Você está na página 1de 7

Particle Swarm Optimization-based

Functional Link Artificial Neural


Network for Medical Image Denoising

Manish Kumar and Sudhansu Kumar Mishra

Abstract In this paper, a new computationally fast and efficient adaptive digital
image filter has been proposed for denoising of digital medical image corrupted
with additive white Gaussian noise. A particle swarm optimization-based func-
tional link artificial neural network (FLANN) has been applied for this interesting
and challenging problem. The three others competitive networks based on artifi-
cial neural network such as multilayer perceptron (MLP), direct linear artificial
feed-through neural network (DLFANN), and LMS-based FLANN have also been
applied for this purpose. The quantitative analysis of the proposed algorithm has
been carried out by taking the peak signal to noise ratio (PSNR) and mean square
error (MSE) as two parameters. Experimental results demonstrate the efficiency
and effectiveness of the proposed algorithm.

KeywordsMedical imageFunctional link artificial neural networkParticle


swarm optimization Peak signal to noise ratio

1Introduction

The medical image may be corrupted during telemedicine due to the noisy com-
munication channel. The presence of noise degrades the medical image q uality
affecting many of its inherent features that alleviates the problem of disease

Fifth International Conference on Computational Vision and Robotic, 10th August, Bhubaneswar,
India.

M. Kumar(*) S.K. Mishra


Department of Electrical and Electronic Engineering, Birla Institute of Technology,
Mesra, Ranchi, India
e-mail: manish.guptasssss007@gmail.com
S.K. Mishra
e-mail: sudhansu.nit@gmail.com

Springer India 2015 105


I.K. Sethi (ed.), Computational Vision and Robotics, Advances in Intelligent
Systems and Computing 332, DOI 10.1007/978-81-322-2196-8_13
106 M. Kumar and S.K. Mishra

diagnosis, recognition, etc. Hence, to eliminate this problem, medical image


should be free from different noises. This important job of medical image denois-
ing can be done by different fixed filters such as mean, median, and rank order
mean. However, all these filters are suitable for fixed channel where system
parameters are not changing with time. But, in real practice, most of the channels
are time varying. To circumvent different limitations of fixed filters, adaptive fil-
ters are designed that adapt themselves to the changing conditions.
ANN has proved to be a potential network to perform the tasks in a highly
nonlinear environment. MLP is one of the variants of ANN whose weights are
trained by applying back propagation (BP) algorithm [1]. In DLFANN struc-
ture incorporates, a set of linear terms is introduced in a network for modeling
both linear and nonlinear systems simultaneously [2]. The functional link arti-
ficial network (FLANN) was proposed by Pao [3]. In the structure of FLANN,
hidden layer is absent and whole structure can be represented as a single layer
ANN. In this work, it is shown that the FLANN can be used for function pattern
classification, quicker rate of convergence, etc. The FLANN model has lesser
computational complexity than both MLP and DLFANN networks because of no
hidden layer. It is quite capable of forming arbitrarily complex decision regions
by generating nonlinear decision boundaries [4]. A FLANN using the Chebyshev
and exponential function expansion has been proposed and reported to denoise
the image [5, 6]. In this paper, Chebyshev expansion is used for the functional
expansion of FLANN, and the particle swarm optimization (PSO) is used for
selecting extended inputs and weights. The primary purpose of this paper is to
highlight the effectiveness of the proposed PSO-based single layer ANN struc-
ture to solve the problem of Gaussian noise cancelation from the medical images.
Applications of some other adaptive filters for medical image denoising have also
been reported [711].
The paper is organized as follows: Different structures of the ANN filter and
their implementation for artifact cancelation in image data have been discussed in
Sect.2. Section3 presents the simulation studies of proposed methods followed by
the conclusion in Sect.4.

2Structure of ANN Filters

2.1Multilayer Perceptron

The multilayer perceptron is one of the feed-forward networks, which have one
or more hidden layers. All the nodes in all layers (except the input layer) of the
MLP contain a nonlinear function. The weighted sum of outputs of a lower layer
is passed through the nonlinear function of a node in the upper layer to produce its
output. The mean square error (MSE) is used as cost function and the back propa-
gation (BP) algorithm attempts to minimize the cost function by updating all the
weights of the network [2] (Fig.1).
Particle Swarm Optimization-based Functional Link 107

