Scholarly article on topic 'Improved PSO Based Multi-level Thresholding for Cancer Infected Breast Thermal Images Using Otsu'

Improved PSO Based Multi-level Thresholding for Cancer Infected Breast Thermal Images Using Otsu Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{Thermogram / Otsu / "Improved PSO" / segmentation / PSNR / "CPU time."}

Abstract of research paper on Materials engineering, author of scientific article — N. Sri Madhava Raja, S. Arockia Sukanya, Y. Nikita

Abstract In this paper, an Improved Particle Swarm Optimization (IPSO) algorithm based bi-level and multi-level thresholding is proposed to segment the cancer infected breast thermal images using Otsu's function. In the proposed image segmentation work, histogram of the image is analyzed and the optimal thresholds are attained by maximizing Otsu's between class variance function. The performance of IPSO based segmentation process is demonstrated by consideringthermograms, being compared with state-of-the-art alternatives, such as Particle Swarm Optimization (PSO) and Darwinian PSO (DPSO). The proposed image segmentation procedure is directly implemented on RGB images. The performance assessment between algorithms is carried out using parameters, such as objective function, PSNR, SSIM and CPU time. The results confirm that, IPSO shows an overall enhancement over the alternatives, illustrating a tradeoff betweenCPU time and performance measure values.

Academic research paper on topic "Improved PSO Based Multi-level Thresholding for Cancer Infected Breast Thermal Images Using Otsu"

(8)

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 48 (2015) 524 - 529

International Conference on Intelligent Computing, Communication & Convergence

(ICCC-2015)

Conference Organized by Interscience Institute of Management and Technology,

Bhubaneswar, Odisha, India

Improved PSO based Multi-Level Thresholding for Cancer Infected

Breast Thermal Images using Otsu

N. Sri Madhava Rajaa*, S. Arockia Sukanyab, Y. Nikitaa

aDepartment of EIE, St. Joseph's College of Engineering,Chennai 600 119, Tamil Nadu, India. bDepartment of CIE, St. Joseph's College of Engineering, Chennai 600 119, Tamil Nadu, India.

Abstract

In this paper, an Improved Particle Swarm Optimization (IPSO) algorithm based bi-level and multi-level thresholding is proposed to segment the cancer infected breast thermal images using Otsu's function. In the proposed image segmentation work, histogram of the image is analyzed and the optimal thresholds are attained by maximizing Otsu's between class variance function. The performance of IPSO based segmentation process is demonstrated by consideringthermograms, being compared with state-of-the-art alternatives, such as Particle Swarm Optimization (PSO) and Darwinian PSO (DPSO). The proposed image segmentation procedure is directly implemented on RGB images. The performance assessment between algorithms is carried out using parameters, such as objective function, PSNR, SSIM and CPU time. The results confirm that, IPSO shows an overall enhancement over the alternatives, illustrating a tradeoff betweenCPU time and performance measure values.

© 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review underresponsibilityofscientificcommittee oflnternational Conference onComputer, Communication and Convergence (ICCC 2015)

Keywords: Thermogram; Otsu; Improved PSO; segmentation; PSNR; CPU time.

1. Introduction

Breast cancer refers to unreliable, uncontrolled growth and explosion of the cells that originate in the breast tissue. A

* Corresponding author. Tel.: +91-9790936295 E-mail address:nsrimadhavaraja@stjosephs.ac.in

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of scientific committee of International Conference on Computer, Communication and Convergence (ICCC 2015) doi:10.1016/j.procs.2015.04.130

recent study made by United States National Cancer Institute states that one among eight women gets a chance to be diagnosed with breast cancer [1].

Thermographyis among the various modalities used for early detection of breast cancer, but due to limitation in interpretation of thermal images, it has gone unnoticed. Thermograms have the ability to detect changes in breast tissue which records the temperature distribution of that body in accordance with infrared variations using high speed thermal cameras [17].Recent advancements in image processing technology, paves way for using thermal imaging for diagnostic purposes, which is a non-invasive, cost effective method and less harmful. A detailed description about the breast thermal image can be found in the literature [3,6,8].

In recent years, heuristic algorithm based bi-level and multi-level image segmentation technique is widely adopted by the researchers for RGB and gray scale images [4,5,10,11,13,14]. In the proposed work, an Improved Particle Swarm Optimization (IPSO) algorithm based bi-level and multi-level thresholding is used to segment the RGB thermal images using Otsu's function.The performance of IPSO is demonstrated by considering a, being compared with state-of-the-art alternatives, such as Particle Swarm Optimization (PSO) and Darwinian PSO (DPSO). Further,the performance assessment between algorithms is carried out using parameters, such as objective function, PSNR, SSIM and CPU time.

