You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/319163534

Analysis of denoising filters on MRI brain images

Article  in  International Journal of Imaging Systems and Technology · August 2017


DOI: 10.1002/ima.22225

CITATIONS READS

12 740

2 authors:

Saritha Saladi Amutha Prabha Nagarajan


VIT University VIT University
16 PUBLICATIONS   26 CITATIONS    35 PUBLICATIONS   40 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

segmentation of brain View project

All content following this page was uploaded by Saritha Saladi on 21 March 2018.

The user has requested enhancement of the downloaded file.


Received: 22 December 2016
| Revised: 30 March 2017
| Accepted: 16 May 2017
DOI: 10.1002/ima.22225

RESEARCH ARTICLE

Analysis of denoising filters on MRI brain images

Saritha Saladi | N. Amutha Prabha

School of Electrical Engineering, VIT


Abstract
University, Vellore, India
The magnetic resonance imaging (MRI) modality is an effective tool in the diagnosis
Correspondence of the brain. These MR images are introduced with noise during acquisition which
Saritha Saladi, School of Electrical reduces the image quality and limits the accuracy in diagnosis. Elimination of noise
Engineering, VIT University, Vellore, in medical images is an important task in preprocessing and there exist different
India.
methods to eliminate noise in medical images. In this article, different denoising algo-
Email: saladi.saritha2014@vit.ac.in
rithms such as nonlocal means, principal component analysis, bilateral, and spatially
adaptive nonlocal means (SANLM) filters are studied to eliminate noise in MR.
Comparative analysis of these techniques have been with help of various metrics
such as signal-to-noise ratio, peak signal-to-noise ratio (PSNR), mean squared error,
root mean squared error, and structure similarity (SSIM). This comparative study
shows that the SANLM denoising filter gives the best performance in terms of better
PSNR and SSIM in visual interpretation. It also helps in clinical diagnosis of the
brain.

KEYWORDS
bilateral filter, denoising, MRI brain, NLM, PCA, SANLM

1 | INTRODUCTION images including linear filters and nonlinear filters. The


mean, median, and adaptive filters are linear methods and
The major problem occurring in biomedical images is noise, nonlinear methods include anisotropic diffusion, bilateral,
which affects the coherent nature of the images. This image wavelets, principal component analysis (PCA), nonlocal
denoising is applicable to other fields such as astronomy and means (NLM), and so forth.3
forensic sciences to obtain the useful potential information. The linear filters cause the image to blur by suppressing
Many researchers concentrate on different fields of magnetic the fine details. The performance of these filters is not
resonance imaging (MRI) brain images like preprocessing, acceptable because they blur the sharp edges, rescind lines,
enhancement, segmentation,1 and feature extraction. In gen- and other details. To avoid these difficulties nonlinear filters
eral, the noise interfered may be either Gaussian or speckle are used. The anisotropic diffusion filter is used to reduce
noise and the image capturing equipment itself incorporates noise without removing the important details from the image
salt and pepper noise. This noise degrades the images and (edges, lines) which are significant for the elucidation.4 Bilat-
leads to the incorrect diagnosis of the disease. Owing to this eral filtering is a nonlinear filter and noniterative method
coherent noise, it is more difficult to distinguish the adequate which combines domain and range filters simultaneously.5
details of the images during the diagnosis of the disease by a This technique is an alternative to wavelet thresholding
human expert. Various denoising filters are used to remove achieved by combining two Gaussian filters; first filter works
the noise and improve the image quality of computed tomog- in spatial domain and the second in intensity domain.6
raphy images.2 Denoising methods are intended to eliminate Wavelets challenges to remove noise by denoising the
the noise from an image without losing the original attributes noisy images, noise present in the signal while holding all
of the image. Thus, denoising the noisy image has become the signal characteristics irrespective of its frequency.7 The
the imperative step in processing the medical MR images. wavelet filter for denoising the MR images is precisely
The number of denoising techniques has performed on MR designed to lever the Rician noise.8 The main drawback of

Int. J. Imaging Syst. Technol. 2017;27:201–208. wileyonlinelibrary.com/journal/ima V


