Classification Of Malaria Parasite Species Based On Thin Blood Smears Using Multilayer Perceptron Network

Transcription

1 Classification Of Malaria Parasite Species Based On Thin Blood Smears Using Multilayer Perceptron Netor Noorhidayati Abu Seman, * Nor Ashidi Mat Isa, Lim Chia Li, Zeehaida Mohamed, Umi Kalthum Ngah, Kamal Zuhairi Zamli School of Electrical & Electronic Engineering, Universiti Sains Malaysia, Engineering Campus, 4300, Nibong Tebal, Pulau Pinang, Malaysia. * ashidi@eng.usm.my Abstract This paper discusses the application of the MLP netor to classify the malaria parasite into three species, namely Plasmodium falciparum, Plasmodium vivax and Plasmodium malariae. Six features (i.e. size of RBC infected per size of normal RBC, shape of parasite, number of chromatin, number of parasite per RBC, texture of RBC and location chromatin of parasite) from thin blood smear ere used as input data. In order to determine the applicability of the MLP netor, three different training algorithms ere employed to train the MLP netors. In this study, the MLP netor trained using bac propagation algorithm produced the best performance ith 89.80% accuracy as compared to Levenberg- Marquardt and Bayesian Rule algorithms. The result significantly demonstrates the suitability of the MLP netor for classifying the malaria parasite. Keyords: Malaria parasite, intelligent system, multilayered perceptron netor, thin blood smear. Introduction Malaria is one of the leading causes of death in the orld. It is estimated that there are million ne cases every year, ith.5 to.7 million deaths orldide []. Research has shon that malaria is caused by four species of the genus malarial parasite Plasmodium (i.e. Plasmodium falciparum, Plasmodium vivax, Plasmodium ovale and Plasmodium malariae) []. Plasmodium is actually a small, single-cell blood organism or protozoan hich originated from a species of mosquito called Anopheles []. It is the Anopheles mosquito bites that can spread the Plasmodium into the human blood stream. Currently, there are to tests to detect the malaria parasite; i.e. thic and thin blood smears. The thin blood smear is used to determine the species of malaria parasite, hile thic blood smear is used to find the density of malaria parasite per microliter of red blood cells (RBC). A number of useful researches into analyzing and classifying malaria parasite using neural netor do exist. For instance, Jayavanth and Singh [3] have used an artificial neural netor to analyze malaria severity through aggregation and deformability parameters of erythrocytes. The research successfully classifies high severity malaria ith 00% accuracy. Nevertheless, the success rate for the classification of non-severe and mildlysevere malaria spans the range from 60% to 80%. Gao et. al [4] advocates that neural netor as effective for forecasting malaria. In their study, the neural netor yield promising result ith the efficiency of 84.85%. The study also concludes that neural 46

2 Classification of Malaria Parasite Species Based on Thin Blood Smears Using Multilayer Perceptron Netor netors are useful for malaria forecasting particularly due to flexible analysis possibility, loer claim for data, convenient and easy to apply. Bender et. al [5] describes the application of neural mitochondrial transit peptides (mtps) from the malariacausing parasite Plasmodium falciparum Multilayered perceptron (MLP) netor trained using bac propagation (BP) algorithm is the most popular choice in neural netor applications. The current study proposes the MLP netor to classify the malaria parasite and compare the performance of various available training algorithms namely bac propagation, Levenberg-Marquardt and Bayesian rule. This paper is organized as follos. Section.0 discusses the basic concept of the MLP netor. Section 3.0 outlines the training algorithms employed in this research. Section 4.0 gives our methodology. Section 5.0 gives our result and discussion. Finally, section 6.0 outlines our conclusion. here the denotes the eights that connect the input and the hidden layer; and b i denote the input that are supplied to the input layer and thresholds in hidden nodes respectively; n i and n h are number of input nodes and hidden nodes respectively. n 0 $y n n 0 n ŷ n o n n i n i n 0 n n n i x i Output Layer Hidden Layer Input Layer. Multilayered Perceptron Netor x x x n i A MLP netor is a feed forard neural netor ith one or more hidden layers. Cybeno [6] and Funahashi [7] have proved that the MLP netor is a general function approximator and the MLP netor ith one hidden layer (as shon in Figure ) is sufficient to approximate any continuous function. Based on Figure, the input layer acts as an input data buffer that distributes the input to the hidden layer. The outputs from the hidden layer then become the inputs to the output layer, hich provides the netor output. A hidden neuron performs to functions, i.e. the combining function and the activation function. Consider a MLP netor ith n i input nodes, the output of the -th neuron of the hidden layer is given by: n i () () v = t F i xi t + b ; for n h () i= Figure : Multilayered Perceptron Netors The output of the -th output neuron, y in the output layer is given by: n h () t = v () t yˆ ; for n o () = here n o is the number of output nodes; denotes the eights of the connections beteen the hidden and output layer. It can be derived from equations () and () that the MLP netor ith one hidden layer can be expressed by folloing equation: nh n i () () yˆ t = F i xi t + b ; = i= for no (3) F(.) is an activation function that is normally selected as a sigmoid function, hich is given by: International Journal of the Computer, the Internet and Management Vol. 6. No. (January-April, 008) pp

