New Bioinspired Filter of DICOM Images
Arata Andrade Saraiva
, N. M. Fonseca Ferreira
and Antonio Valente
UNICEUMA, UTAD University, Vila Real, Portugal
Institute of Engineering of Coimbra, Polytechnic Institute of Coimbra, Portugal
Knowledge Research Group on Intelligent Engineering and Computing for Advanced Innovation and Development
(GECAD) of the Institute of Engineering, Polytechnic Institute of Porto, Portugal
INESC TEC (formerly INESC Porto), School of Science and Technology, UTAD University, Vila Real, Portugal
Filtering, Genetic Algorithm, Hybrid, DICOM.
The article refers to a new model of genetic algorithm. The method used has finality of optimize the filtering
of artifacts in DICOM images in two-steps.The first step is constituted by filterings with BM4D, 3d median
filter and ellipsoid filter. The second step is formed by the application of operators of simple mutations in
the previously recovered image, for that was used: intensity change, gaussian filter and mean filter. As a
result, a better performance filter was obtained and which provides an improvement in diagnosis, in diseases
assessment and in decisions making by the professional.
Digital images have been used for various purposes,
from just storing remembrances until accurate exams
in medicine (James and Dasarathy, 2014). Over the
years, the use and popularization of the digital image
made it possible the great increase of the volume of
images, just like it’s over by to make available new
advances and challenges in its use. As example there
is an introduction of images processing solutions in
industrial environment for visual inspection in envi-
ronments at risk for a physical integrity of employees
(Gonalves et al., 2014).
Despite of various technological advances, during
the captation process and posteriorly the transmission
of the digital images can acquire artifacts in innume-
rable ways.
Each artifacts filter model adapts differently to
each noise, thus forming its advantages and disadvan-
tages in relation to a determined type of noise.
As an example, there is the image in figure 1, here
white (Gaussian) noise has been added. This type of
noise is quite common in communications.
The white noise it comes from the stirring of the
electrons in the metallic conductors. Its level is in
function of the temperature, being evenly distributed
in all the frequencies of the spectrum.
The challenge of suppressing or attenuating has
provided the search of enhancement of techniques
Figure 1: Image increased Gaussian noise.
to reduce imperfections, in way to preserve impor-
tant information of the image such as corners, bor-
ders and textures. In the literature, the BM3D (Dabov
et al., 2006) noises attenuation technique was inten-
sively researched and tested on real problems such as
suppress artifacts in images of ultrasonography (Gan
et al., 2015), however, there is no solution available to
completely solve the problem.
Genetic Algorithms (GA) are metaheuristics ba-
sed on the theory of evolution of the species of Char-
les Darwin, where by natural selection the fittest in-
dividual tends to survive and reproduce descendants
(Barbosa, 2014). In this context, the individual is a
representation of the solution of the problem.
This work proposes and analyzes a 3D hybrid ge-
netic algorithm (HGA3D) for noise attenuation in DI-
COM medical images, integrates genetic algorithm to
Saraiva, A., Ferreira, N. and Valente, A.
New Bioinspired Filter of DICOM Images.
DOI: 10.5220/0006723802580265
In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 1: BIODEVICES, pages 258-265
ISBN: 978-989-758-277-6
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
some literature attenuation methods: BM4D (Maggi-
oni et al., 2013), 3D median filter (Jiang and Crookes,
2006) and ellipsoid (Yang et al., 2008).
Each individual of the population corresponds to
an image initially restored by one of these three met-
hods and the others individuals of the population are
created through the application of different mutation
operators in the initial image. HGA3D evolves the
entire population during a determined amount of time
and at the end the best individual is returned as the
restored image.
The hypothesis of the work is that the proposed
genetic algorithm model be able to find quality so-
lutions when compared to other methods present in
the literature for the smoothing of artifacts in DICOM
Thus, the article starts with the state of the art, a
review of evaluation methods used, explains the met-
hodology, it is made to exhibition of results and a dis-
cussion of the results.
Different solution methods for the noise attenuation
problem were proposed. The BM4D (Maggioni et al.,
2013) method for example, use sliding voxels cubes in
a first stage for the stacking of similar cubes, in a se-
cond phase each cube is filtered by a Wiener type filter
(Gonzalez and Woods, 2006). At the end, the image
is reconstructed using adaptive weights for each cube
added in its original position.
