ORGANIZERS AND SUPPORTERS

(1)

(2)

ORGANIZERS AND SUPPORTERS

Organized by

Research Center for Informatics, Indonesia Institute of Sciences - Indonesia Supported by

Research Center for Informatics, Indonesia Institute of Sciences - Indonesia IEEE Section Indonesia

Ministry of Communication and Information - Kominfo

iii

(3)

SCIENTIFIC AND ORGANIZING COMMITTEE

Scientific Committee

Ana Hadiana (Research Center for Informatics, Indonesian Institute of Sciences) - Chair Agus Subekti (Research Center for Informatics, Indonesian Institute of Sciences) - Co-Chair

Roberto Togneri (Electrical, Electronic and Computer Engineering, The University of Western Australia, Australia) Bernadetta Kwintiana Ane ( Institute of Computer-aided Product Development Systems, Universitat Stuttgart, Germany)

Mohd Yusoff Mashor (University of Malaysia Perlis, Malaysia)

Md. Mahmud Hasan (Kazakh-British Technical University, Kazakhstan) Moncef Gabbouj (Tampere University of Technology, Finland)

Anitawati Mohd Lokman (Universiti Teknologi MARA, Malaysia)

Hilman Ferdinandus Pardede (Research Center for Informatics, Indonesian Institute of Sciences, Indonesia) Laksana Tri Handoko (Research Center for Informatics, Indonesian Institute of Sciences, Indonesia )

Andriyan Bayu Suksmono (School of Electrical Engineering and Informatics, Bandung Institute of Technology, Indonesia)

Kuncoro Wastuwibowo (Chair IEEE Indonesia Section, Indonesia)

Anto Satriyo Nugroho (Center for Information & Communication Technology Center for the Assessment & Appli- cation of Technology, BPPT, Indonesia)

Estiko Rijanto (Research Center for Electrical Power and Mechatronics, Indonesian Institute of Sciences, Indonesia) John Batubara (School of Electrical Engineering, Bina Nusantara University, Indonesia)

Chan Basaruddin (Faculty of Computer Science, Indonesia University, Indonesia)

Rifki Sadikin (Research Center for Informatics, Indonesian Institute of Sciences, Indonesia)

iv

(4)

2013 International Conference on Computer, Control, Informatics and Its Application - Frontmatters

Semi Automatic Detector of Plasmodium Falciparum on Microscope Image Based

Iis Hamsir Ayub Wahab

Jurusan Teknik Elektro dan Teknologi Informasi Universitas Gadjah Mada

Yogyakarta, Indonesia

Email: [email protected]

Paulus Insap Santosa

Jurusan Teknik Elektro dan Teknologi Informasi Universitas Gadjah Mada

Yogyakarta, Indonesia

Adhi Susanto

Jurusan Teknik Elektro dan Teknologi Informasi Universitas Gadjah Mada

Yogyakarta, Indonesia

Maesadji Tjokronegoro Fakultas Kedokteran Universitas Gadjah Mada

Yogyakarta, Indonesia

Abstract—The inspection standard for diagnosis of active malaria for parasites in blood slides through a microscope. Although microscopy has good sensitivity and allows species identification and parasite counts it requires microscopy expertise and involves a labor intensive repetitive task which is time consuming. The main aim of this research is to build an automatic system for identify and classify the parasite plasmodium falciparum based on a digital image scheme. The results of this research show the color features can give higher accuracies based on the architecture of LVQ network with four hidden neurons.

Keywords—computer aid; detection; plasmodium falciparum;

identify; classication

I. I NTRODUCTION

Standard definitive diagnosis of malaria infection is to find parasites in blood slides through a microscope. However, it is a routine and time-consuming task. In addition, recent studies in the field suggest that the level of agreement among clinical experts for diagnosis is quite low [1]. Therefore, it is essential to generate neither the tools nor the general standard algorithm that is able to diagnose.

Automatic detection of malaria parasites have been made by some researchers that the number of malaria parasites in the blood can be calculated based on the acquisition of blood samples were converted into digital images. Detection systems are generally built through several processing stages, namely:

image acquisition, image pre-processing, image segmentation, feature extraction and classification.

