US 6,204,853 B1 · 4D KAPPA5 Gaussian noise reduction

Abstract

The noise reduction system of the present invention takes into account that noise is random with little between images adjacent physically, and in sampling time. It also takes into account the fact that some sequences of images are cyclical and “wrap around” in time with the beginning closely resembling the end of the cycle. A filter was developed which would smooth noise in a direction along an edge, but will not blur across an edge. It operates by determining vectors tangential to a surface point p, at a current voxel, and projecting 4D data onto the tangential vectors. A curvature matrix Bαβ is determined. The eigenvalues of curvature matrix Bαβ are determined to result in three curvatures for 4 dimensions. If the sign of all of the eigenvalues is the same, the current voxel is filtered, else, it is unchanged. This filtering is repeated for a number of voxels as the current voxel within a desired region for a single iteration. Preferably, this is repeated for several iterations to result in filtered data with reduced noise and little change in detail. The filtered data may then be segmented to separate structures. The segmented structures may then be used in analysis, such as in medical diagnosis.

Claims7

Dependent claims

A method of reducing noise in images while preserving edge information comprising the steps of:

a) acquiring a multidimensional set of images S each which vary at least over D dimensions, comprised of a set of voxel values;
b) selecting a current voxel at a point p to be filtered;
c) determining a plurality of first differences Si=dS/di at the current voxel, where i represents each of D dimensions;
d) determining a plurality of second differences Sij, from the first differences, where indices i, j represent each of D dimensions;
e) determining D−1 tangent vectors e1, e2, . . . eD-1 at point p;
f) projecting second differences Sij onto tangent vectors e1, e2, . . . eD-1 to result in a curvature matrix Bαβ;
g) diagonalizing curvature matrix Bαβ to determine a diagonal matrix with diagonal elements being curvatures K1, K2, . . . KD-1;
h) testing the signs of the curvatures K1, K2, . . . KD-1, and if the curvature signs are the same:
i. calculating a cofactor Δij from Sij;
ii. modifying the current voxel value Sw at point p according to:
S w + 1 = S w + ∑ i = 1 D ⁢ ∑ j = 1 D ⁢ ( S i ⁢ S j ⁢ Δ ij ) 1 D + 1
to result in the new filtered value Sw+1;
and
i) repeating steps “c” through “h” for a plurality of different voxels as the current voxel in a desired region to perform a filtering iteration resulting in filtered data Sw+1 having reduced noise with little change in detail.

11 more elements

Dependencies are inferred from the claim text.

4D KAPPA5 Gaussian noise reduction

Life of this patent

4D KAPPA5 Gaussian noise reduction

Life of this patent

Abstract

Claims7