The proposed BM4D algorithm proved to be ef-
fective for gaussians noise and its performance is re-
markable in PSNR statistics generated during the aut-
hor’s tests.
Approaches based on the 3D median filter (Jiang
and Crookes, 2006) were also suggested. Widely ap-
plied in images processing, this filter is known for
its edge conservation nature. The filter demonstrated
in this paper uses a median calculation of a window
with sliding mask size NxNxN voxels. Results de-
monstrate its efficiency for removal of splashes in 3D
medical images, in addition to having low computati-
onal cost.
In addition to the previously cited methods, there
is also the ellipsoidal filter (Yang et al., 2008). In this
paper the author proposes a three-dimensional me-
dian filtering method and then an adaptive ellipsoidal
Gaussian filtering method for local preservation of the
image characteristics. According to the research the
filter is ideal in the meaning it reduces the magnitude
spatial of the high frequency in an image.
There are also methods based on genetic algo-
rithm of great relevance currently, as the hybrid ge-
netic algorithm for noises suppression in images pro-
posed in Paiva’s thesis (Paiva, 2016). It is proposed
the combination of a genetic algorithm with various
algorithms for the removal of artifacts from images
found in the literature.
The HGA was tested on images corrupted by a
gaussian additive noise with different levels of stan-
dard deviation. At the end of the work, the effective-
ness of the proposed method is demonstrated by me-
ans of statistical and visual data, showing better re-
sults in several cases in relation to literature methods.
In addition to all the methods already mentioned
above, a search was made in the literature for other ap-
proaches of great current impact in the research area,
for this was considered the google scholar metrics op-
tion. Initially a work was used available in IEEE
Transactions on Image Processing whose index is
the same.
In the work proposed in (Moore and Lopes, 1999)
is proposed a general methodology to create and op-
timize a wide group of algorithms for the destruction
of a mixed artifact between poisson noise and gaus-
sian noise. To remove of the artifact is demonstrated,
an algorithm denominated PURE-LET where in par-
ticular the best results are obtained. With the tests in
images and posteriorly the comparison between this
proposed method and other competing methods it is
verified the effectiveness of the restoration of several
textures present in the image.
In (Danielyan et al., 2012) is proposed an analy-
sis and synthesis for the family of BM3D algorithms
aiming to develop new iterative algorithms of deblu-
ring. The BM3D is a non-local modeling technique
based on adaptive models, it is divided into three steps
where initially, similar image blocks are collected in
groups, then the obtained groups spectra are filtered,
and lastly the filtered spectra are inverted providing
estimates of blocks that were returned to their original
positions and finally occurs the image recostruction.
Based on the researchs carried out and described,
a genetic algorithm based on BM4D, 3D median and
ellipsoid was developed.
The image filtering search aims to reduce the num-
ber of artifacts to represent an image, removing the
noises, as much as possible. The ideal is to get the
resulting image it’s close to the original image. One
of the ways to quantify is given by the measurement
New Bioinspired Filter of DICOM Images
of proximity with the Mean Square Error (MSE) can
be defined mathematically by equation 1 (Talbi et al.,
(I(x, y) K(x, y))
In this equation I represents the original image and
K the final image to be compared. The x and y are two
matrices of size MxN, respectively representing the
original x-channel and the y-channel to be compared
(after filtering).
Another way to compare the quality of the ima-
ges is the Peak Signal to Noise Ratio (PSNR) what is
usually a measure of image quality and can be repre-
sented by equation 2 (Fedorov and Rodyhin, 2016).
The PSNR ideal of comparison presents an optimum
value the higher its is your value.
PSNR = 10 log
= 20 log
In which, MAX represents the maximum possible
value of the pixel in the image and MSE is the value
resulting from equation 1.
The MSE may present problems when used to
compare similarity. The main from them is that large
distances between pixel intensities do not necessarily
mean that the content of the images be dramatically
different. It is important to note that a value of 0 for
MSE indicates perfect similarity. A value greater than
1 implies smaller similarity and will continue to grow
as the mean difference between pixel intensities in-
creases as well.