Selena, et al. [2], develop an automated system for counting malaria parasites quickly and accurately based image analysis.

The image obtained from the acquisition using a video camera 3-CDD. There are four steps being taken to be able to count the number of parasitemia, namely: edge detection, edge linking, clump splitting and parasite detection. Parasites can be identified based on the characteristic properties of the edge of the red blood cells by using Sobel operator and the transformation Hough [3]. This is also done by Diaz et al. [4], but the system is built using a low pass filter on the pre- processing of the image, and then extracts the image pre-

processing results in order to obtain information which the blood cells infected with malaria parasites.

Detection of malaria parasite can be improved when using machine learning strategies [5]. The system is a system built classifiers based on feature of each type of parasite. Features that have been extracted from previous experiments are feature histogram [6][7], morphological [8], texture [9]. The classifier system is used multilayer perceptron neural network (MLP) [5], support vector machine (SVM) [4][10], k-nn [3], and the type of back propagation[3].

The main objective of this research is to design a computer- assisted system for detecting and classifying parasite Plasmodium falciparum in the blood preparations are indicated based digital microscopic images using learning vector quantization neural networks and image segmentation.

Analysis of the methods tested is the detection and classification accuracy, the speed of the system is built. For this study, we also consider the developmental cycle of Plasmodium falciparum, i.e. gametocyte, ring, trophozoite and schizonts.

II. M ATERIAL AND M ETHOD A. Material

In this study, the data samples used were obtained from blood samples of patients on glass microscope slides. Then preparations in the form of a digital image file in Jpg format with 24-bit color depth and size of 256x256 pixels. The number of data is 96 with the number of class 4, according to the type of variation in the condition of Plasmodium falciparum.

B. Method

The proposed method can be seen in Figure 1. This method consists of: pre-processing of the image based on the following techniques are discussed i.e. the process of image filtering, segmentation, feature extraction and classification.

2013 International Conference on Computer, Control, Informatics and Its Application

59

(10)

Fig. 1. The proposed method

Here's an explanation of the stages of the overall system:

1) Image aqcuisition: this process aims to obtain images of the preparations plasmodium falciparum which will be further processed as an object to be classified.

2) Pre-proccesing: image pre-processing is done to improve the quality of the image obtained by the change of optical to electrical likely caused a decline in the quality of the sensor noise, blur due to camera or a lack of focus because the camera relative movement of the object. Image enhancement techniques used in this study is the bilateral filter.

Bilateral filter is an improved method of image quality of non- linear adaptive. This method is a combination of two filter methods, namely domain filter and range filter [13]. Combined the two enables a more effective improvement to the quality of the image is degraded by noise. Bilateral filter method used two kinds of weighting the pixels in the image, thus resulting in a more accurate calculation than previous methods [12].

The first weight in the bilateral filter is a method of spatial weights, which calculates the geometric proximity of pixels, while the second weight is the weight that measures the photometric color intensity differences between pixels in the image. Both of these weight calculations will result in effective weight values for proximity into account the distance between pixels and color intensity differences between pixels. Domain low pass filter used in image g(x) to produce the output image is defined as follows.

( ) ( )

∫ ∫

−^∞∞

∞

−

= ^k ^x ^g ξ ^c ξ ^x ^d ξ x

f ( )

_d¹

( ) (1)

Where ^c ( ^ξ ⁻ ^x ) is a closed geometric measure between the center of neighbors x and the point of the proximityξ, f and g are both factually a multiband image input and output. To maintain a low pass component of the signal, then:

∫ ∫

−^∞∞

( )

∞

=

−

^c ξ ^d ξ x

k

_d

( ) (2)

if the shift invariant filter, then the ^c ( ^ξ ⁻ ^x ) is simply a different function of vector ξ - x, and k

d

is constants. The same manner filter range obtained using the following equation.