C 2017 Wiley Periodicals, Inc. | 201
202 | SARITHA AND AMUTHA PRABHA

this filtering method is that any small structures that are simi- 2.2 | Bilateral filter
lar in size to the noise are also eliminated.9Whereas PCA
offers an efficient and simple way to analyze multivariate Bilateral filter is well-defined as a weighted mean of adjoin-
statistical distribution to reduce optimal dimensionality, it ing pixels/voxels that is analogous to Gaussian convolution.
aids as one of the best approaches to produce statistic sam- The use of this filter is grown-up promptly and it is used in
ples.10,11 It functions as an effective technique to perform a applications such as image denoising and image enhance-
linear mapping with maximal variance and minimal mean ment.6 It is an efficient method to smooth the MR images
squared error (MSE).12 while preserving their disjoints and to separate the image
Among these solutions, the NLM filter is gained with an structures of different regions. Bilateral filtering is a method
increasing attractiveness due to its outstanding performance to smooth the images while preserving the edges. The bilat-
and verified for Gaussian, additive, and multiplicative eral filter considers the modification in values of the neigh-
noise.13 It not only eradicates noise from the input image in bors to preserve edges while smoothing.19
an effective way but also preserves fine details. The main
idea is to exploit the natural redundancy in the input image
to restore better image quality. The MR images are tainted 2.3 | Nonlocal means
by random noise which appears during the process of analy- NLM method uses an innovative idea in removing the noise
sis. This leads to the difficulty in extracting the features and from the images. It averages the similar pixels/voxels in the
segmenting the brain images.14 The significant property of a image based on the intensity distance to regain the single
good denoising method is to provide the noise free images pixel/voxel value. Estimation of each individual pixel/voxel
along with preserving nature of the edges. is based on the weighted mean of all pixels/voxels in the MR
The evaluation of denoising methods is compared in terms image. These weights are ascertained based on the compari-
of peak signal-to-noise ratio (PSNR), MSE, signal-to-noise son of the pixels/voxels.13 The NLM is a robust method for
ratio (SNR), root mean squared error (RMSE), and structure removing the noise, but limits the performance on wide
similarity (SSIM) values. Image denoising methods provide applications with high computational loads. Owing to this
better image quality by improving the visibility of fine details, drawback, NLM lacks the efficiency for larger noise remov-
which in turn helps in clinical diagnosis of the disease to als in the image.
obtain the accurate information. Different denoising MRI fil-
ters are discussed and compared with different metrics. This
helps in future processing like segmentation, registration and 2.4 | Principal component analysis–nonlocal
classification used for computer aided diagnosis.15 This article means
is organized as follows: PCA, NLM, bilateral, PCA-NLM,
PCA-NLM method computes neighborhood resemblances
and SANLM denoising filters are described in Section 2. Per-
once the PCA projection is done. This method works on the
formance metrics used in evaluation are described in Section
hypothesis that the image contains a widespread volume of
3. Comparative analysis of results is presented in Section 4
self-similarity. It performs best when compared with the
and concluded in Section 5.
other denoising methods in edge preserving and in enhancing
the visual appearance of the denoised MR image. This
2 | METHODS AND MATERIALS amendment to the NLM filter results in better accuracy and
the computational performance is improved.20 This PCA-
2.1 | Principal component analysis NLM method is compared with wavelets and conventional
PCA is a statistical method which takes a large number of PCA and acquires better results.21
interconnected variables and transmutes this data into a
smaller number of uncorrelated variables while recollecting
2.5 | Spatially adaptive nonlocal means
highest amount of dissimilarity. These uncorrelated varia-
bles are termed as principal components and this method SANLM filter improves the eminence of the image. SANLM
recognizes the outlines in data and assigns the emphasized deals with the noise level, spaciously erratic within the MR
resemblances and variances. PCA is used as the most dom- image.22 It can be estimated locally by processing data with
inant tool for analyzing data.16 PCA-based denoising static or spaciously erratic noise fields in fully automatic
method calculates the locally fitted base and transmutes the manner. Several adaptive filtering techniques are used to
image.17 PCA is applied directly on the MR images to improve the quality of magnitude MR images. The idea of
eliminate the noise and visual artifacts occur in the noisy spatially distributed noise fields has specific augmentation to
images.18 more interest and demands closer attention.22
SARITHA AND AMUTHA PRABHA | 203

