Proceedings of the World Congress on Engineering and Computer Science 2019

WCECS 2019, October 22-24, 2019, San Francisco, USA

Processing MRI Brain Image using OpenMP

and Fast Filters for Noise Reduction
Luis Cadena, Darwin Castillo, Aleksandr Zotin, Patricia Diaz, Franklin Cadena, Gustavo Cadena,
Yuliana Jimenez

Abstract— The medical diagnosis of most pathologists short time periods. Parallel computing provides concurrency
requires the analysis of the image studies. Therefore, it is and by this we can use non-local recourses very efficiently. It
important to get the best quality of the images without noise and also removes the limit of serial computing.
highlight the details of tissues. The principal aim of this work is
Medical image processing has experienced dramatic
to apply different algorithms and filters to reduce the noise of
magnetic resonance brain images, due to the noise in these can expansion and has been an interdisciplinary research field
cause to give a difficult diagnosis. The algorithms considered in attracting expertise from applied mathematics, computer
this work are the fast mean filter, fast Gaussian filter, and fast sciences, engineering, statistics, physics, biology and
median filter; also was used parallel programming in OpenMP. medicine. Computer-aided diagnostic processing has already
The results show that the parallel implementation of algorithms become an important part of clinical routine. Accompanied
has more performance in the time processing, localization, and
by the rush of new development of high technology and use
noise reduction than sequential and classic implementation.
of various imaging modalities, more challenges arise; for
Index Terms— MRI brain image, fast mean filter, fast median example, how to process and analyze a significant volume of
filter, fast gaussian 2D, parallel programming, OpenMP. images so that high quality information can be produced for
disease diagnoses and treatment.
I. INTRODUCTION One of the recent innovations in computer engineering has

M EDICAL imaging is the technique and process used to

create images anatomic, physiological or functional of
the human body for clinical purposes or medical purpose.
been the development of multicore processors, which are
composed of two or more independent cores in a single
physical package. Today, many processors, including digital
There are many different medical image modalities like CT, signal processor (DSP), mobile, graphics, and general-
PET, MRI, X-ray, Ultrasound imaging, etc. purpose central processing units (CPUs) have a multicore
These modalities have different features and are used as design, driven by the demand of higher performance. Major
pre requirements. Due the size of each image is very large; CPU vendors have changed their strategy away from
analyze these modalities takes much time to process increasing the raw clock rate to adding on-chip support for
sequentially. So, if we divide this sequential processing to multi-threading by increasing the number of cores; dual- and
efficient parallel processing then we can find good results in quad-core processors are now commonplace. Signal and
a very reasonable time or if we are able to process basic steps image processing programmers can benefit dramatically from
like image enhancement, morphological operation, feature these advances in hardware by modifying single-threaded
calculation quickly then it will be beneficial for medical code to exploit parallelism to run on multiple cores.
practice. Hence, by using parallel computing we can save This work present medical images processing of MRI with
time and money and we can solve large problems in very the principal aim to show a simple and efficient technique to
remove noise from the medical images, which combines
median filtering, mean filtering and Gauss filter to determine
Manuscript received July 02, 2019. This work was supported by the pixel value in the noise less image.
Vicerrectorado de Investigación, Universidad Técnica Particular de Loja
Magnetic resonance imaging (MRI) has created a lot of
UTPL, Av. Marcelino Champagñat s/n, Loja Ecuador.
L. Cadena, is with Electric and Electronic Department at Universidad de interest between Medical Professionals and patients, because
las Fuerzas Armadas ESPE, Av. Gral Ruminahui s/n, Sangolqui Ecuador. it provides anatomical and physiological information in a
(phone: +593997221212; e-mail: [email protected]). non-invasive way. MRI does not use any kind of ionizing
D. Castillo is with the Chemistry and Exact Sciences Department,
Universidad Técnica Particular de Loja UTPL, Av. Marcelino Champagñat radiation. MRI creates images of structures through the
s/n, Loja Ecuador (e-mail: [email protected]) and also with I3M, interactions of magnetic fields and radio waves with tissues.
Universitat Politècnica de Valencia, Camino de Vera, 46022 Valencia-Spain. Knowledge of the probable pathology is fundamental,
A. Zotin is with Department of Informatics and Computer Techniques,
Reshetnev Siberian State University of Science and Technology, 31 choosing the appropriate technique and analyzing the correct
krasnoyarsky rabochу av., Krasnoyarsk 660037, Russian Federation (e-mail: region of the body.
[email protected]). Further, MRI is an imaging technique that produces high
F. Cadena is with College Unidad Educativa Atahualpa. Oyacoto-Quito,
Ecuador (e-mail: [email protected])
quality images of the anatomical structures of the human
G. Cadena is with Medicine Faculty at Universidad Central del Ecuador, body, especially in the brain, and provides rich information
Iquique 132, Quito-Ecuador (e-mail: [email protected]) for clinical diagnosis and biomedical research [9,10]. It is the
P. Diaz Guzman is with Health and Medical Science Department at
Universidad Técnica Particular de Loja UTPL, Av. Marcelino Champagñat
most commonly used imaging modality as it offers high-
s/n, Loja Ecuador. (e-mail: [email protected]) resolution images in a noninvasive and safe method, without
Y. Jimenez is with the Chemistry and Exact Sciences Department, exposing patients to ionizing radiation. Significant attention
Universidad Técnica Particular de Loja UTPL, Av. Marcelino Champagñat is given also to two applications for which MRI has unique
s/n, Loja Ecuador (e-mail: [email protected]).