( ) ( )

∫ ∫

−^∞∞

∞

−

= ^k ^x ^g ξ ^s ^f ξ ^g ^x ^d ξ x

f ( )

_r¹

( ) ( ) ( ) (3)

In this case, the kernel will measure photometric similar between the pixels. Normalization constant in this case is

∫ ∫

−^∞∞

( )

∞

−

= ^s ^f ξ ^g ^x ^d ξ x

k

_r

( ) ( ( )) (4)

Spatial distribution of image intensity is not so instrumental to filter the range by himself. The combination of the intensity of the whole image would make a little sense of change, because the distribution of image values far away from the x and otherwise not to affect the final value at x [13]. Additionally, it can be shown that the filter without the filter range domain is just changing the color of an image map. Appropriate solution is to combine domain filter and range, thus strengthening the geometric and photometric aspects. Combination filter range and domain filter can be shown in the following equation:

∫ ∫

−^∞∞

( ) ( ) ( )

∞

−

− −

= ^k ^x ^g ξ ^c ξ ^x ^s ^g ξ ^g ^x ^d ξ x

f ( )

_d¹

( ) ( ) ( ) (5)

With normalization:

( ) ( )

∫ ∫

−^∞∞

∞

−

− −

= ^c ξ ^x ^s ^g ξ ^g ^x ^d ξ x

k ( ) ( ) ( ) (6)

Combination filter range and domain filter is called the bilateral filter.

The image quality has been improved, and then will be separation of objects based on color intensity. This is because the objects which will split the intensity of color tend to have different and each object has a nearly uniform color [6]. In this study, k-mean segmentation method is used to separate objects.

K-mean method is basically to divide the image histogram.

The stages are as follows.

- Find the maximum and minimum intensity in the image - From minimum to maximum intensity do the division

number N. N determines the number of objects that are expected to exist in the image.

- After the distribution, the histogram will be divided into sections called clusters. Then do a search for the image of the whole point, every point will be assigned to the nearest cluster so that the end result of this process is the number of colors in the image into N.

- Then look for the results rata-rata/mean whole point in each cluster,

- Then change the color of all the points in the clusters is the average of each cluster.

3) Feature extraction: to obtain the specific characteristics of each image acquired malaria parasite preparations and give each pattern in the image. To extract the characteristics of each image can be used characterize the color histogram

4) Classification: in this study, the classification method used is a neural network learning vector quantization (LVQ).

LVQ is a method to conduct the training competitive layers are supervised learning [14]-[16]. LVQ consists of two layers, namely layers competitive and linear layers as shown in Figure 2.

60

(11)

R1

Gi

Bn

c1

cj

cm

w11

w1j

w1m

wi1

wij

wim

wn1

wnj

wnm

input hidden neuron

Tn

T1

Tk

Tl

output v11

v1k

v1l

v1n

vj1

vjk

vjl

vjn

co

w1o

wio

wno

vm1

vmk

vml

vmn

vo1 vok

vol

von

Fig. 2. LVQNN Architecture

The first layer is a competitive layer classifies the input with a competitive manner. The second layer is linear layer transforms the sub-classes to be output first tier classes that have been defined previously the target. Each subclass is represented by a single neuron in the output layer are competitive and each class is represented by a linear neuron in the output layer. Output neurons in the competitive layer is usually referred to as the hidden neurons and linear neurons in the output layer is referred to as the output neuron. Subclass the competitive layer is the layer that competition results in the classroom while the linear layer is defined by the user class. Processing that occurs at each neuron is finding the shortest distance between an input vector to the corresponding weight. In this case w is the weight vector connecting each neuron in the input layer to the hidden neurons in the output layer. Activation function in the output layer will map the weight vector of the linear layer to the output layer.

From Figure 2 it can be seen in the competitive layer neuron is connected to every input neuron to the hidden certain weight (w). Whereas in each hidden layer neuron is connected linearly with the output neurons with a certain weight (v). The output of the plasmodium falciparum gametocyte be categorized in class 1, the plasmodium falciparum ring will be categorized in class 2, the plasmodium falciparum schizonts will be categorized in classes 3 and the plasmodium falciparum trophozoite be categorized in class 4.

