Retrospective Evaluation and Correction of Intensity Inhomogeneity - Advance Image Processing in Magnetic Resonance Imaging - page 143

important correction methods. In the next sections, we first discuss early
attempts of retrospective correction, and then we present the two main classes
of methods routinely used nowadays in many image processing studies.

5.2 EARLY SOLUTIONS

Starting in the mid-1980s, MRI researchers began to develop methods for the
correction of intensity inhomogeneity. The need for correction was most imminent
in images acquired using surface coil images with strong inhomogeneity caused
by the falloff of coil sensitivity with distance from the coil center. In 1986,

Haselgrove and Prammer [11] (see also Merickel et al. [ 3 ]) suggested the use of

smoothing in order to reduce the inhomogeneity. They proposed that each MR
slice is divided by its spatially smoothed copy. This reduces the low-frequency
inhomogeneity effects of the surface coil. In order to correct for step-like interslice
intensity inhomogeneities, they linearly normalized each slice to have the same
average intensity values for a selected set of manually identified tissue classes
visible in all slices. Smoothing was also proposed by Lim and Pfefferbaum [4]
in 1989 to correct for inhomogeneities in brain MRI scans. After the manual
extraction of the head, the intensity values were extended radially toward the
image boundaries and smoothed with a Gaussian filter of large kernel size. They
assumed that the resulting blurred image represents one homogeneous region that
is only distorted by the scanner inhomogeneities. The images were corrected with
this approximation of the inhomogeneity characteristics.
Many more researchers proposed different methods using smoothing filters
or homomorphic unsharp masking (e.g., [12]). These early filtering methods
undesirably reduce the contrast between tissues, and they often generate new
artifacts in the corrected images. Homomorphic filtering also falls into the same
class of correction methods [13,14]. Like smoothing methods, homomorphic
filtering assumes a separation of the low-frequency inhomogeneity field from the
higher frequencies of the image structures. The assumption is often valid in
microscope images of small particles but often fails for the structured MR images.
A scene, such as a head structure, contains a considerable amount of low-
frequency components. Homomorphic filtering assumes further that the local
intensity statistics are constant across the whole image. This assumption is not
true for most of the inhomogeneity effects observed in MR.
In 1988, Vannier et al. [2] were one of the first to model the intensity inhomo-
geneity as a parametric inhomogeneity field. They used a simple linear ramp in the
transverse anatomic plane as the parametric field model. This model was fitted to
the linewise average intensity values of the image. Later the same year, they
proposed an improved model that fitted a fourth-order polynomial to the line-by-
line histogram [15]. In both approaches, only vertical distortions were taken into
account. The fitted inhomogeneity field was then subtracted from the original image.
Several researchers proposed the estimation of the correction from a prior
segmentation. Dawant et al. [8] proposed that users select typical samples of each