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2019
WCECS 2019, October 22-24, 2019, San Francisco, USA

potential: blood flow imaging and quantification, and Classic and fast median filter
functional neuroimaging based on exploiting dynamic Classic median filter replaces the value of a pixel spectrum
changes in the magnetic susceptibility. by the median of the spectrum levels in the neighborhood of
Additionally, MRI technique has become a critically that pixel.
important tool in diagnosis and differentiation of different Median filtering is a commonly applied non-linear filtering
demyelinating disorders, because it offers high-resolution technique that is particularly useful in removing speckle and
images in a noninvasive and safe method, without exposing salt and pepper noise. It works by moving through the image
patients to ionizing radiation. Thus, MRI uses magnetic field pixel by pixel, and replacing each value with the median
gradients to modify the frequency and phase of the MR signal
in a controlled manner. The images are reconstructed through
mathematical algorithms to convert the collected MR signals
into spatial information [1-12].

II. DIGITAL PROCESSING OF IMAGES

The digital processing of images consists of algorithmic
processes that transform an image into another in which
certain information of interest is highlighted, and/or the
information that is irrelevant to the application is attenuated
or eliminated. For remove noise we used fast filters, and to Fig. 3. Classic median filter.
evaluate quality used SSIM measure. Following we describe
these methods. value of neighbor pixels.
Classic and fast mean filter The fast median filter is obtained through the histogram of
The arithmetic classic mean filter is defined as the mean of spectrum for median calculation can be far more efficient
all pixels spectrum within a local region of an image. because it is simple to update the histogram from window to
window. Thus the histogram used for accumulating pixels in
the kernel and only a part of it is modified when moving from
one pixel to another [8-13, 16].

Fig. 1. Classic mean filter.

Fig. 4. Fast median filter.

The fast mean filter obtained by accumulation of the
neighborhood of pixel P(y,x), shares a lot of pixels in
Classic and fast 2D Gauss filter
common with the accumulation for pixel P(y,x+1). This Gauss 2D classic filter calculate kernel Gauss bell G(x,y),
means that there is no need to compute the whole kernel for take pixels from gray value image A in kernel area and add to
all pixels except only the first pixel in each row. Successive sum considering Gaussian coefficient, and put obtained value
pixel filter response values can be obtained with just add and in study pixel in image B.
a subtract to the previous pixel filter response value. The Gaussian filter uses a Gaussian function (which also
expresses the normal distribution in statistics) for calculating
the transformation to apply to each pixel in the image.
The equation of a Gaussian function in one dimension is
x2
1 
G ( x)  e 2 2
2  
In two dimensions, it is the product of two such Gaussians,
one in each dimension:
x2  y2
1 
G ( x, y )  e 2
2