3 | PERFORMANCE METRICS closeness of the two images x and y in terms of mean lumi-
nance (lx and ly ).
The performances of different filters are evaluated using dif- Contrast is the difference in color that makes an object
ferent performance metrics such as MSE, PSNR, RMSE, distinguishes the image representation. In visual perception,
SNR, and SSIM. Different perceptual quality assessments contrast is determined by the difference in the color and
are clearly explained for two-dimensional images with differ- brightness of the object and other objects within the same
ent evaluation metrics.23 field of view. The term c(x,y) is the contrast comparison
function, measures the contrasts in the images x and y are
how close to each other in terms of standard deviations
3.1 | MSE and PSNR (rx and ry ).
MSE and PSNR are the two quality measures used in the Structure provides the structural information in an
evaluation of the performance on images.24 The MSE image is defined as the attributes that represent the structure
between the original image and noisy image is defined as of objects in the scene, independent of the average luminance
and contrast. The term s(x,y) is the structure comparison
1 NX X
21 M 21
function, measures the correlation coefficient between the
MSE5 ½xði; jÞ2yði; jÞ2 (1)
N  M i50 j50 images x and y in terms of covariance(rxy ).
where x(i,j) is the original image, y(i,j) is the noisy image, SSIMðx; yÞ5½lðx; yÞa3½cðx; yÞb3½sðx; yÞg (6)
and N,M are the number of rows and columns in the image. where,
The PSNR is given as
  2lx ly 1C1
MAX2 lðx; yÞ5 (7)
PSNR510log10 (2) l2x 1l2y 1C1
MSE
2rx ry 1C2
where MAX-255, small MSE value means lesser the error, cðx; yÞ5 (8)
r2x 1r2y 1C2
results in high PSNR. Acceptable PSNR range is 30–50 dB.
rxy 1C 3
sðx; yÞ5 (9)
rx ry 1C3
3.2 | Root mean squared error
where lx is the local mean of x, ly is the local mean of y, rx
RMSE is defined as is the standard deviation of x, ry is the standard deviation of
pffiffiffiffiffiffiffiffiffiffi y, and rxy is the cross-covariance of x, y for the image. There-
RMSE5 MSE (3)
  fore, a 5 b 5 g 5 1 and C1 5ðL3K1 Þ2 andC2 5ðL3K2 Þ2 ,
MAX
PSNR520log10 (4) where L is the dynamic range of the pixel values (255 for 8-
RMSE bit grayscale images), C1 is a very small constant at K1  1,
C2 is non-negative constant at K2  1 and C3 5 C2=2.
The similarity index is simplified to
3.3 | Signal-to-Noise Ratio   
2lx ly 1C1 2rxy 1C2
SNR is defined as the ratio of the noisy image to that of the SSIMðx; yÞ5   (10)
difference between the original image and noisy image. l2x 1l2y 1C1 r2x 1r2y 1 C2
PN21 PM21 2
i50 j50 yði; jÞ
SNR5 PN21 PM21 2
(5)
i50 j50 ½xði; jÞ2yði; jÞ
3.5 | Analytical relationship between PSNR
and SSIM
3.4 | Structure similarity The analytical relationship between PSNR and SSIM is men-
SSIM is based on the multiplication of three terms, the lumi- tioned.25 First we obtain the relation between SSIM and
nance, the contrast, and the structural terms.24 MSE, and then we relate SSIM to PSNR. Equation 1 is
Luminance is the amount of energy an observer per- rewritten as
ceives from the light source. It is also defined as the product X
N X
21 M 21
of the illumination and the reflectance an object being NM  MSEðx; yÞ5 ½xði; jÞ2yði; jÞ2
i50 j50
observed in the image. But the structures of the objects in
the scene are independent of the illumination. The term l (x, by defining lX is the mean values of x and lY is the mean
y) is the luminance comparison function, measures the value of y. We derive
204 | SARITHA AND AMUTHA PRABHA

X
N X
21 M 21 X X
N 21 M 21      
2 1 1 1 1 aðx; yÞ3MSE1bðx; yÞ
½xði; jÞ2yði; jÞ2 5 xij 2lx 2 yij 2ly 1 lx 2ly 5 5 (18)
i50 j50 i50 j50
SSIM lðx; yÞ cðx; yÞ sðx; yÞ lðx; yÞsðx; yÞ
X
N X
21 M 21   X
N X
21 M 21 From Equation 2,
5 ½ðxij 2lx Þ2 yij 2ly  2 1 ðlx 2ly Þ2  
i50 J50 i50 j50 MAX2
PSNR510log10
X
N X
21 M 21   MSE
12 ½ðxij 2lx Þ2 yij 2ly ðlx 2ly Þ
i50 j50 MSE5 MAX2 3e2PSNR3lnð10Þ=10 (19)
(12) We show the relationship between SSIM and PSNR
Let us consider, from the equations:

