In [24], a novel scale-based filtering method is presented which utilizes scale-dependent

diffusion conductance. Unlike reported scale-based filtering methods, the proposed

method employs a concept called generalized scale, which imposes no shape, size, or
anisotropic constraints to obtain object scale information. The object scale information is
used to adaptively perform smoothing in the interior of homogeneous regions more than
smoothing in regions with fine details. In [25], the Optimized Block wise NL-means filter
is proposed to overcome computation time problem of NLmeans in 3D images. This
approach divides the image volume into blocks with overlapping supports and applies
NL-means on these blocks. This approach restores the voxels values based on the
restored values of the blocks they belong to. This approach decreases the computational
time while preserving the performances of the NL-means filter. In [26], the best values
for the parameters of Non-Local Means (NLM) algorithm are estimated. NLM is
parametric filter and highly dependent on the setting of its parameters. The authors
performed an exhaustive search to find the optimal parameter for NLM in context of MR
image de-noising. This paper estimates the optimum parameters for different noise
levels.

In [27], a new noise estimation method based on the adaptation of the Median Absolute
Deviation (MAD) estimator in the wavelet domain for Rician noise is proposed. Also the
impact of noise estimation accuracy on de-noising performance of the 3D Non-Local-
Means filter (NLM) is studied. In [28], three different sequential Wiener filters are
presented: isotropic is a sequential filter which similar to the classical Wiener filter uses
an isotropic neighborhood to estimate its parameters, orientation which uses oriented
neighborhoods to estimate the structure orientation present at each voxel and
anisotropic which selects locally either isotropic or oriented neighborhoods adaptively.
Section II explains two popular noise reducing methods in details. Section III represents
comparative study of state-of-arts works in noise reducing area. Table 1 lists the state-of-
art denoising algorithms.

1. TWO POPULAR NOISE REDUCING

METHODS
2. A. Anisotropic Filter
Perno and Maik [6] proposed anisotropic diffusion process. The equation is as the
follows:

I t ( x , y , t)=¿ ⁡(C ( x , y , t) ∇ I )=C(x , y , t) Δ I +∇ c . ∇ I

where div is divergence operator, I t (x , y , t ) is intensity of input image, t is the iteration
number, ∇ , Δ are the gradient and Laplacian operators and C ( x , y , t) ∇ I is a
monotonically decreasing diffusion function of the image gradient magnitude. The
gradient magnitude in boundary of region is higher than interior, and diffusion function
is monotonically decreasing. Therefore, the diffusion process happens in regions' interior
faster and the boundaries of regions would remain sharp. Different diffusion functions
are proposed [29]. Zhigeng et al. [30] used the following gauss function as diffusion
function:
Table 1. Advantages and disadvantages of noise reduction methods

De-noising method Advantages Disadvantages

 2 2

C (∇ I )=e ¿∇ I ¿ /2 K
where parameter K is the average gradient magnitude in the neighbor of each pixel and
specify degree of diffusion. Cattle et al. [31] used ∇∨G σ . u∨¿ as input for diffusion
function which uses smoothed image using Gaussian filter. Also following equation is
proposed [28]: C (s )=1/(1+ K )

where parameter K is the average of difference of gradient magnitude and maximum

gradient magnitude in the neighbor of each pixel.

3. B. NL-Means Method
In an image I , a neighborhood ηi of pixel i could be defined as an m× m window around
pixel i and the values of pixels in ηi is denoted by v ( η i) , a vector of intensities.
Furthermore, the pixel j , such that η j is similar to ηi , is a non local neighbor of pixel i .
Moreover, similarity of ηi and η j is defined by similarity of v ( η i) and v ( η j ), which is
measured by Euclidean distance. NL-means obtains s j, the signal value in the pixel i , by
averaging the value of pixels in non-local neighborhood of i .
si=∑ wi j I j
j ∈I

where w presents the weight of each non-local neighbor in averaging and is defined as
follow:
∥v ( ηi ) −v ( η j )
1 − h 2
w i j= e
Zi
where Zi is an normalizing factor and defined as follow:
∥ v ( ηi ) − v ( η j )
−
Zi j = ∑ e h2

j
where the parameter h controls the decay of exponential function. In other words, noise-
free value of pixel i is computed by weighted average of all the pixels in image, but pixels
with similar non-local neighborhood have larger weights in the average. Also, due to
decay factor, the pixels with large distance have weights near zero.

4. COMPARATIVE STUDY
The reported results for noise reducing algorithms are presented. In order to investigate
their effectiveness, the quantitative and qualitative results of the algorithms are
presented.