The test results will be used to determine the type of parasite detection.

Here are given training algorithm LVQ network if the data is given by m n input variables.

LVQ network training algorithms:

Set:

Initial weight of the j-th input variable to the class heading into the Wij-i, with i = 1, 2, ..., K and j = 1, 2, ..., m.

Maximum epoch = Maxepoh.

Parameter learning rate = α ^. Reduction in learning rate = Dec. α ^. Minimal learning rate allowed α ^{= Min.}

Enter:

Data input Xij with i = 1, 2, ..., n and j = 1, 2, ..., m.

Targets in the form of classes, ie Tk with k = 1, 2, ..., n.

Set initial conditions epoch, the epoch = 0.

Do if (epoch ≤ Maxepoh) and ( ≥ α α ^Min) epoch = epoch 1

Done for i = 1 to n

Define J such that | Xi - Wj | j = minimum with 1, 2,, K.

Fix Wj with the following provisions:

If T = Cj then Wj = Wj α (Xi - Wj) If T ≠ Cj then Wj = Wj - α (Xi - Wj) Subtract the value of α

( α value reduction can be done with α ⁼ α ^{- Dec.} α ^or α ⁼ α way Dec.* α ^.

end

After the training will be obtained final weights (W). A weight is what will be used to perform testing or simulation with different data (not the same as the data used for training).

LVQ network testing algorithm:

Enter the data to be tested, for example, Xij with i = 1, 2, ..., np and j = 1, 2, ..., m.

Done for i = 1 to np

Define J such that | Xi - Wj | j = minimum with 1, 2, ..., K.

J is the class for Xi

The parameters used in the LVQ neural network in this study:

the rate of initial training α = 0.01; decrease in the rate of training Dec. α = 0.1; minimum training rate Min α = 10

^-6

; maximum epoch = 5000;, and Hidden neurons = 4. The terms used to stop training are the convergence of the weight vector value to a fixed value.

III. R ^ESULT

The resulting digital image of the acquisition process in this study still have lighting effects, focusing on the optical lens and the camera, and the effect of noise so the recovery should be done by using the bilateral filter. In the screening process, first color space is converted to the CIELAB color space. This is done because the results of image filtering in CIELAB color space using the bilateral filter will produce natural colors images [13].

The results of this filtering will be reduce the noise on the image so that the image of the color components red, blue and green for all images undergo significant changes. The best results from the use of domain filter parameter equal to 3 pixels together with the results of experiments carried out [6], [12] and [13] which uses the same values to calculate the weight of proximity pixels geometrically, while the weight of the photometric measures the difference between the color intensity pixel in the image is what changed the value to get the best results. Three dimensional visualization of the image before and after filtered using a bilateral filter can be seen in Figure 3.

After the restoration of the image of the sample preparation, the next step is to separate the parts of blood parasites on image segmentation using k-mean. In this study, the segmentation process performed prior image color space conversion from RGB color space to CIELAB color.

Grouping colors in CIELAB color space is obtained by using the Euclidean distance formula of elements of L , a , and b *, and thus will get the average color values corresponding closest

61

(12)

distance of each color. K values used in the simulation are k = 3, 4, 5, 6, and 7. Segmentation results for sample preparation for each of plasmodium falciparum can be seen in Figure 4.

(a)

(b)

Fig. 3. 3D modeling for image (a) before and (b) after filtered

(a) (b)

(c) (d)

Fig. 4. Results of image segmentation plasmodium falciparum (a) gametocyte, (b) ring (c) schizonts and (d) trophozoite

Separation of objects for plasmodium falciparum gametocyte and schizonts of red blood cells can be done either when the value of k = 3. Separation can be done well because of the morphological form of the two types of plasmodium is very clear where plasmodium falciparum gametocyte shaped like a banana and plasmodium falciparum schizonts shaped blobs. For a sample image of plasmodium falciparum ring, separation of objects with red blood cells can be done with the best value of k = 3 because of the morphological shape flaciparum plasmodium is shaped like a ring that is sometimes very thin. In the experiments that have been done by adding the value of k in the sample group, the object will come with reduced color other objects so that the object will be damaged or lost when given a value of k greater than 3 [4][5]. As with the plasmodium falciparum trophozoite, to separate the object from the red blood cells required divider value different k is 3, 5 and 7.