X
N X
21 M 21    
1 aðx; yÞ3MAX2 3e2PSNR3lnð10Þ=10 1bðx; yÞ
xij 2lx 2 yij 2ly
2 5 (20)
A5 SSIM lðx; yÞsðx; yÞ
i50 J50  
X M21
X Let us assume that C2  max rx ; ry and C3 
N21 2  
B5 lx 2ly max rx ; ry , we obtain
i50 j50
1
X X
21 M 21  aðx; yÞ5
N   
  2rx ry
C52 xij 2lx 2 yij 2ly lx 2ly (13)
 2
i50 j50 2rxy 2 lx 2ly
bðx; yÞ5
Then, 2rx ry
rxy
X
N X
21 M 21 sðx; yÞ5
MN3MSEðx; yÞ5 ½xði; jÞ2yði; jÞ2 5 A1B1C rx ry
i50 j50
Substitute aðx; yÞ; bðx; yÞ and s(x,y) in expression,
(14)
2rxy 2ðlx 2ly Þ
2

Simplifying Equations 11, 12, and 13 yields 1


1
2rx ry 3MAX2 3e2PSNR3lnð10Þ=10 1 2rx ry
5 r
   2 SSIM lðx; yÞ rx xyry
A5NM r2x 1r2y 22rxy ; B5NM lx 2ly ; C50 (15)  2
1 MAX2 3e2PSNR3lnð10Þ=10 12rxy 2 lx 2ly
5 (21)
where rx is the standard deviation of the image x, ry is the SSIM 2rxy lðx; yÞ
standard deviation of the image y, rxy is the covariance
Rewriting the equation in terms of PSNR as,
between x and y. Finally from Equations 14 and 15, we
derive 2rxy lðx; yÞ  2
MAX2 3e2PSNR3lnð10Þ=10 5 22rxy 1 lx 2ly
SSIM
 2  
MSE5r2x 1r2y 22rxy 1 lx 2ly (16) 2PSNR3lnð10Þ=10 1 2rxy lðx; yÞ  2
e 5 22rxy 1 lx 2ly
MAX2 SSIM
From SSIM, consider term contrast, cðx; yÞ5 r2 1r
x y 2
2 1C is
2r r 1C
 2 !
x y 2
2rxy lðx; yÞ 2rxy 1 lx 2ly
rewritten based on the MSE (Equation 16): PSNR5210log10 2
MAX2 SSIM MAX2
1 r2x 1r2y 1C2
5 2 2rxy ðlðx; yÞ2SSIMÞ lx 2ly 2
cðx; yÞ rx 1r2y 1C2 PSNR5210log10 1
MAX2 SSIM MAX
 2
1 MSE 2rxy 2 lx 2ly 1C2 (22)
5 1
cðx; yÞ 2rx ry 1C2 2rx ry 1C2 (17) Let the luminance of the image lðx; yÞ51 and, assume
5aðx; yÞ3MSE1bðx; yÞ mean of the image x is equal to mean of the image y
where, (lx 5ly ). Then the PSNR is written as

1 2rxy ð12SSIMÞ
aðx; yÞ5 ; PSNR5210log10 (23)
2rx ry 1C2 MAX2 SSIM
 2 Now PSNR is written as
2rxy 2 lx 2ly 1C2
bðx; yÞ5    
2rx ry 1C2 2rxy 12SSIM
PSNR5210log10 210log
From the above expressions, SSIM is written as MAX2 SSIM
SARITHA AND AMUTHA PRABHA | 205

FIGURE 1 (a) Original MR image. (b) Noisy image (noise deviation 5 5)

FIGURE 2 Denoised images: (a) PCA, (b) NLM, (c) bilateral filter, (d) PCA-NLM, and (e) SANLM