2. Methodology

Otsu is one of the commonly used methods in the segmentation of Gray scale and RGB images [9]. This method provides the best possible thresholds by maximizing the between class variance function. This procedure is defined as follows:

For a given RGB image, let there is L intensity levels in the range {0, 1,2,..., L-1}. Then, the probability distribution Pi can be defined as:

hc L-1

f-N &f-1 (■)

where*- specific intensity level in the range {0<i <L-l}, C = {R,G,B}, N = total number of pixels in the image, and hy = number of pixels for the corresponding intensity level I in component C.

A detailed description about Otsu can be found in [4,5,9].

During the RGB segmentation process, the m - level thresholding is reduced to an optimization problem to search for tj , that maximize the objective function of each image component C can be defined as:

qf= max (2)

l<tf<,...,L-l * J

Along with the above objective function, the common image quality measures, such as the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Matrix (SSIM) are also considered and its mathematical expression is given below:

PSNRfx,^ = 20 log 10

; dB (3)

SSIM(x,y) = (^Z^lli^Sll (4)

(jux +JUy -Cj)(ctx2 + CT2 +C2)

wherex and y are original and segmented images; ¡uxand ¡uyare the average values, ox2and oy2 are the variance, oxyis the covariance, and C1= (k1L)2 and C2= (k1L)2 stabilize the division with weak denominator, with L = 256, fy = 0.01,

and k2 = 0.03 [11-13].

The performance of the heuristic algorithms is assessed using the CPU time. In this work, each image is examined for m = 2,3,4 and 5.

3. Particle Swarm Optimization Algorithm

PSO was initially proposed by Kennedy and Eberhart in 1995[7]. Due to its optimization accuracy, it is widely adopted by most of the researchers to solve variety of engineering optimization problems [3,12]. In recent years, PSO based approaches are widely considered in image segmentation application [4,5,10,14]. In this work, traditional PSO, Darwinian PSO and Improved PSO algorithms are adopted to segment the considered RGB image data set. The basic code for PSO based segmentation is available in [18].

3.1 Traditional PSO

A detailed explanation of the traditional PSO can be found in the literature. PSO has two mathematical relations, such as velocity update (eqn. 5) and position update (eqn. 6).

j-/¡t+l) = wtyt +CjRjfpbest-Sj )+C2R2(gbest-Sti) (5)

s(t+1> ^Sl+V^ (6)

whereC7= cognitive parameter; C2= social parameter; R and R2 = random numbers [0,1]; W = inertia weight.

3.2 Darwinian PSO

The Darwinian PSO was initially proposed by Tillett et al. (2005) [15]. A detailed description of the DPSO is existing in the literature [2]. Recently, this algorithm was widely adopted by most of the researchers for solving the multilevel image segmentation problems [4,5].

3.3 Improved PSO

The Improved PSO was initially proposed by Chang and Shih [16]. In this work, they modified the velocity update equation as presented below:

Vp+V = wtyt +CjRj(pbest-S\ )+C2R2(gbest-S\ )+C3R3(ibest-S\) (7)

Compared with the traditional PSO, it has additional parameters, such as C3 and R3. The position update is similar to the traditional PSO algorithm [12].

3.4 Implementation

In the proposed method, the considered algorithms continuously explore the histogram of the image until the objective is satisfied. In the proposed work, maximization of Otsu's between-class variance function is chosen as the objective function (Jmax). The initial algorithm parameters are assigned based on Table 1.

Table 1. Initial parameters of PSO algorithms

Parameter IPSO PSO DPSO

Number of Iterations 300 300 300

Population 50 50 50

C1 1.8 1.8 1.8

C2 1.2 1.2 1.2

C3 1.8 - -

W 0.75 0.75 -

Xmax 255 255 255

Xmin 0 0 0

Min population - - 15

Max population - - 50

Number of swarms - - 4 Min swarms - - 2 Max swarms - - 6 Stagnancy - - 20 Stopping criteria_¿max_¿max_■¿max.

4. Results and Discussions

Otsu guided, PSO algorithm based multi-level thresholding have been tested on achosen RGB image data set. The simulation work is executed on a work station with Intel Core i3 2.2 GHz CPU with 2 GB of RAM and equipped with Matlab R2010a software.

Initially, the proposed method is tested on the 481 x 321 sized standard test image obtained from the database [19]. Fig 1 (a) depicts the considered image, the RGB and gray scale histograms are show in Fig 1(b) and (c) respectively.From Fig 1 (b) and (c), one can observe that, the pixel distribution of the RGB image is approximately similar. The considered image is segmented using the IPSO, PSO and DPSO algorithms. This image segmentation process is repeated 10 times for each 'm' and the mean value is chosen as the optimized result.