The output of the segmentation is then taken as the peak color component characteristic vector input pattern classification plasmodium falciparum. An image of the object has a specific color. To state the color characteristics of the image of the object to note how the behavior of the distribution of the color itself. Granted many image objects have the same basic color but the color distribution in it have turned out differently[4]. Distribution can be obtained using the color histogram of colors. In this study, the peak value of the feature of color histogram is used as input to the classification system. Results of the peak value of the RGB color components can be seen in Table 1. After the feature extraction process of the sample image, the next step is to use the maximum color component feature of as input vector pattern classification parasites using LVQ neural network.

LVQ is designed to classify four types of plasmodium falciparum parasite image so that the network will have four target encoding pattern with the number of data samples of 96 copies. Prior to conducting experiments training and testing system, initial weights are determined in advance. Initial weights are determined from the average value across the data for each of the color feature shown in Table 2. Experiment is to observe the performance of the network by varying the rate of training. LVQ will be tested with the training given the rate of α of 0.01; 0.02; 0.03; 0.04; 0.05, and 0.1.

Table 3 showed that the accuracy of training data and test data for all the same training the rate of, i.e. 100% of the trained image data 96 and 91.67% of the 48 new data were tested on this network. Reaching network to converge at epoch 66 with MSE values approaching fault tolerance value that is equal to 1,045 × 10-6 (Figure 5). This suggests that the rate of addition of training on LVQ network with input feature do not affect the color of the center point of the neurons, so that the training data is not moved to another class [6]. So the network has a steady state for input color feature.

62

(13)

TABLE I. R

ANGE OF COLOR FEATURE FOR FALCIPARUM PLASMODIUM

Sample of Plasmodium

falciparum Rmaks Gmaks Bmaks

gametocyte 1603 - 15097 1155 - 4996 1486 - 3678 ring 3715 - 13751 1574 - 4192 2326 - 10768 schizonts 1790 - 14163 1492 - 3521 1809 - 10999 trophozoite 4527 - 18854 2417 - 14471 1782 - 16481

TABLE II. C

ODE OF TARGET PATTERN OF NEURAL NETWORK AND THE INITIAL WEIGHT

R

ANGE OF COLOR FEATURE

Sample of Plasmodium

falciparum

Bit output of neuron

initial weight of the color feature

1 2 3 4 Rmax Gmax Bmax gametocyte 1 0 0 0 10654 2194 4777

ring 0 1 0 0 4040 2230 8083 schizonts 0 0 1 0 2223 1962 1899 trophozoite 0 0 0 1 16227 9751 9352

Fig. 5. Performance of training LVQ

TABLE III. E

FFECT OF THE TRAINING RATE OF COLOR FEATURE

training

rate MSE

Accuracy training data data test

0,01 9,55 × 10

^-6

100% 91,67%

0,02 9,135 × 10

^-6

100% 91,67%

0,03 9,989 × 10

^-6

100% 91,67%

0,04 9,71 × 10

^-6

100% 91,67%

0,05 9,831 × 10

^-6

100% 91,67%

0,1 9,404 × 10

^-6

100% 91,67%

I. C ONCLUSION

The analytical results of the tests performed on the system identification of plasmodium falciparum is known that the use of bilateral filter, segmentation using k-mean clustering method and LVQ neural network with feature color as vectors enter it was able to identify the types of parasite plasmodium falciparum well.

R EFERENCES

[1] K. Mitiku, G. Mengistu, and B. Gelaw. The reliability of blood film examination for malaria at the peripheral health unit. Ethiop.J.Health Dev, 17(3):1pp. 97–204, 2003.