   
MAX2 SSIM images lack the visual interpretation and leading to incorrect
PSNR510log10 110log (24) diagnosis of the disease by the clinician.
2rxy 12SSIM
The different performances metrics are used for evaluate
From these expressions, we can conclude the analytical and compared for noisy and denoised images in terms of
relation between PSNR and SSIM. PSNR, SNR, MSE, RMSE, and SSIM. In Table 1, the com-
parison of different denoising methods on noisy images for
4 | RESULTS AND DISCUSSIONS different measures is tabulated. These results are clearly in
the indicate that the SANLM method produces better results
A comparative analysis of the SANLM method with the con- in terms of all metrics that are compared with the other
denoising filters. Table 2 is clearly indicating that the
ventional denoising methods (PCA, NLM, bilateral filter,
and PCA-NLM) is implemented. MR image datasets of mul-
tiple sclerosis are acquired form many subjects of different T A BL E 1 Comparison of different denoising methods on noisy
age groups [http://www.medinfo.cs.ucy.ac.cy/]. These data- images for different measures
sets are then subjected to noise and a comparative analysis of
the existing methods is done to remove the noise. Figure 1 Filters Bilateral PCA-
measures PCA NLM filter NLM SANLM
shows the original image and the noisy image. Figure 2
shows the visual interpretation of the obtained denoised out- PSNR 34.3214 34.1687 35.4512 36.3270 37.2451
put MRI brain images. SNR 9.2783 9.1895 9.1727 9.2974 9.4543
The comparative analysis of PCA, NLM, bilateral, PCA-
NLM, and SANLM filters is performed with different noise MSE 9.5582 9.5734 9.5806 9.2145 9.0214
deviations r510; 15; 20, and 25. The PSNR and SSIM RMSE 3.0916 3.0941 3.0953 3.0355 3.0035
values of the noisy and denoised images are decreased and
SSIM 0.8849 0.8851 0.8878 0.8941 0.9014
the MSE and RMSE values are improved. The resultant
206 | SARITHA AND AMUTHA PRABHA

T A BL E 2 Comparison of different denoising methods on denoised


images for different measures

Filters Bilateral PCA-


measures PCA NLM filter NLM SANLM

PSNR 41.2317 42.2083 42.6807 42.7771 43.0942

SNR 13.4979 13.3142 13.2764 13.8742 14.4954

MSE 3.9202 3.9414 3.5352 3.4576 3.2141

RMSE 1.9799 1.9241 1.8802 1.8595 1.7928 F I G U R E 4 Comparison of SNR for different denoising filters.
[Color figure can be viewed at wileyonlinelibrary.com]
SSIM 0.9472 0.9486 0.9567 0.9597 0.9701

SANLM filter provides efficient and effective results com-


pared to other filters for the denoised images.
Figure 3 shows the graphical representation of PSNR
comparison between noisy and denoised MRI brain images.
NLM filters produce the least value for noisy image com-
pared to SANLM filter and its PSNR is improved by 9.00%.
PCA filters gives the least value for denoised image among
other filters and its PSNR is increased by 4.51%. With the
comparison of all filters, SANLM filter leads to the highest
PSNR for both noisy and denoised images on an average of F I G U R E 5 Comparison of MSE for different denoising filters.
6.27% and 20.7%, respectively. [Color figure can be viewed at wileyonlinelibrary.com]
Figure 4 explains the comparison of SNR between the
noisy and denoised MRI brain images; bilateral filter gives an average of about 2.40% and 6.11% as compared with
the lowest SNR in terms of both noisy and denoised images PCA, NLM, bilateral, and PCA-NLM denoising filters. The
improved by 3.06% and 9.18%. The highest SNR values are bilateral and PCA filters produce the highest value for a
given by the SANLM filter compared to other filters, the noisy image and denoised image, its RMSE is reduced by
SNR increased by a mean of 2.37% and 7.29% for both noisy 2.96% and 9.40%, respectively.
and denoised MR images. Figure 7 explains the comparison of SSIM between noisy
Figure 5 displays the comparison between noisy and and denoised MRI brains in which PCA filter yields the least
denoised images in terms of MSE, bilateral filters produces SSIM values in both images. The highest SSIM value is pro-
the highest value for noisy image compared to SANLM filter duced by the SANLM filter which is improved by 1.86%
and MSE is reduced by 5.83%. NLM filters give the maxi- and 2.41% for noisy and denoised MR images.
mum value for denoised image compared other filters and its The anisotropic diffusion, bilateral, and sparse represen-
MSE is decreased by 18.4%. SANLM filter leads the best tation methods are used for eliminating noise in MR images.5
results among other filters by an average of 4.79% and From the results, the anisotropic diffusion executes at lesser
13.1%. time and sparse representation executes at long time with
Figure 6 shows that SANLM filter yields better RMSE
value for both noisy and denoised images with a reduced by