In order to remedy some of the problems asso-
ciated with MSE for image comparison, one has the
Structural Similarity Index (SSIM). The SSIM Is ob-
served by equation 3.
SSIM(x, y) =
+ c
+ c
+ µ
+ c
+ σ
+ c
In the equation 3 o µ represents the mean, the σ
symbolizes the standard deviation and σ
the cova-
riance. And c
with c
represent constants that avoid
the instability of values.
Unlike MSE, the SSIM value can range from -1 to
1, where 1 indicates perfect likeness.
The essence of SSIM is to model the perceived
change in the structural information of the image,
while the MSE is actually estimating the perceived
errors. There is a subtle difference between the two,
but the results can be great.
In addition, the SSIM is used to analyze small sub-
samples instead of the entire image as in MSE. The
parameters used are the mean of the pixel intensities,
the variance of the intensities, together with the cova-
riance. In this way, a more robust approach is obtai-
ned capable of explaining the changes in the structure
of the image, instead of just the perceived change.
For the quantitative comparison of the filtering
methods in this article, the objective metrics evalua-
tion methods MSE, PSNR and SSIM were used. Such
methods are known as full reference, because they
consider the original image as a reference.
These methods are applied over a DICOM image.
Being that, in this work MatLab software was used
to manipulation of the presented algorithms and the
visualization of the results.
The hybrid genetic algorithm (HGA) of this work is
based on the genetic algorithm (GA) proposed by To-
leto (Toledo et al., 2013) and in the method proposed
by Paiva (Paiva, 2016), where each individual of the
population is an image itself, represented by a set of
pixels whose values are integers in the range of 0 to
255. In a similar way to this method starts the propo-
sed algorithm, where a noisy image is used as input
for the method and the other individuals in the popu-
lation are created from applied mutation operators.
Based on the analysis and results demonstrated in
(Paiva, 2016), it was decided that the same parameters
already tested by the author were used as standard va-
lues in the model proposed here. In the step by step of
choosing the best parameters by the author is demon-
strated the effectiveness of each change in metric data
PSNR and SSIM, in addition a whole argumentation
of each result is provided.
According to Paiva (Paiva, 2016), during the
choice of tournament size the worst case of tourna-
ment size 3 tends to be better than the worst case of
the others. However the test of the different local se-
arch rates, although all the results were very close,
the value rate 0.6 was the one that reached the best
results compared to the others. The differences bet-
ween the results with different population sizes sho-
wed up clearer in their 2D approach, however, there is
a superiority of size 15 population where it obtained
good results in about 70 % of cases in both PSNR and
SSIM. Also is demonstrated the effectiveness of two
other parameters, beta whose best value was 1.5 and
execution time equal to 20 minutes.
The proposed algorithm combines the GA method
approach (Toledo et al., 2013), with noises smoothing
techniques in 3d images. In which, the pseudocode is
described in Algorithm 1.
BIODEVICES 2018 - 11th International Conference on Biomedical Electronics and Devices
Algorithm 1 : Hybrid Genetic Algorithm for Dicom
1: function HGA3D(Dicom path)
2: images ReadingPath(path)
3: Population createPopulation(images)
4: best
5: while elapsedTime < maxTime do
6: cont 0
7: while cont < maxIter do
8: IntermPop Population
9: for i 1 to Population.size do
10: ind1 Parents(Population)
11: ind2 Parents(Population)
12: ind3 Crossover(ind1,ind2)
13: if (Λ [0, 1]) LocalSearchRate
14: localSearch(ind3)
15: end if
16: IntermPop.append(ind3)
17: end for
18: Sort(IntermPop)
19: Population IntermPop[1..Popula-
20: if (best then
21: cont cont + 1
22: else
23: cont 0
24: end if
25: end while
26: end while
27: end function
The beginning of the HGA consists of creating
the initial population in two steps: first, the image
with noise is used as input for three methods of noi-
ses smoothing. Thus, at the end of the first stage, the
population has three individuals. Next are cited the
techniques used:
BM4D (Maggioni et al., 2013)
3D median filter (Jiang and Crookes, 2006)
Ellipsoid (Yang et al., 2008)
After the first stage, one of the outputs of these
techniques is chosen randomly. Then it is passed by a
mutation operator, also in a random way and changes
are realized in the image initially recovered by one of
the initial methods. As Mutation operators was used
three types:
Intensity change: is a linear operation that con-
sists of multiplying all the pixels of the image by
the same numerical factor.