[2] Selena W.S. Sio, Weiling Sun, Saravana Kumar, Wong Zeng Bin, Soon Shan Tan, Sim Heng Ong, Haruhisa Kikuchi, Yoshiteru Oshima, Kevin S.W. Tan. MalariaCount: An image analysis-based program for the accurate determination of parasitemia. Elsevier. 2006.

[3] Minh-Tam Le, Timo R Bretschneider, Claudia Kuss and Peter R Preiser.

A novel semi-automatic image processing approach to determine Plasmodium falciparum parasitemia in Giemsa-stained thin blood smears. BMC Cell Biology 2008, 9:15 doi:10.1186/1471-2121-9-15.

2008.

[4] Gloria Díaz, Fabio A. González, Eduardo Romero. A semi-automatic method for quantification and classification of erythrocytes infected with malaria parasites in microscopic images. Journal of Biomedical Informatics 42 (2009) 296–307. 2009.

[5] Charles Ma, Paul Harrison, Lina Wang, Ross L Coppel. Automated estimation of parasitaemia of Plasmodium yoelii-infected mice by digital image analysis of Giemsa-stained thin blood smears. Malaria Journal, 9:348. 2010.

[6] Wahab, Iis Hamsir Ayub. Identifikasi Parasit Malaria Dalam Darah Menggunakan Metode Segmentasi Citra Digital Dan Jaringan Syaraf Tiruan. Tesis. Teknik Elektro. Universitas Gadjah Mada. 2008.

[7] Wahab, Iis H A., Adhi Susanto and Litasari. Classification of Plasmodium falciparum Using Learning vector quantization Neural Network. Proceeding of International Conference on Biomedical Engineering, pp 17. Surabaya, Indonesia. 2008.

[8] Subhamoy Mandal, Amit Kumar, J Chatterjee, Manjunatha and Ajoy K.

Segmentation of Blood Smear Images using Normalized Cuts for Detection of Malarial Parasites. Annual IEEE India Conference (INDICON) 978-1-4244-9073-8/10. 2011.

[9] Vishnu V. Makkapati, Reghuveer M. Rao. Segmentation of Malaria Parasites in Pheriperal Blood Smear Images. ICASSP 2009. 978-1- 4244-2354-5/09. 2009.

[10] F Boray Tek, Andrew G Dempster and Izzet Kale. Computer vision for microscopy diagnosis of malaria. Malaria Journal. 2009

[11] Dian Anggraini, Anto Satriyo Nugroho, Christian Pratama, Ismail Ekoprayitno Rozi, Aulia Arif Iskandar5, Reggio Nurtanio Hartono.

Automated Status Identification of Microscopic Images Obtained from Malaria Thin Blood Smears. International Conference on Electrical Engineering and Informatics. Bandung, Indonesia. 2011.

[12] Adrianto, Hari Agung., Fahren Bukhari., dan Laura Prescilla.

Penggunaan Kualitas Citra Warna Menggunakan Metode Bilateral Filter. Jurnal Departemen Ilmu Komputer Vol. 5 - No. 2 . Bogor Agricultural University. 2007.

[13] Tomasi, C and Manduchi, R. Bilateral Filtering for Gray and Color Images. Proceedings of the IEEE International Conference on Computer Vision. Bombay, India. 1998.

[14] J. C. Bezdek and N. R. Pal. Two soft relatives of learning vector quantization. IEEE Trans. Neural Networks, vol. 8, no. 5, pp. 729–743.

1995.

[15] Kohonen T. New developments of learing vector quantization and the self-organizing map. Symposium on Neural Networks; Alliances and Perspectives in Senri 1992 (SYNAPSE' 92), Osaka, JAPAN. 1992.

[16] Kohonen, T. Self-organizing maps, Springer, Berlin, Bibliography on the Self-Organizing Map (SOM) and Learning vector quantization (LVQ). Neural Networks Research Centre, Helsinki University of Technology. 1995.

0 10 20 30 40 50 60 70

0 1 2 3 4 5 6 7 8

9x 10^-3 Performance is 9.55e-006 Goal is 1e-006

66 Epoch

Training-Blue Goal-Black

ORGANIZERS AND SUPPORTERS