Fig.1Multilayer neural network

2.2Direct Linear Feed-through Neural Network

The process of noise cancelation may have both linear and/or nonlinear behav-
ior over its working range. Hence, it appears more appropriate to have a network
structure that is capable of handling both linearity and nonlinearity. The DLFANN
configuration which combines conventional NN architecture with a set of linear
terms, to produce a network, can handle both linear and nonlinear behavior at the
same time [5] (Fig.2).
In this network, neurons associated with input layer are connected directly with
neuron of hidden as well as output layer. The arrangements are shown by thick
and thin solid lines, respectively. These direct interconnections between input
and output neurons are accountable for handling linear terms. The computational
complexity of the network is increased as weights corresponding to direct inter-
connection between the input and output neurons are added. This increases the
complexity of network and the training of network also take more CPU time to
update.
108 M. Kumar and S.K. Mishra

Fig.2The DLFANN
network

Input Layer Hidden Layer Output Layer

2.3Particle Swarm Optimization-based Functional Link


Artificial Neural Network

The FLANN is one of the variants of ANN which has single layer architecture
and has expanded inputs based on some functional expansions such as trigono-
metric expansion, Chebyshev, and power series expansion. In contrast to linear
weighting of the input pattern produced by the linear links of an MLP, the func-
tional link acts on the entire pattern by generating a set of linearly independent
functions [12]. The architecture becomes simple and the need of back propaga-
tion is eliminated. In this paper, Chebyshev polynomial expansion is used and the
structure C-FLANN is shown in Fig.3. These higher Chebyshev polynomials for
1<x<1 may be generated using the recursive formula given by
Tn+1 = 2Tn (x) Tn1 (x) (1)
The first few Chebyshev polynomials are given by

T0 (x) = 1
T1 (x) = x
T2 (x) = 2x 2 1 (2)
3
T3 (x) = 4x 3x
T4 (x) = 8x 4 8x + 1

In this paper, some of the expanded inputs of FLANN model are chosen by apply-
ing PSO. The weights of the model are chosen by applying PSO. The idea of par-
ticle swarm optimization was first introduced by Kennedy and Eberhart in [13].
This algorithm is formulated by observing the social behavior of school of fish
or focus of birds. Due to the robustness, simplicity and low computational com-
plexity PSO is applied to many other interesting fields of research [4, 14]. In
PSO, each solution is considered by a particle and ith particle (Xid), where d is
the dimension of the search space. The ith particle of the swarm population has its
Particle Swarm Optimization-based Functional Link 109

+1
X1

T1

Functional Expansion
X1
T2
X2
:
y
: (.)
:
X2 -

T1 Adaptive e
X9 Algorithm
T2

+
d

Fig.3Chebyshev functional link multilayer perceptron architecture

best position (Pi) that generates the highest fitness value. The global position (Pg)
is the best particle that gives best fitness value in the entire population. Vi is the
current velocity of ith particle [15]. The position and velocity can be expressed in
the next iteration are mathematically expressed as
 
Vi (d) = wVi (d) + Ci (d) rand (Pi (d) Xi (d)) + C2 rand Pg (d) Xi (d)
(3)

Xi (d) = Xi (d) + Vi (d) (4)

3Simulation Studies

The MATLAB environment has been implemented for simulation studies.


Simulation analysis was carried out with several iterations and to compare perfor-
mance of various variants of MLPs, DLFANN, and FLANN for denoising of med-
ical image. The proposed method is carried out with X-ray image of chest which
degraded by additive white Gaussian noise. Gaussian noise having a continuous
energy spectrum of all frequencies and for simulation mean, variance value, i.e.,
0 and 0.01 has been selected respectively. In our simulation, we set the structure
of MLP, DLFANN as {9-4-1} and FLANN as {9-1}. The order of the polynomial
used to obtain expanded inputs is gradually increased so that a minimum number
of inputs are used and complexity of the neural network is minimized.
The learning rate of MLP and DLFANN is set to 0.03 and that of FLANN is
set to 0.1. The number of iterations was set to be 3,000 for all models. The BP
learning algorithm has been used for MLP. In DLFANN, the weight of a direct
link between input layer and output layer is trained by LMS algorithm and others
110 M. Kumar and S.K. Mishra

Original Image Noisy Image Denoise Image

Fig.4X-ray image of chest

Table1Result for filters Image Noisy MLP DLFANN FLANN PSO-


interms of PSNR value FLANN
Image1 20.18 25.62 25.66 26.5 26.6
Image2 20.34 25.74 25.54 26.2 26.3
Image3 20.13 25.17 25.32 26.1 26.1

are trained by using the BP algorithm. The PSO technique is used for training
of FLANN. The training targets were normalized within value [0, 1]. This pro-
cess continues iteratively until full pattern of the image get processed and whole
process continues for 100 times to find MSE as well as PSNR value. The size of
X-ray image has been taken 256256 for the training and testing (Fig.4).