Gaussian filter: the filter that has the effect of
smoothing the image artifact through a Gaussian
Average filter: the technique that allows the smoo-
thing of noises in images by means of calculating
the average of all the filters of a given vicinity for
each pixel of the original image.
At the end of this stage the resulting image is ad-
ded to the population. Then the mutation process is
repeated until the population reaches the chosen size.
Thus, a hybrid population is formed, constituted of
the output of the three methods of suppression of ini-
tial noises plus the images that went through the mu-
tation process.
The HGA runs for a fixed time, in which the popu-
lation continues to evolve while there is no changes in
the best individual to a maximum number of interacti-
ons. By reaching maximum number of interactions,
the entire population is restarted while only the best
individual is preserved. Posteriorly the population is
created again by the same process already mentioned.
An intermediate population twice the size of the
initial population is created during the process of evo-
lution formed by the current population plus the new
individuals generated. These new individuals are cre-
ated through operators crossover where the parents’
selection is made via the tournament. Shortly after
the parents were chosen, a new operator crossover is
randomly selected for the generation of a new indi-
vidual (son). For this are cited below the three types
available for the choice:
Operator of a line point: randomly choose a line
of pixels in the image, then all the pixels above
it will come from one parent and the other pixels
that are below it will come from the other parent.
Operator of a column point: approach similar to
the first, but the image is divided by a column rat-
her than a line.
Uniform Operator: each pixel of the image is cho-
sen randomly from one of the parents with 50 %
chance of the value chosen to be from either pa-
Once created, the new individual can still be sub-
mitted by a local search operator, a process that has
purpose improve the final quality of the solution by
means of transformations in the individual, case sa-
tisfied the condition that a real number generated by
the algorithm in the execution that is equal to a value
within the range of 0 to 1 in the algorithm be less than
the local search rate chosen, it will go through one of
the artifact suppression operators already mentioned
in the initial step: BM4D, Median Filter or Ellipsoid.
New Bioinspired Filter of DICOM Images
With all intermediate population completed, indi-
viduals are ordered by fitness, from the first individu-
als selected in a population of the size chosen at the
outset to form the main population of the HGA for
the next evolution step, where the algorithm verifies if
there are no changes in the best individual of the po-
pulation during a chosen number of executions of the
evolution. Case the best individual does not change
after a maximum number of iterations, so this popu-
lation is restarted. A flow diagram of the execution of
the algorithm is shown in figure 2.
Figure 2: Flowchart of the algorithm. Source: Own author.
At the outset they begin to form the main popu-
lation of the HGA3D for the next evolutionary step,
where the algorithm checks if there is no change in
the best individual of the population during a chosen
number of evolution executions. If the best individual
does not change after a maximum number of Iterati-
ons, then this population is restarted. A flowchart of
the algorithm execution is shown below.
In this chapter the results of the quantitative analysis
of the results will be presented through the evalua-
tion metrics. The comparison established is related to
other methods of filtering three-dimensional images:
3D median and ellipsoid.
The table 1 refers to the amount of MSE for each
image after the filtering, establishing values. In the
column 1 shows the percentage of image degradation,
in column 2 the noise mean, and columns 3, 4 and
5 the respective MSE values obtained for the filters
of the median 3d, ellipsoid and the proposed filtering
method HGA3D.
Table 1: Evaluation of the result through MSE.
Gaussian additive noise
Noise (MSE) Median Ellipsoid HGA3D
10% (0.0062) 0.0102 0.0102 0.0747
20% (0.0268) 0.0112 0.0189 0.0025
30% (0.0307) 0.0120 0.0115 0.0747
40% (0.0336) 0.0145 0.0073 0.0165
Average 0.0199 0.0119 0.0187
Table 2: Evaluation of the result through PSNR.