(a) Snake image

(b) RGB histogram Fig 1. 481 x 321 sized RGB test image

0 50 100 150 200 250

(c) Gray histogram

(a) BTI1 (b) BTI2 (c) BTI3

Fig 2. 543 x 345 sized breast thermal image data set

Table 2.Segmented image data set with IPSO

Table 3 Objective values and optimal thresholds obtained with PSO algorithms for m = 2,3,4,5

m Objective function Optimal threshold

IPSO PSO DPSO IPSO PSO DPSO

2 138.74 148.65 155.72 116, 137 114, 134 114, 137

Snake 3 168.32 167.04 169.82 107,128,144 105,126,146 104,130,143

4 178.94 177.46 180.95 102, 133,140,151 101, 136,144,150 100, 133,138,152

5 187.77 182.39 193.44 98,129,140,148,155 96,125,141,147,154 96,127,142,147,157

2 1889.17 1794.64 1869.26 126,177 123,179 125,180

Ö 3 1905.66 1862.22 1912.04 91,141,182 90,138,186 91,142,184

pq 4 1941.30 1916.07 1944.62 58,108,155,192 55,103,158,190 54,106,154,191

5 1982.42 1962.46 1981.88 50,74,133,171,204 51,76,130,174,206 48,78,134,173,202

2 1202.53 1213.92 1225.07 105,154 101,152 104,160

3 1271.84 1263.79 1284.66 88,143, 181 86,144, 183 84,149, 180

pq 4 1282.41 1281.66 1293.18 65,93.136,207 66,90.137,208 62,95.134,202

5 1306.47 1298.05 1311.25 42,71,135,188,214 44,73,138,184,216 42,76,134,182,212

2 1302.61 1286.92 1312.06 107, 162 104, 166 101, 168

m 3 1337.28 1328.24 1334.93 92, 155, 181 90, 152, 184 95, 156, 188

pq 4 1391.24 1396.03 1402.11 72,134,169, 196 70,136,166, 199 72,133,165, 203

5 1416.02 1422.02 1418.62 61,118,138,185,208 60,121,135,187,211 58,126,131,182,213

Table 4. Performance measure values obtained with PSO algorithms

m PSNR (dB) SSIM CPU time (min)

IPSO PSO DPSO IPSO PSO DPSO IPSO PSO DPSO

2 24.86 23.47 23.11 0.638 0.652 0.627 0.493 0.583 0.447

<D 3 29.16 27.81 28.60 0.694 0.704 0.713 1.283 1.308 1.003

e 4 30.42 29.05 31.78 0.728 0.719 0.742 1.542 1.400 1.674

5 31.18 30.33 31.92 0.762 0.744 0.758 1.584 1.592 1.516

2 28.16 29.04 29.82 0.672 0.704 0.698 0.442 0.477 0.502

£ 3 30.75 30.68 30.27 0.711 0.718 0.714 0.583 0.564 1.035

pq 4 31.44 31.82 30.90 0.725 0.744 0.735 1.046 1.003 1.082

5 33.01 32.60 32.16 0.770 0.792 0.781 1.392 1.228 1.402

2 30.04 29.04 29.82 0.658 0.690 0.688 0.594 1.004 0.533

(N 3 30.52 30.82 31.16 0.683 0.704 0.694 1.046 1.068 1.115

pq 4 31.75 31.42 31.99 0.718 0.711 0.703 1.283 1.305 1.416

5 31.88 32.15 32.41 0.726 0.718 0.715 1.490 1.552 1.494

2 28.01 29.62 28.62 0.704 0.694 0.708 0.521 0.552 0.507

m 3 28.66 29.81 29.24 0.718 0.702 0.713 1.060 1.224 1.168

pq 4 29.37 30.44 29.86 0.720 0.711 0.718 1.184 1.307 1.224

5 29.92 31.06 30.27 0.728 0.716 0.721 1.296 1.486 1.307

Later, the segmentation procedure is tested on 543 x 345 sized Breast Thermal Images (BTI) depicted in Fig 2 [20]. Table 2 shows the segmented images using the IPSO algorithm for m = 2, 3, 4, 5 and the corresponding maximal objective values, optimal thresholds and the performance measures are presented in Table 3 and 4 respectively. From Table 2, it can be observed that, the segmentation procedure enhances the cancer region compared with the original test images shown in Fig 2.

From Table 3, it is noted that, the objective function obtained with the DPSO algorithms is better than PSO and IPSO. For most of the test images, the PSNR, SSIM and CPU time obtained with the considered IPSO algorithm is better than the PSO.