2   2

where x is the distance from the origin in the horizontal axis,

y is the distance from the origin in the vertical axis, and σ is
the standard deviation of the Gaussian distribution.
Since the image is represented as a collection of discrete
pixels it is necessary to produce a discrete approximation to
the Gaussian function before perform the convolution.
Depends on kernel size and σ some of coefficients can be out
range of kernel. Theoretically the Gaussian distribution is
non-zero everywhere, which would require an infinitely large
Fig. 2. Fast mean filter. convolution kernel. In practice it is effectively zero more than

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2019
WCECS 2019, October 22-24, 2019, San Francisco, USA

about three standard deviations from the mean. Thus it is OpenMP is an API that implements a multi-threaded,
possible to truncate the kernel size at this point. Sometimes shared memory form of parallelism. It uses a set of compiler
kernel size truncated even more. Thus after computation of directives that are incorporated at compile-time to generate a
Gaussian Kernel, the coefficients must be corrected that way multi-threaded version of program code. OpenMP is designed
that the sum of all coefficients equals 1. Once a suitable for multi-processor/core, shared memory machines [20-24].
kernel has been calculated, then the Gaussian smoothing can Measure PSNR and SSIM
be performed using standard convolution methods. The Any processing applied to an image may cause an
convolution can in fact be performed fairly quickly since the important loss of information or quality. Image quality
equation for the 2-D isotropic Gaussian is separable into y evaluation methods can be subdivided into objective and
and x components. In some cases the approximation of subjective methods. Subjective methods are based on human
Gaussian filter can be used instead of classic version [13-19]. judgment and operate without reference to explicit criteria.
Objective methods are based on comparisons using explicit
numerical criteria, and several references are possible such as
the ground truth or prior knowledge expressed in terms of
statistical parameters and tests.
The next equations show the relationship between the
SSIM (structural similarity index measure) and the PSNR
(peak-signal-to-noise ratio) for grey-level (8 bits) images.
Given a reference image f and a test image g, both of size
M×N, the PSNR between f and g is defined by:
2552
𝑃𝑆𝑁𝑅(𝑓, 𝑔) = 10𝑙𝑜𝑔10 ( )
𝑀𝑆𝐸(𝑓, 𝑔)
where
𝑀 𝑁
1 2
𝑀𝑆𝐸(𝑓, 𝑔) = ∑ ∑(𝑓𝑖𝑗 − 𝑔𝑖𝑗 )
𝑀𝑁
𝑖=1 𝑗=1

The PSNR value approaches infinity as the MSE

Fig. 5. Separable 2D Gauss filter. approaches zero; this shows that a higher PSNR value
provides a higher image quality. At the other end of the scale,
Parallel programming in OpenMP a small value of the PSNR implies high numerical differences
Parallel Programming may speed up code. Today between images.
computers have one or more CPUs that have multiple The structural similarity (SSIM) index is designed to
processing cores (Multi-core processor). This helps with improve on traditional methods such as peak signal to noise
desktop computing tasks like multitasking (running multiple ratio (PSNR) and mean squared error (MSE), which have
programs, plus the operating system, simultaneously). For proven to be inconsistent with human visual perception.
scientific computing, this means the ability in principle of Structural information is the idea that the pixels have
splitting up computations into groups and running each group strong interdependencies especially when they are spatially
on its own processor. close. These dependencies carry important information about
Two main paradigms talk about here are shared memory the structure of the objects in the visual scene. Luminance
versus distributed memory models. In shared memory masking is a phenomenon whereby image distortions (in this
models, all multiple processing units have access to the same context) tend to be less visible in bright regions, while
memory space. This is the case on desktop or laptop with contrast masking is a phenomenon whereby distortions
multiple CPU cores. In a distributed memory model, multiple become less visible where there is significant activity or
processing units each of their have their own memory store, "texture" in the image.
and information is passed between them. This is the model The mean structural similarity index is computed as
that a networked cluster of computers operates with. A follows:
computer cluster is a collection of standalone computers that Firstly, the original and distorted images are divided into
are connected to each other over a network, and are used blocks of size 8 x 8 and then the blocks are converted into
together as a single system. vectors. Secondly, two means and two standard derivations
The methodology in our case of the algorithms (filters) for and one covariance value are computed from the images as:
processing images is: 𝑇 𝑇
1 1
1.- Select Kernel 𝜇𝑥 = ∑ 𝑥𝑖 𝜇𝑦 = ∑ 𝑦𝑖
2.- Evaluate denoise filter with parallel OpenMP 𝑇 𝑇
𝑖=1 𝑖=1
𝑇
#pragma omp parallel for 1
for (int y=0; y< Image_Height; y++) 𝜎𝑥2 = ∑(𝑥𝑖 − 𝑥̅ )2
𝑇−1
for (int x=0; x< Image_Width; x++) 𝑖=1
𝑇
{ 1
// do denoise filters 𝜎𝑦2 = ∑(𝑦𝑖 − 𝑦̅)2
𝑇−1
} 𝑖=1