Gaussian additive noise
Noise (PSNR) Median Ellipsoid HGA3D
10% (65.61) 72.58 72.93 73.85
20% (61.22) 69.83 69.93 72.76
30% (58.00) 67.07 67.68 71.24
40% (56.06) 65.16 65.69 69.88
Average 68.66 69.05 71.93
Table 3: Evaluation of the result through SSIM.
Gaussian additive noise
Noise (SSIM) Median Ellipsoid HGA3D
10% (0.2651) 0.4810 0.5053 0.8600
20% (0.1076) 0.2946 0.2428 0.8004
30% (0.0955) 0.2319 0.2456 0.6303
40% (0.0887) 0.1826 0.2357 0.5374
Average 0.2875 0.3773 0.7070
In the tables 2 and 3 are related to the qualitative
analysis of PSNR and SSIM. Featuring a design simi-
lar to table 2.
The first analysis was done by the MSE metric,
presented in only one case the filter type HGA3D as
best. However, taking into account the relevance of
this type of meter the best results are those whose
values are the smallest, on the other hand it has the
contestable confidence level of this metric, making it
present the need of comparison with new forms.
In the table 2 is shown an evaluation using a better
metric, this metric demonstrates in numerical data an
approximation of the human perception of the quality
of reconstruction, where not necessarily, but in most
cases the larger PSNR values represent a better recon-
struction of the image.
When comparing the resulting values demonstra-
ted below it is clear the superiority of the data re-
sulting from the proposed method. With efficiency
in 100 % of the cases tested in this approach, it is de-
monstrated in the table that in only one case, the va-
BIODEVICES 2018 - 11th International Conference on Biomedical Electronics and Devices
lues were close to the genetic algorithm model. After
is realized the difference in values resulting from the
methods stay distant, in addition, it is also remarka-
ble the difference between the average of the HGA3D
with the means of the other competing methods.
The table 3 demonstrates the analysis using the
most accurate evaluative metric currently used, SSIM.
This metric improves traditional methods, that show
inconsistent with human visual perception.
The results demonstrated here by the tables prove
that the combination of several artifact removal
techniques show up very favorable in most images, in
addition, the few amount of limitations of the HGA3D
provides a multitude of options for changing parame-
ters and providing improvements in the final image.
As a visual example of the obtained results, it is
shown in figures 3, 4, 5 and 6. In each figure, four
images are observed, one referring to the slice added
with noise and another three are results of the me-
dium, ellipsoid and HGA3D filtering. In figure 3 it
has been the image corrupted with gaussian artifact
and a standard deviation of 10%. In the other figures
differ in the standard deviation of 20%, 30% and 40%
Figure 3: Image corrupted with white additive Gaussian ar-
tifact and standard deviation = 10%.
In figure 3 was observed that when applying the
noise with low deviation, 20 % , it is not visually per-
ceptible the difference of HGA3D in relation to the
others. However, in Figures 4, 5 and 6 the difference
between the proposed filter and the other two filters
that serve as a basis for verifying the quality.
Figure 4: Image corrupted with white additive Gaussian ar-
tifact and standard deviation = 20%.
Figure 5: Image corrupted with white additive Gaussian ar-
tifact and standard deviation = 30%.
With the introduction of the filter it is evident that
there is an improvement of resolution in both images,
making them more interesting for the observation of
the image.
In the table 1, it was observed that in three items
the ellipsoid obtained the most efficient filtration con-
dition. The proposed method presented only a signifi-
New Bioinspired Filter of DICOM Images
Figure 6: Image corrupted with white additive Gaussian ar-
tifact and standard deviation = 40%.
cant result in the percentage of image degradation of
20 % the best performance. However, the MSE may
exhibit similarity failure.
Thus, the efficiency of the HGA3D method is de-
monstrated when compared to the others exposed in
tables 2 and 3, using PSNR and SSIM. Demonstra-
ting the final image after filtering that most closely
resembles the original image and provides an incre-
ase in quality.
There are several techniques for developing DICOM
image filtering, this study applies the hybrid method
of genetic algorithm, in which the method obtains op-
timal filtering and minimizes artifacts.