5. A. Simulated Images
In [32], Gaussian mixtures [33] with and without MRF is applied on phantom based
image from Brainweb. Average Dice similarity index for different algorithms with variant
noise levels (3%, 5%, 7%, 9%) are Gaussian mixtures ( 0.927 0.918 0.853 0.832 ) and
Gaussian mixtures+MRF (0.956, 0.949, 0.936 and 0.929). In the presence of all noise
levels, incorporating MRF improves segmentation results. The similarity index of the
Gaussian mixtures+MRF decreases more slowly than Gaussian mixtures algorithm when
noise level decreases. In [25], the proposed Optimized Blockwise NL-means filter is
applied on phantom based image from Brainweb with variant Gaussian levels.

The proposed filter is compared with three noise recuing algorithms: standard NL-
means filter [34], AD filter [6] and TV minimization [35]. The best values for the
parameters of AD filter and TV minimization scheme are estimated by exhaustive search.
The Peak Signal to Noise Ratio (PSNR) was used to qualitatively evaluate the new de-
noising algorithm. At studied noise levels, the best PSNR values are for proposed
algorithm. Also, the histograms of the de-noised images and the ground truth are
compared.

The produced histogram obtained by proposed algorithm is most similar one to

histogram of ground truth and has most sharp peaks (highest contrast). The distances
between the histograms of the de-noised images and the ground truth are also compared.
The distance obtained by proposed algorithm is the least one. Also, denoised image and
removed noised produced by competing methods are compared visually.
The proposed algorithm produces the most homogenous white matter. NL-means-based
algorithms remove the high frequencies related to noise while preserve the high
frequency information of anatomical structures more than other competing methods.
The AD filter spoils the edges especially on the skull. The TV minimization algorithm
preserves the edges slightly better but fails to remove all the noise.

The same experiments are performed with Rician noise. Again, the proposed algorithm
outperforms the classical ones. The NL-means-based algorithms de-noise corrupted
image and emphasizes the three main peaks corresponding to three main tissues in
histogram better than other competing algorithms. In compare to previous experiments
with Gaussian noise, the denoising of background is worse in the Rician case, but the
NLmeans filter de-noises correctly the cerebral structures, especially the white matter.
The filters perform almost similar to Gaussian case but since for the same level, Rician
noise is more pronounce than Gaussian case, the PNSR values slightly decrease.

In [26], the best values for the parameters of NonLocal Means (NLM) algorithm are
estimated. The NLM algorithm has three parameters: Rsearch (the radius of search
window), Rsim (the radius of the neighborhood window which is used to calculate the
similarity between each two pixels), h (which controls the degree of smoothing). In this
experiments, a 11− 11 search window ( Rsearch =5 ) is used as a reasonable size for medical
images and 3 phantom based MR images ( T 1 , P D and T 2 ) from the Brainweb were used
to perform experiment. The Root Mean Squared Error (RMSE) was used to qualitatively
evaluate NLM algorithm. In presence of variant noise levels (1%, 3%, 5%, 7 % and 9 % ),
for each Rsim value, RMSE is used to perform an exhaustive search for the optimum h
value.

When Rsim increases, the RMSE decreases but computation cost increases. Rsim =2 is
proposed because further increasing Rsim do not improve RMSE noticeable but increase
computation time notably. The optimum h value has linear relation with noise level.
Based on the results, the authors proposed general value of 1.2 r for h with Rsim =2 and
Rsearch =5 . The NLM with the proposed parameters are applied on different MR image. In
the residuals (difference between de-noised and original image), almost no anatomical
information can be noticed. Also, the Rician noise almost is removed. The performance
of NLM with optimal estimated parameters is compared qualitatively and quantitatively
with two denoising algorithms, the ADF (parameters manually tuned to get the best
possible results) [6] and a wavelet-based de-noising algorithm (the parameters proposed
by the authors) [36]. In almost all the cases, the NLM with proposed parameters values
produces lower RMSE.
In [24], a novel b-scale-based filtering method is presented which utilizes scale-
dependent diffusion conductance. The proposed b-scale-based diffusion (bD) is
compared with nonlinear complex diffusion (NCD) [37] , and g-scale-based diffusion
(gD) [38] qualitatively and quantitatively. Several brain MRI data sets including
phantom based images from Brainweb are used for qualitative evaluation. The three
methods smooth well the interior of homogeneous regions. The bD and gD methods
perform better than NCD in preserving fine details and the edges, independent of the
body region, the imaging modality and protocol. Most diffusion based methods smooth
across edges as side effect. The gD method diffuses well along edges and minimizes
diffusion across them even better than bD.

A single g-scale region along a boundary is likely to consist of a run of the boundary
voxels due to unrestricted shape of g-scale regions, whereas, every boundary voxel is
likely to be in a different b-scale region due to small size of b-scale regions. Therefore, gD
achieves diffusion along the boundary more than bD method. 45 MRI phantom based
image volumes from Brainweb with three levels of noise (3%, 7%, and 9%), three
protocols ( P D ,T 1 , and T 2 ) , and five slice thicknesses ( 1 m m, 3 m m, 5 m m, 7 m m, and
9 mm ) are used for quantitative comparison of the methods. Also, the relative contrast
of the object regions (RC), residual noise (RN) and the area under the curve (AUC) of
these values are used to evaluate the methods. The higher value of AUC is the more
effective the method is.