3.1Peak Signal to Noise Ratio

In this work, computer simulation is carried out to compare the PSNR value of fil-
tered images obtained from these adaptive models. In this case, the Gaussian noise
having mean 0 and variance 0.01 is taken. Let X(m, n), Y(m, n) is the noise-free
image, noisy image, respectively. The size of the image of MN pixels, where
m=1, 2, 3M is the number of rows and n=1, 2, 3N is the number of col-
umns. Then the formula for PSNR and MSE is shown as (Table1)
M N 2  
m=1 n=1 ((m, n) Y (m, n)) 1
MSE = and PSNR = 10. log10 (5)
MN MSE

4Conclusion

The PSO-based FLANN model is proposed for denoising of for medical image
corrupted by Gaussian noise. To solve this challenging problem, three other vari-
ants of ANN such as MLP, DLFANN, and LMS-based FLANN have also been
Particle Swarm Optimization-based Functional Link 111

successfully applied. The comparison of performance of these networks includes


evaluation of two performance metrics such as MSE and PSNR. It can be con-
cluded that the proposed PSO-FLANN model reduces the computational complex-
ity with improved performance.

References

1. Haykin, S.: Neural Networks. Maxwell Macmillan, Ottawa (1994)


2. Dash, P.P., Patra, D.: Evolutionary neural network for noise cancellation in image data. Int.
J. Comput. Vision Robot. 2(3), 206217 (2011)
3. Pao, Y.H.: Adaptive Pattern Recognition and Neural Network. Addison-Wesley Reading, MA
(1989)
4. Mishra, S.K., Panda, G., Meher, S., Majhi, R., Singh, M.: Portfolio management assessment
by four multiobjective optimization algorithm. In: IEEE International Conference (RAICS-
2011), pp. 326331. Trivandrum, India, 2325 Sept 2011
5. Mishra, S.K., Panda, G., Meher, S.: Chebyshev functional link artificial neural networks
for denoising of image corrupted by salt and pepper noise. Int. J. Recent Trends Eng. 1(1),
413417 (2009) (Academy publisher, Finland. ISSN:1797-9617)
6. Mishra, S.K., Panda, G., Meher, S., Sahoo, A.K.: Exponential functional link artificial neural
networks for denoising of image corrupted by gaussian noise. In: IEEE Computer Society
International Conference on Advanced Computer Control, pp. 355359 (2009)
7. Muller, H., Michoux, N., Bandon, D.: A review of content-based image retrieval systems in
medical applications-clinical benefits and future directions. Int. J. Med. Inform. 73(1), 123
(2004)
8. Sharif, M., Jaffar, M.A., Mahmood, M.T.: Optimal composite morphological supervised fil-
ter for denoising using genetic programming: application to medical resonance image. In:
Elsevier Conference on Engineering Applications of Artificial Intelligence, pp. 7889 (2014)
9. Suzuki, K., Horiba, I., Sugie, N.: Efficient approximation of neural filters for removing quan-
tum noise from images. IEEE Trans. Signal Process. 50(7), 17871799 (2002)
10. Jiang, J., Trundle, P., Ren, J.: Medical image analysis with a artificial neural network.
Elsevier J. Comput. Med. Imaging Graph. 34, 617631 (2010)
11. Mohamed, S.M.R., Jayanthi, R.B.: Speckle noise removal in ultrasound images using parti-
cle swarm optimization technique. In: IEEE-International Conference on Recent Trends in
Information Technology Chennai, pp. 926931 (2011)
12. Majhi, B., Sa, P.K.: FLANN based adaptive threshold selection for detection of impulse noise
in image. Elsevier Int. J. Electron. Commun. 61, 478484 (2007)
13. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of the Fourth IEEE
International Conference on Neural Networks, pp. 19421948. IEEE Service Centre, Perth,
Australia (1995)
14. Mishra, S.K., Panda, G., Manjhi, R.: A comparative performance assessment of a set of
multiobjective algorithm for constrained portfolio assets selection. Elsevier Swarm Evol.
Comput. 16, 3851 (2014)
15. Jansi, S., Subasini, P.: Particle swarm optimisation based total variation filter for image
denoising. J. Theor. Appl. Inf. Technol. 57, 169173 (2013)

Você também pode gostar