian
distributions from the image intensity distribution.
The properties of both information minimization and N3 are quite similar,
whereas their implementations vary quite a bit. Both methods have a minimal set
of parameters and are sometimes also called “nonparametric.” The remaining
parameters mainly control the smoothness of the parametric inhomogeneity field
and the histogram computation. The most relevant parameter of histogram-based
methods is the size of the histogram bins (in the discrete case for information
minimization), or the variance of the kernel (in the continuous case for N3).
Choosing this parameter too low will result in a flat inhomogeneity field estima-
tion, whereas choosing it too high will basically low-pass-filter the image to
produce an inhomogeneity field estimation.
Both information minimization methods and N3 perform well on standard
MR images, and have proven to be very useful in small and large MR studies.
However, N3 is reported to be less suited than information minimization for
images with large-scale structures [53,55].

5.5 DISCUSSION AND CONCLUSION

In this chapter, we reviewed the large body of literature concerning the retrospective
evaluation and correction of intensity inhomogeneities in MRI. The proposed
approaches range from early solutions, through combined segmentation and inho-
mogeneity correction methods, to histogram-based techniques. Whereas the early
methods were instrumental in advancing the state of the art in the field, they are now
largely abandoned because they typically need a high degree of user interaction (e.g.,
for masking or tissue labeling), and are often based on inadequate assumptions (e.g.,
additive inhomogeneity models or single homogeneous region assumptions after large
kernel size smoothing). In contrast, both histogram-based techniques and combined
segmentation and inhomogeneity correction methods are currently being used as
part of everyday image processing pipelines in institutions around the world.
Whether the histogram-based approach is to be preferred over the combined
segmentation and inhomogeneity correction approach or vice versa remains the
subject of continuous debate. On the one hand, approaches that explicitly segment
images while estimating the inhomogeneity field have the distinct advantage that
inhomogeneities are estimated using extensive domain information, rather than using
voxel intensities alone. Furthermore, the goal of inhomogeneity correction in image
processing pipelines is typically to obtain accurate image segmentations, anyway,
in which case solving the segmentation and the inhomogeneity estimation problem
simultaneously makes perfect sense. However, combined segmentation and inhomo-

geneity correction approaches rely heavily on the availability of accurate image

* N3 is freely available at http://www.bic.mni.mcgill.ca/software/distribution.

*