3.- Processed pixel put in study pixel of image denoise

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2019
WCECS 2019, October 22-24, 2019, San Francisco, USA

𝑇
2
1
𝜎𝑥𝑦 = ∑(𝑥𝑖 − 𝑥̅ ) (𝑦𝑖 − 𝑦̅)
𝑇−1
𝑖=1
Thirdly, luminance, contrast, and structure comparisons
based on statistical values are computed, the structural
similarity index measure between images x and y is given by:
(2𝜇𝑥 𝜇𝑦 + 𝑐1 )(2𝜎𝑥𝑦 + 𝑐2 )
𝑆𝑆𝐼𝑀(𝑥, 𝑦) = 2
(𝜇𝑥 + 𝜇𝑦2 + 𝑐1 )(𝜎𝑥2 + 𝜎𝑦2 + 𝑐2 )
where c1 and c2 are constants [25].

III. EXPERIMENTAL RESULTS a)

Different images, with different sizes were processed with
classic and fast filters: mean, median, Gaussian 2D. For the
experiment used PC based on Intel Core i5 3.1 GHz with 8
GB RAM. The results were obtained by measuring the
processing time of 80 different images (for each image, 400
measurements were taken). The results are presented in Fig.
6-8 and Tables I-IV.

Brief description of medical images used:

MRI Brain: The image presents a calcification between the
cortex and the stem. These images are very sensitive for study b)
of congenital and acquired structural anomalies such as
infectious, hemorrhagic, tumor, degenerative, and chronic
pathologies.
Fig. 6 shows processing time of classic and fast or
optimized filters for different image size (4 threads OpenMP).
In addition, a study of fast filter implementations was made.
It showed the magnitude of the acceleration relative to the
sequential implementation of the classical version of the
filters. The result of this study was represented as the maps of
acceleration of fast filters, showing acceleration coefficient
depending on the kernel size (Fig. 6). Also, the acceleration
stability of fast algorithms was evaluated depending on the c)
Fig. 6. Evaluation of optimized (fast) filters acceleration: (a) Mean filter; (b)
size of the core and the number of threads used. To estimate Gaussian filter; (c) Median filter.
the acceleration, the mean values obtained during the 300
measurements were taken for each combination of the kernel We conducted experimental research on MRI brain
size and the number of threads. medical images Fig. 7.
The experimental results show that the increase in the
processing speed for different kernel sizes is almost the same.
Some stability is observed in the acceleration for two threads
as well as one can see the increase of the acceleration
coefficient in the case with more than two threads having the
kernel size larger than 5×5. Acceleration with the usage of
four threads demonstrates poor efficiency as parts of the CPU
resources are spent on background tasks (Fig. 6).
In addition to the speed evaluation, evaluations of noise
suppression characteristics were also performed. To simulate
the noise which may occur in the equipment, the following
noise filter were used. Thus, in the noise filter was layered
over the image, the additive noise part was 80%, while the
impulse noise part was 20%. The magnitude of additive noise
component considered at maximum as 60% of the dynamic
range of the experimental data.
The experiments were carried out as follows. For each Fig. 7 Test image.

noise level, 300 noise maps were generated, which were

superimposed on the original image. Next, filters (with Table I demonstrates average PSNR values in dB for fast
different kernel sizes) were applied to the noisy image and the filters when processing noise images with size 630×630
PSNR and SSIM metrics were calculated. pixels. Table II demonstrates average SSIM values for the
same images in range 0 to 100%.