F I G U R E 3 Comparison of PSNR for different denoising filters. F I G U R E 6 Comparison of RMSE for different denoising filters.
[Color figure can be viewed at wileyonlinelibrary.com] [Color figure can be viewed at wileyonlinelibrary.com]
SARITHA AND AMUTHA PRABHA | 207

improved PSNR of about 2.07% for denoised images. Bilat-


eral filters preforms least among other filters in noisy and
denoised MR images. The MSE and RMSE values are
reduced by a mean of 13.1% and 6.11% for denoised brain
images. The SNR and SSIM metrics are optimized by an
average of 7.20% and 2.32% for the denoised MR images.
From these results, it can be observed that variety of filters
work to eliminate noise in MR images. Depending on the
different types of noises and filtering methods are employed.
This analysis of denoising methods enhances the accurate-
ness of the MRI brain images for further processing such as
F I G U R E 7 Comparison of SSIM for different denoising filters.
[Color figure can be viewed at wileyonlinelibrary.com] segmentation and feature extraction.

high PSNR and low PSNR values are obtained for bilateral R EFE RE NC ES
filters. The noise in MR images is eliminated using linear fil- [1] Saritha S, Amutha Prabha N. A comprehensive review: Segmen-
ters and wavelets for denoising.26 Among these linear filters, tation of MRI images—Brain tumor. Int J Imag Syst Technol
wiener filter eliminates the noise with high PSNR but the 2016;26:295–304.
resolution of the images is enhanced by the wavelets by pre- [2] Suganthi M, Deepa P. Performance evaluation of various denois-
serving the edge and contour information. Study of denoising ing filters for medical image. Int J Comput Sci Information
filters is accomplished using linear filters such as adaptive, Technol 2014;5.
mean, and median on T2-weighted MR images.27 This pro- [3] Isaac B, ed. Handbook of Medical Image Processing and Analy-
vides the best evaluation as salt and pepper noise is elimi- sis. Academic Press; 2008.
nated by median filter and a speckle noise is reduced by [4] Perona P, Malik J. Scale-space and edge detection using aniso-
average filter. tropic diffusion. IEEE Trans Pattern Anal Mach Intelligence
The different MR images are denoised using various per- 1990;12:629–639.
formance metrics including picture quality scale and unique [5] Zhang L, Chen J, Zhu Y, Luo J. Comparisons of Several New
quality index along with PSNR and RMSE.28 The hybrid De-Noising Methods for Medical Images. In: Bioinformatics and
Biomedical Engineering, 2009. ICBBE 2009. 3rd International
median, bilateral, wavelets, and independent component
Conference on, pp. 1–4. IEEE, 2009.
analysis (ICA) filters are used in which ICA provides the
[6] Tomasi C, Manduchi R. Bilateral filtering for gray and color
better result in removing salt and pepper noise and wavelets
images. In: Computer Vision, 1998. Sixth International Confer-
works best for eliminating speckle noises. Various denoising ence on. IEEE, 1998.
methods such a NLM, bilateral, and linear filters are used for
[7] Savaji S, Arora P. Denoising of MRI images using thresholding
eliminating Rician noise in brain slices.29 The conclusion techniques through wavelet. Int J Sci Eng Technol Res 2014;1.
from the obtained results indicates that images are blurred [8] Nowak RD. Wavelet-based Rician noise removal for magnetic res-
and lacks the information while using linear filters. The onance imaging. IEEE Trans Image Process 1999;8:1408–1419.
NLM and bilateral filters produce better performance in pre- [9] Weaver JB, Xu Y, Healy DM, Cromwell LD. Filtering noise
serving the edges. In this article, SANLM filter is imple- from images with wavelet transforms. Magn Reson Med 1991;
mented that yields best performance by reducing noise to the 21:288–295.
maximum extent which is measured in terms of PSNR and [10] Abdi H, Williams LJ. Principal component analysis. Wiley Inter-
SNR. This SANLM filter produces the better results when discip Rev Comput Stat 2010;2:433–459.
compared to the conventional denoising methods in terms of [11] Hotelling H. Analysis of a complex of statistical variables into
all performance metrices. The improvement is due to the principal components. J Educ Psychol 1933;24:417.
nature of eliminating the noise level, which is spaciously [12] Jolliffe I. Principal Component Analysis. John Wiley & Sons,
erratic within the MR image along the noise fields. Ltd; 2002.
[13] Buades A, Coll B, Morel J-M. A non-local algorithm for image
denoising. In: 2005 IEEE Computer Society Conference on
5 | CONCLUSION Computer Vision and Pattern Recognition (CVPR’05). Vol. 2.
IEEE, 2005.
This article explores the implementation of various denoising [14] Devasena CL, Hemalatha M. Noise removal in magnetic resonance
approaches verified with Gaussian noise on the MRI brain images using hybrid KSL filtering technique. Methodology 2011;27.
images. Spatially adaptive NLM filter outperforms very effi- [15] Mohan J, Krishnaveni V, Guo Y. A survey on the magnetic res-
ciently and effectively, compared to other conventional onance image denoising methods. Biomed Signal Process Con-
denoising filters. This filter delivers better results with trol 2014;9:56–69.
208 | SARITHA AND AMUTHA PRABHA