The efficiency of the model adopted as a filter is
the result of the architecture that is found distributed
in a selective and evolutionary way in two stages. The
first stage consists of the BM4D filtering, the 3d me-
dian filter and the ellipsoid filter. The second stage is
formed by the application of operators of simple mu-
tations in the previously recovered image, for that was
used: intensity change, gaussian filter and average fil-
As comparison the MSE, PSNR and SSIM was
used to estimate the filtering efficiency of the restored
images. It was observed experimentally that the adop-
ted filter is efficient and robust presenting indexes bet-
ter than the others in the PSNR and SSIM.
With the study of the HGA3D can generate more
advances and minimize the artifacts, resulting in a
better performance in the system. The disadvantage
is the limitations of techniques for the random values,
that make it difficult the optimal value defined in the
In order to apply more efficient methods of recon-
struction of DICOM images, it is intended in future
works to approach the methods with the application
of new filters to increase efficiency. As an example
one has the artificial intelligence in one of the stages.
The elaboration of this work would not have been pos-
sible without the collaboration of the Research Group
on Intelligent Engineering and Computing for Advan-
ced Innovation and Development (GECAD) of the In-
stitute of Engineering, Polytechnic Institute of Porto,
Barbosa, R. C. (2014). Metaheuristic algorithm genetic
application in optimization of distribution of delivery
routes physics products in fortaleza county. In UFC.
Dabov, K., Foi, A., Katkovnik, V., and Egiazarian, K.
(2006). Image denoising with block-matching and 3d
filtering. In Electronic Imaging. International Society
for Optics and Photonics, p. 606414-606414-12.
Danielyan, A., Katkovnik, V., and Egiazarian, K. (2012).
Bm3d frames and variational image deblurring. vo-
lume 21, pages 1715–1728.
Fedorov, O. and Rodyhin, M. (2016). A referenceless psnr
estimator of compressed jpeg images. In 2016 26th In-
ternational Conference Radioelektronika (RADIOE-
LEKTRONIKA), pages 227–230.
Gan, Y., Angelini, E., Laine, A., and Hendon, C. (2015).
Bm3d-based ultrasound image denoising via brushlet
thresholding. In 2015 IEEE 12th International Sympo-
sium on Biomedical Imaging (ISBI), pages 667–670.
Gonalves, M., Rodrigue, J., Rodrigues, C., and Rocha, A.
(2014). Desenvolvimento de uma solucao de proces-
samento de imagem em ambiente industrial. In ENE-
GICOIMBRA2014: inovacao, tecnologia e excelen-
cia: atas do 3 Encontro Nacional de Engenharia e
Gestao Industrial, p. 28-29.
Gonzalez, R. C. and Woods, R. E. (2006). Digital Image
Processing. Upper Saddle River, NJ, USA: Prentice-
Hall, 3rd edition.
James, A. P. and Dasarathy, B. V. (2014). Medical image
fusion: A survey of the state of the art. In Information
Fusion, v. 19, p. 4-19.
Jiang, M. and Crookes, D. (2006). High-performance
3d median filter architecture for medical image
despeckling. volume 42, pages 1379–1380.
BIODEVICES 2018 - 11th International Conference on Biomedical Electronics and Devices
Maggioni, M., Katkovnik, V., Egiazarian, K., and Foi, A.
(2013). Nonlocal transform-domain filter for volume-
tric data denoising and reconstruction. volume 22, pa-
ges 119–133.
Moore, R. and Lopes, J. (1999). Paper templates. In TEM-
PLATE’06, 1st International Conference on Template
Production. SCITEPRESS.
Paiva, J. L. (2016). A hybrid genetic algorithm for image
denoising. In Tese de Doutorado. USP.
Talbi, M., Ftima, S. B., and Cherif, A. (2015). Image wa-
termarking using data compression. In 2015 World
Symposium on Computer Networks and Information
Security (WSCNIS), pages 1–9.
Toledo, C. F. M., Oliveira, L. D., da Silva, R. D., and Pe-
drini, H. (2013). Image denoising based on genetic
algorithm. In 2013 IEEE Congress on Evolutionary
Computation, pages 1294–1301.
Yang, F., Zuo, W. M., Wang, K. Q., and Zhang, H. (2008).
3d cardiac mri data visualization based on volume
data preprocessing and transfer function design. In
2008 Computers in Cardiology, pages 717–720.
New Bioinspired Filter of DICOM Images