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2019
WCECS 2019, October 22-24, 2019, San Francisco, USA

TABLE I. larger than 5×5. Acceleration with the usage of four threads
PSNR EVALUATION_OPTIMIZED FILTERS IN dB
demonstrates reduced efficiency as parts of the CPU
Noise data Kernel Filter
Level PSNR Size Mean Gauss Median resources are spent on background tasks.
3×3 26,4499 26,6162 31,4632
5×5 25,4441 26,3348 26,0941
10% 19,4337 7×7 24,4246 25,9878 24,0914
9×9 23,9051 25,9163 23,5251
11×11 23,4213 25,9019 23,0034
3×3 25,0951 25,2115 31,3586
5×5 24,6368 25,4533 25,7129
15% 17,7498 7×7 24,0103 25,3044 24,1180
9×9 23,5739 25,2658 23,5490
11×11 23,1413 25,2572 23,0291
3×3 23,9026 23,9860 31,0633
5×5 24,0798 24,6117 26,1548
20% 16,4937 7×7 23,5122 24,5513 24,1792 a) b)
9×9 23,1564 24,5395 23,6019
11×11 22,7774 24,5358 23,0754
3×3 23,0048 23,0674 30,8252
5×5 23,4580 23,8787 26,1362
25% 15,6150 7×7 23,0348 24,2929 24,2111
9×9 22,7388 24,1965 23,6280
11×11 22,4073 23,8957 23,1053

TABLE II
SSIM VALUES (%) OPTIMIZED FILTERS
Noise data Kernel Filter
Level SSIM Size Mean Gauss Median
3×3 52,5261 52,9341 98,9732 c) d)
5×5 56,8031 58,4583 96,8460 Fig. 8. Processing image with noise using openMP: (a) original noise image
10% 30,8309 7×7 55,3762 59,2824 94,6519 (b) Mean filter; (c) Gaussian Filter; (d) Median filter.
9×9 52,2012 59,1818 92,5806
11×11 48,9455 59,1173 90,5115 TABLE III
3×3 40,5656 40,9190 98,7282 THE ACCELERATION COEFFICIENT OF FAST FILTERS IN
5×5 45,6737 47,3892 96,3257 PARALLEL MODE RELATIVE TO CLASSIC IMPLEMENTATION FOR
15% 20,3248 7×7 45,0116 48,2220 94,5730 IMAGE 2140×1740
9×9 42,2631 48,1409 92,5113 Kernel size
11×11 39,3629 48,0626 90,4537 Threads 3×3 5×5 7×7 9×9 11×11
3×3 32,5422 32,8148 98,1682 Mean Filter
5×5 38,5104 39,4721 96,6738 2 2,675 4,056 5,713 6,950 8,548
20% 14,6082 7×7 38,5305 40,8086 94,4997 3 3,784 5,334 7,872 10,037 12,173
9×9 36,5191 40,8399 92,4367 4 3,850 6,454 7,737 10,816 11,979
11×11 34,1926 40,7849 90,3801 Gaussian Filte
3×3 27,5511 27,7515 97,3491 2 1,862 3,447 5,517 6,883 9,382
5×5 34,0901 34,7222 96,5549 3 2,379 4,775 7,544 10,080 11,866
25% 11,4076 7×7 34,9439 36,4118 94,4181 4 2,723 5,243 8,101 10,994 13,182
9×9 33,5629 36,5819 92,3665 Median Filter
11×11 31,6682 36,5630 90,3168 2 9,947 21,101 33,756 47,442 64,712
3 13,221 26,773 41,607 68,399 88,632
4 14,220 30,719 47,108 79,472 94,799
Visual results of filtered noise modeled MRI brain image
shown in Fig. 8. To create examples of noise suppression for TABLE IV
fast filters demonstrated on noise map processing, a second THE ACCELERATION COEFFICIENTS OF FAST FILTERS FOR
image was produced at 20% of the original image size for IMAGE 2140×1740
demonstration of filters processing selected (5×5). Kernel size
Threads 3x3 5x5 7x7 9x9 11x11
The overall average acceleration of fast filters (parallel Mean Filter
mode) in comparison with sequential implementation of 2 2,038 2,122 2,083 2,112 1,923
classic filters for model image (size 2140×1740) shown in 3 2,883 2,791 2,870 2,817 2,738
Table III. The data of processing speed acceleration of fast 4 3,134 3,277 3,420 3,336 3,394
Gaussian Filter
filters for model image is presented in Table IV. 2 1,832 1,862 1,867 1,982 1,921
3 2,340 2,410 2,653 2,610 2,698
IV. CONCLUSIONS 4 2,663 2,856 2,742 2,847 2,899
Experiments were conducted to estimate the processing Median Filter
time of fast filtering algorithms (Mean filter, Median filter, 2 2,023 2,002 1,970 1,994 1,930
3 2,689 2,740 2,758 2,745 2,743
Gaussian Filter) and evaluation of noise suppression. The 4 2,892 2,914 3,049 2,951 2,927
experimental results show that the increase in the processing
speed for different kernel sizes is almost the same. Some Since the medical point of view any medical image
stability is observed in the acceleration for two threads as well presents an acceptable amount of noise, however, the
as one can see the increase of the acceleration coefficient in important thing to considerate is that this amount of noise
the case with more than two threads having the kernel size does not affect the quality of the image and especially the

