A nonparametric method for automatic correction of intensity nonuniformity in MRI data

TLDR: A novel approach to correcting for intensity nonuniformity in magnetic resonance (MR) data is described that achieves high performance without requiring a model of the tissue classes present, and is applied at an early stage in an automated data analysis, before a tissue model is available.

Abstract: A novel approach to correcting for intensity nonuniformity in magnetic resonance (MR) data is described that achieves high performance without requiring a model of the tissue classes present. The method has the advantage that it can be applied at an early stage in an automated data analysis, before a tissue model is available. Described as nonparametric nonuniform intensity normalization (N3), the method is independent of pulse sequence and insensitive to pathological data that might otherwise violate model assumptions. To eliminate the dependence of the field estimate on anatomy, an iterative approach is employed to estimate both the multiplicative bias field and the distribution of the true tissue intensities. The performance of this method is evaluated using both real and simulated MR data.

Figures

Figure 2.3: Simulated spin echo images of a cylinder with elliptic section for various values of Er and p. Compare with Figure 2.1 where Er = 80 and p = 2 nnz. i.e. normal values.

Figure 4.3: Field estimates for a sirnulated .\IR volume. (a) :\. slice from a simulated :\tIR volume. (b) The histogranl of the volunle shawn in (a). This is considered an estimate of the distribution ~'. (c) Bias field estimates Je created by applying the mapping in (d) ta the image in (a). (cl) The mapping between image intensity vand bias field estimate fe~ derived frOIn the histogram in (b). (e) A smoothed estimate fs of the bias field created by SIIloothing the volume shawn in (c). (f) The actual bias field present in the volunle shawn in (a).

Figure 5.2: Error measure versus iterations for four different widths of the deconvolution kernel F. The F\YH.\[s of the field distributions F used in deconvolution were 0.1, 0.2, 0.3, and 0.-1. The rnaximum coefficient of variation 4.5% corresponds to no correction. The horizontal line is the result of smoothing the uncorrected data directly using B splines.

Figure 2.2: Tl weighted brain scans showing diagonal non-uniformity~brighter at the top left and bottom right. The gray scale has been compressed to exaggerate the intensity variations.

Figure 5.1: (a) Intensity distribution U for an idealized volume. Random samples from this distribution \vere llsed to create the volume in (cl). (b) Distribution F for a parabolic non-uniformity field. (c) Distribution F corresponding to the volume shown in (f). (d) A slice through a random volume. (e and g) A parabolic non-uniformity field. (f) The \"OIUllle in (d) multiplied by the parabolic field in (e).

Figure 6.1: .-\. comparison using simulated data of the four non-uniformity correction methods. (a) through (c) are for increasing noise levcl and (d) through (e) are for increasing levels of intensity non-llniformity. The trace '~E~'I - ideal" is the performance of the Ej\I algorithm \Vith an ideal training set (see text). The crror bars are at ±l standard deviation based on four trials.