[16] Smith LI. A tutorial on principal components analysis. Cornell [25] Hore A, Ziou D. Is there a relationship between peak-signal-to-
Univ USA 2002;51:52. noise ratio and structural similarity index measure? IET Image
[17] Muresan DD, Parks TW. Adaptive principal components and Process 2013;7:12–24.
image denoising. ICIP 2003;1. [26] Rajeshwari S, Sree Sharmila T. Efficient quality analysis of MRI
[18] Kumar RH, Kumar BV, Gowthami S. Performance analysis of LPG image using preprocessing techniques. In: Information & Com-
PCA algorithm in medical images. In: Machine Vision and Image munication Technologies (ICT), 2013 IEEE Conference on, pp.
Processing (MVIP), 2012 International Conference on. IEEE, 2012. 391–396. IEEE, 2013.
[19] Paris S, Kornprobst P, Tumblin J, Durand F. Bilateral Filtering: [27] Isa IS, Sulaiman SN, . Mustapha M, Darus S. Evaluating denois-
Theory and Applications. Now Publishers Inc.; 2009. ing performances of fundamental filters for T2-weighted MRI
[20] Ruikar SD, Doye DD. Wavelet based image denoising tech- images. Proc Comput Sci 2015;60:760–768.
nique. Int J Adv Comput Sci Appl 2011;2. [28] Deepa B, Sumithra MG. Comparative analysis of noise removal
[21] Bansal M, Devi M, Jain N, Kukreja C. A proposed approach for techniques in MRI brain images. In: Computational Intelligence
biomedical image denoising using PCA-NLM. Development 2014;6. and Computing Research (ICCIC), 2015 IEEE International
Conference on, pp. 1–4. IEEE, 2015.
[22] Manjon JV, Coupe P, Marti-Bonmati L, Louis Collins D, Robles
M. Adaptive non-local means denoising of MR images with spa- [29] Oza SD, Joshi KR. Performance analysis of denoising filters for
tially varying noise levels. J Magn Reson Imag 2010;31:192–203. MR images. In: Advances in Computing Applications. Singapore:
Springer; 2016:87–96.
[23] You J, Xing L, Perkis A, Wang X. Perceptual quality assess-
ment for stereoscopic images based on 2D image quality metrics
and disparity analysis. In: Proc. of International Workshop on
Video Processing and Quality Metrics for Consumer Electronics, How to cite this article: Saladi S, Amutha Prabha N.
Scottsdale, AZ, USA. 2010. Analysis of denoising filters on MRI brain images. Int.
[24] Hore A, Ziou D. Image quality metrics: PSNR vs. SSIM. In: J. Imaging Syst. Technol. 2017;27:201–208. https://doi.
Pattern Recognition (ICPR), 2010 20th International Conference org/10.1002/ima.22225
on, pp. 2366–2369. IEEE, 2010.

View publication stats

You might also like