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2019
WCECS 2019, October 22-24, 2019, San Francisco, USA

medical diagnosis. [15] A. Lukin, “Tips & Tricks: Fast Image Filtering Algorithms,” 17-th
International Conference on Computer Graphics GraphiCon'2007:
Different medical specialists have interpreted the filters 186–189, 2007.
that were used in this study and depending on the pathology, [16] G. A. Pascal, “A Survey of Gaussian Convolution Algorithms,” Image
they validated the importance of image noise processing. The Processing On Line 3: 286–310, 2013.
[17] S. Perreault, P. Hebert, “Median filtering in constant time,” IEEE
experimental results show that Median filter demonstrates the
Transactions on Image Processing 16(9): 2389–2394, 2007.
best noise reduction, though in some cases it suppresses [18] R.C. Gonzalez, R.E. Woods, “Digital Image Processing,” 3rd edition,
details. Prentice-Hall, 2008. ISBN-13: 978-0131687288, 2008.
In order to the processing, the experimental results show [19] A. Zotin K. Simonov, F. Kapsargin, T. Cherepanova, A. Kruglyakov,
L. Cadena, “Techniques for Medical Images Processing Using Shearlet
that the increase in the processing speed for different kernel Transform and Color Coding,” In: Favorskaya M., Jain L. (eds)
sizes is almost the same. Some stability is observed in the Computer Vision in Control Systems-4. Intelligent Systems Reference
acceleration for two threads as well as one can see the Library, vol 136. Springer, Cham. Chapter First Online: 27 October
2017 DOI https://doi.org/10.1007/978-3-319-67994-5_9.
increase of the acceleration coefficient in the case with more [20] L. Huang, et al., “Parallelizing Ultrasound Image Processing using
than two threads having the kernel size larger than 5×5. OpenMP,” on Multicore Embedded Systems. 978-1-4673-5085 -
Acceleration with the usage of four threads demonstrates 3/12/$31.00 ©2012 IEEE.
[21] S. Patel, “A Survey on Image Processing Techniques with OpenMP,”
reduced efficiency as parts of the CPU resources are spent on © 2015 IJEDR | Volume 3, Issue 4 | ISSN: 2321-9939.
background tasks. [22] R. Chandra, L. Dagum, D. Kohr, D. Maydan, J. McDonald, R. Menon,
Using OpenMP, we made parallel implementation of fast “Parallel programing in openmp,” Academic Press. USA. 249p ISBN
1-55860-671-8, 2001.
algorithms, which gives performance boost up in almost two [23] A. Kiessling, “An Introduction to parallel programming with
times for two threads and around 3, 2 times for 3 and 4 OpenMP,” A Pedagogical Seminar. The University of Edinburgh. UK,
threads. Experimental results demonstrate that the fast 2009.
[24] G. Slabaugh, et al., “Multicore Image Processing with OpenMP,”
version of filter algorithms can well do with the noise
unpublished.
reduction at appropriate minimum processing time compared [25] A. Horé, D. Ziou, “Image quality metrics: PSNR vs. SSIM,” 2010
to classical implementation. The greatest increase of International Conference on Pattern Recognition. 2010 IEEE. DOI
processing speed was gained for the median filter. For quality 10.1109/ICPR.2010.579, 1051-4651/10
processing used SSIM measure with good result, which
showed in table II.