From Table 1, it can be observed that, the initial algorithm parameters to be assigned in the case of DPSO are more compared with the PSO and IPSO. From this work, it is confirmed that, IPSO algorithm can offer a better result compared with the PSO and satisfactory result compared with the complex DPSO algorithm.

5. Conclusion

In this paper, histogram assisted bi-level and multi-level image thresholding problem is addressed for RGB images using IPSO, PSO and DPSO algorithms. The simulation study is carried using Matlab R2010a software. Maximization of Otsu's between class variance is chosen as the cost function. To evaluate the performance of proposed method, one standard RGB test image and three Breast Thermal Image (BTI) are analyzed. From this study, it is confirmed that the considered IPSO offers better cost function, PSNR, SSIM and lesser CPU time for most of the test images compared with the traditional PSO algorithm.

References

1. Acharya RU, Ng EYK., Sree VS, Chua CK, Chattopadhyay S. Higher order spectra analysis of breast thermograms for the automated identification of breast cancer, Expert Systems, 2014, 31(1): 37 - 47.

2. Couceiro, M.S., Rocha, R.P., Ferreira, N.M.F. and Machado, J.A.T. 2012a. Introducing the fractional-order Darwinian PSO. Signal Image and Video Processing, 6(3), 343-350.

3. Francis SV, Sasikala M, Bharathi GB, Jaipurkar SD. Breast cancer detection in rotational thermography images using texture features, Infrared Physics & Technology, 2014, 67: 490-496

4. Ghamisi P, Couceiro MS, and Benediktsson JA. Classification of hyperspectral images with binary fractional order Darwinian PSO and random forests,SPIERemote Sensing, 2013: 88920S-88920S-8.

5. Ghamisi P, Couceiro MS, Martins FML and Benediktsson JA. Multilevel image segmentation based on fractional-order Darwinian particle swarm optimization, IEEE T. on Geoscience and Remote sensing, 2014, 52(5): 2382-2394.

6. GolestaniN, Tavakol EM, Ng EYK..Level set method for segmentation of infrared breast thermograms, EXCLI Journal, 2014,13:241-251.

7. Kennedy, J, Eberhart, R. C. Particle swarm optimization, In Proceedings of IEEE international conference on neural networks, 1995:1942-1948.

8. Ng EYK. A review of thermography as promising non-invasive detection modality for breast tumor, International Journal of Thermal Sciences, 2009, 48(5):849-859.

9. Otsu N. A Threshold selection method from Gray-Level Histograms,IEEE T. on Systems, Man and Cybernetics, 1979, 9(1): 62-66.

10. Raja NSM, Kavitha G, Ramakrishnan S. Analysis of Vasculature in Human Retinal Images Using Particle Swarm Optimization Based Tsallis Multi-level Thresholding and Similarity Measures, Swarm, Evolutionary, and Memetic Computing, Lecture Notes in Computer Science (LNCS 7677), Vol. 7677, 2012: 380-387.

11. Raja NSM, Rajinikanth V, Latha K. Otsu Based Optimal Multilevel Image Thresholding Using Firefly Algorithm, Modelling and Simulation in Engineering, vol. 2014, Article ID 794574, 17 pages, 2014.

12. Rajinikanth V, Latha K. Optimization of PID Controller Parameters for Unstable Chemical Systems Using Soft Computing Technique, International Review of Chemical Engineering, 2011, 3(3): 350-358.

13. Rajinikanth V, Couceiro MS. Multilevel Segmentation of Color Image using Levy driven BFO Algorithm, In proceedings of ICONIAAC'14, Article no.19.2014. D01:10.1145/2660925.

14. Rajinikanth V, Raja NSM, Latha K. Optimal Multilevel Image Thresholding: An Analysis with PSO and BFO Algorithms,4ust. J. Basic & Appl. Sci., 8(9): 443-454.

15. Tillett, T., Rao, T.M., Sahin, F. and Rao, R. 2005. Darwinian particle swarm optimization. Proceedings of the 2nd Indian international conference on artificial intelligence, Pune, India, 1474-1487.

16. Wei-Der Chang, Shun-Peng Shih, PID controller design of nonlinear systems using an improved particle swarm optimization approach, Communications in Nonlinear Science and Numerical Simulation, 2010, 15(11): 3632-3639.

17. Suganthi SS, Ramakrishnan S. Evaluation of human hand thermal images using wavelet transform based local spatial features, Biomed Sci. Instrum. 2013,49:1-6.

18. http://www.mathworks.com/matlabcentral/fileexchange/authors/117313.

19. http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/BSDS300/html/dataset/images.html

20. http://visual.ic.uff.br/dmi/prontuario/home.php