In term of AUC for every protocol and each level of noise, the scale-based diffusive
filtering (bD and gD) outperforms the NCD method and the gD method outperforms the
bD method. gD performs smoothing (lower RN) more than bD and NCD for the same
level of boundary blur (RC). Additionally, for the same level of noise suppression (RN), it
produces boundary blur (higher RC) less than bD and NCD. The gD method performs as
quickly as the NCD method for a 256 ×256 . 51 image and averagely takes under 1 mi n
for three iterations. The bD method requires more iteration and takes about 2 min for
roughly the same level of filtering on an image.

6. B. Real Images
In [28], three different sequential Wiener filters, namely, isotropic, orientation and
anisotropic are proposed. The proposed filters is compared in terms of the global MSE
with several reported methods: 3D median filter [39] with window length three and five
voxels in each dimension; an anisotropic diffusion filter [40] and three methods
presented in a flux-diffusion filter [41].

The flux-diffusion filter outperforms other competing methods for MSE measure;
however, the simple median filter with window length three voxels in each dimension
presents satisfactory result for MSE. Wiener filters in compare to median filter produce
de-noised image with less noise; the result of Wiener filters is not more blur than both
median filter and flux-diffusion filter. Additionally, the Wiener filter in compare to
fluxdiffusion filter takes less time and the lower amount of memory in the same machine.
Therefore, the proposed filters can be considered as candidate.
In [26], The Non-local Means (NLM) with optimal estimated values for parameters is
applied on a T 1 weighted sagittal MP-RAGE scan acquired on a Siemens 1.5 Tesla Vision
−

scanner from the fMRI Data Center database (www.fmridc.org) and two body images
acquired on a Philips 3 Tesla scanner from Hospital Quiron of Valencia (Spain). The
performance of NLM with optimal estimated parameters is compared qualitatively and
quantitatively with two de-noising algorithms, the ADF (parameters manually tuned to
get the best possible results) [6] and a wavelet-based denoising algorithm (the
parameters proposed by the authors) [36]. The proposed algorithm gives a residual
(difference between de-noised and original image) that less anatomical information can
be noticed. Also, the Rician noise is removed more than other competing methods. The
wavelet-based algorithm shows artefacts in result. Also, unnatural edge enhancement
and blurring of small edges are noticeable in ADF results.

In [32], 20 normal images from IBSR are catogerized in two groups: contaminated and
uncontaminated image volumes based on their quality. The ten contaminated image
volumes contain either the smooth intensity inhomogeneity, or rapid interslice intensity
variation. Gaussian mixtures [33] with and without MRF is applied on 10
uncontaminated images (11_3, 12_3, 13_3, 100_23, 110_3, 111_2, 112_2, 191_3, 202_3,
205_3). Jaccard similarity index for different methods are: Gaussian mixtures (0.6,
0.615, 0.6, 0.66, 0.625, 0.6, 0.66, 0.58 , 0.61 , 0.61¿ and Gaussian mixtures+MRF (0.725,
0.742 , 0.67 , 0.71, 0.645 , 0.632 , 0.662, 0.692 , 0.725, 0.72). The average Jaccard indexes
are: Gaussian mixtures ¿ 0.62, Gaussian mixtures + M R F=0.69 . The results show that
incorporating MRF improves segmentation quality.

Also, Gaussian mixtures [33] with and without MRF is applied on the eighteen newly
added image volumes from the IBSR. Jaccard similarity index for different methods are:
Gaussian mixtures ¿, 0.74 ,0.73 , 0.7 , 0.7 , 0.625 , 0.74 , 0.55 , 0.68 ,0.735 , 0.75 ,
0.8 , 0.85 , 0.727 , 0.76 ¿ and Gaussian mixtures+MRF (0.67, 0.73, 0.745, 0.68, 0.71, 0.76,
0.685, 0.73, 0.7, 0.737, 0.614, 0.7, 0.69, 0.745, 0.71, 0.69, 0.707, 0.68). Average Jaccard
similarity index for different methods are: Gaussian ¿ 0.715 , Gaussian+MRF ¿ 0.705 . The
Gaussian approach gives better results than those by the Gaussian+MRF model. It seems
that MRF models overregularize the segmentation. When comparing between results for
new dataset and results for 20 normal images, the segmentation results are improved.
The improvement in segmentation results is related to better quality of new added image
data.

In [27], a new noise estimation method based on the adaptation of the Median Absolute
Deviation (MAD) estimator in the wavelet domain for Rician noise is proposed. The
NLM with proposed and state-of-art noised estimators (ML [42], LMB [43], LVB [43],
LVO [43], MAD [44]) has been applied on the T 1 − w phantom corrupted with ghosting
inhomogeneity and variant noise levels from 2 % up to 15 %.