REFERENCES

[1] S. Sanjay, S. Neeraj, S. Shiru, “Image Processing Tasks using Parallel

Computing in Multi core Architecture and its Applications in Medical
Imaging”. International Journal of Advanced Research in Computer
and Communication Engineering Vol. 2, Issue 4, April 2013.
[2] H. Zhu, “Medical image processing Overview,” unpublished.
[3] A. S. Y. Bin-Habtoor et al, “Removal Speckle Noise from Medical
Image Using Image Processing Techniques,” (IJCSIT) International
Journal of Computer Science and Information Technologies, Vol. 7 (1)
2016, 375-377
[4] Y. Zhang, L. Wu, S. Wang, “Magnetic resonance brain image
classification by an improved artificial bee colony algorithm," Progress
In Electromagnetics Research, Vol. 116, 65 -79, 2011.
[5] C. Kadam, S. Borse, “A Comparative Study of Image Denoising
Techniques for Medical Images. image,” (2017) 4(06).
[6] A. Dogra, B. Goyal, “Medical Image Denoising,” Austin J Radiol.
2016; 3(4): 1059
[7] R. Singh, P. Sapra, V. Verma, “An advanced technique of de-noising
medical images using ANFIS,” International Journal of Science and
Modern Engineering, 1(9), 2013.
[8] D. Trinh, M. Luong, J. Rocchisani, C. Pham, F. Dibos,. “Medical image
denoising using kernel ridge regression,” In 2011 18th IEEE
International Conference on Image Processing (pp. 1597-1600). (2011,
September) IEEE.
[9] A. Velayudham, D. Kanthavel, “A Survey on Medical Image
Denoising Techniques,” International journal of Advanced research in
Electronics and Communication Engineering (IJARECE), (2013)
ISSN.
[10] A. Kaneria, "Image Denoising Techniques: A Brief Survey,” The SIJ
Transactions on Computer Science Engineering & its Applications
(CSEA), The Standard International Journals (The SIJ), (2015) Vol. 3,
No. 2, Pp. 32-37.
[11] R. Rajni, A. Anutam. “Image Denoising Techniques - An Overview,”
International Journal of Computer Applications. (2013). 86.
10.5120/15069-3436.
[12] R. Kundu, A. Chakrabarti, “De-Noising Image Filters for Bio-Medical
Image Processing,” CSI Communications. 2014.
[13] Chandel et al., “Image Filtering Algorithms and Techniques: A
Review,” International Journal of Advanced Research in Computer
Science and Software Engineering 3(10), pp. 198-202, 2013.
[14] B. Gupta, N. Singh, “Image Denoising with Linear and Non-Linear
Filters: A Review,” International Journal of Computer Science Issues,
Vol. 10, Issue 6, No 2, pp. 149-154, 2013.

ISBN: 978-988-14048-7-9 WCECS 2019

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)