USAGE:
UnbiasedNonLocalMeans [--returnparameterfile ]
[--processinformationaddress ]
[--xml] [--echo] [--ps ] [--hp ]
[--rc >] [--rs
>] [--sigma ] [--]
[--version] [-h]
Where:
--returnparameterfile
Filename in which to write simple return parameters (int, float,
int-vector, etc.) as opposed to bulk return parameters (image,
geometry, transform, measurement, table).
--processinformationaddress
Address of a structure to store process information (progress, abort,
etc.). (default: 0)
--xml
Produce xml description of command line arguments (default: 0)
--echo
Echo the command line arguments (default: 0)
--ps
To accelerate computations, preselection is used: if the normalized
difference is above this threshold, the voxel will be discarded (non
used for average) (default: 2)
--hp
This parameter is related to noise; the larger the parameter, the more
aggressive the filtering. Should be near 1, and only values between
0.8 and 1.2 are allowed (default: 1)
--rc >
Similarity between blocks is computed as the difference between mean
values and gradients. These parameters are computed fitting a
hyperplane with LS inside a neighborhood of this size (default: 1,1
,1)
--rs >
The algorithm search for similar voxels in a neighborhood of this
radius (radii larger than 5,5,5 are very slow, and the results can be
only marginally better. Small radii may fail to effectively remove the
noise). (default: 3,3,3)
--sigma
The root power of noise (sigma) in the complex Gaussian process the
Rician comes from. If it is underestimated, the algorithm fails to
remove the noise. If it is overestimated, over-blurring is likely to
occur. (default: 5)
--, --ignore_rest
Ignores the rest of the labeled arguments following this flag.
--version
Displays version information and exits.
-h, --help
Displays usage information and exits.
(required) Input MRI volume.
(required) Output (filtered) MRI volume.
Description: This module implements a fast version of the popular
Non-Local Means filter for image denoising. This algorithm filters each
pixel as a weighted average of its neighbors in a large vicinity. The
weights are computed based on the similarity of each neighbor with the
voxel to be denoised.
In the original formulation a patch with a certain radius is centered in
each of the voxels, and the Mean Squared Error between each pair of
corresponding voxels is computed. In this implementation, only the mean
value and gradient components are compared. This, together with an
efficient memory management, can attain a speed-up of nearly 20x.
Besides, the filtering is more accurate than the original with poor
SNR.
This code is intended for its use with MRI (or any other
Rician-distributed modality): the second order moment is estimated, then
we subtract twice the squared power of noise, and finally we take the
square root of the result to remove the Rician bias.
The original implementation of the NLM filter may be found in:
A. Buades, B. Coll, J. Morel, 'A review of image denoising algorithms,
with a new one', Multiscale Modelling and Simulation 4(2): 490-530.
2005.
The correction of the Rician bias is described in the following
reference (among others):
S. Aja-Fernandez, K. Krissian, 'An unbiased Non-Local Means scheme for
DWI filtering', in: Proceedings of the MICCAI Workshop on Computational
Diffusion MRI, 2008, pp. 277-284.
The whole description of this version may be found in the following
paper (please, cite it if you are willing to use this software):
A. Tristan-Vega, V. Garcia Perez, S. Aja-Fenandez, and C.-F. Westin,
'Efficient and Robust Nonlocal Means Denoising of MR Data Based on
Salient Features Matching', Computer Methods and Programs in
Biomedicine. (Accepted for publication) 2011.
Author(s): Antonio Tristan Vega, Veronica Garcia-Perez, Santiago
Aja-Fernandez, Carl-Fredrik Westin
Acknowledgements: Supported by grant number FMECD-2010/71131616E from
the Spanish Ministry of Education/Fulbright Committee