3 Fvt ( ( )) = ( ) (4) + e vt The eights i, and threshold b are unnon and should be selected to minimize the prediction errors, defined as: ρ () t = y () t yˆ (5) here y (t) is the actual output and yˆ is the netor output. 3. Training Algorithms As described earlier, three training algorithms for the MLP netor ill be employed and compared their performance. This section briefly presents the bac propagation, Levenberg-Marquardt and Bayesian Rule algorithms. 3. Bac Propagation Algorithm Bac-propagation, the most commonly used to train The MLP netor, is a gradient descent procedure that computes the derivatives values in a very efficient ay, and modifies the eights according to a parameter non as learning rate [8]. Bac propagation is a steepest decent type algorithm here the eight connection beteen the -th neuron of the hidden layer and the i-th neuron of the input layer are respectively updated according to: = b = b i i ( t ) + Δ ( t ) + Δb i (6) ith the increment Δ i (t) and Δb (t) given by: Δi = ηρ xi + α Δ i ( t ) (7) Δb = ηbρ + αbδb ( t ) here the subscripts and b represent the eight and threshold respectively, α and α b are momentum constants hich determine the influence of the past parameter changes on the current direction of movement in the parameter space, η and η b represent the learning rates and ρ (t) is the error signal of the -th neuron of the hidden layer hich is bac propagated in the netor. Since the activation function of the output neuron is linear, the error signal at the output node is ρ t = y t yˆ t (8) ( ) ( ) ( ) and for the neurons in the hidden layer ( ) = F' ( xi ) ρ i ( t here '( x ) F( x i (t)) ith respect to (t). ρ t ) (9) F i is the first derivative of Since bac propagation algorithm is a steepest decent type algorithm, the algorithm suffers from a slo convergence rate. The search for the global minima may be trapped at local minima and the algorithm can be sensitive to the user selectable parameters [9]. 3. Levenberg-Marquardt Algorithm Levenberg-Marquardt algorithm is a gradient-based, deterministic local optimization algorithm. The Levenberg- Marquardt algorithm has an advantage over the traditional Bac Propagation algorithm, here it can provide faster (second-order) convergence rate and eep relative stability [0][]. Lie the quasi-neton methods, the Levenberg-Marquardt algorithm as designed to approach second-order training speed ithout having to compute the Hessian matrix. When the performance function has the form of a sum of squares (as is typical in training feed forard netors), then the Hessian matrix can be approximated as: T H = J J (0) and the gradient can be computed as: T g = J e () here J is the Jacobian matrix that contains first derivatives of the netor errors ith respect to the eights and biases, and e is a x i 48

4 Classification of Malaria Parasite Species Based on Thin Blood Smears Using Multilayer Perceptron Netor vector of netor errors. The Jacobian matrix can be computed through a standard bac propagation technique that is much less complex than computing the Hessian matrix []. The Levenberg-Marquardt algorithm uses this approximation to the Hessian matrix in the folloing Neton-lie update: T T [ J J + I ] J e Δ = μ () here Δ is a differential eights and µ is a control parameter. When the scalar µ is zero, it is similar to Neton's method, using the approximate Hessian matrix. When µ is large, it becomes gradient descent ith a small step size. Neton's method is faster and more accurate near an error minimum, so the aim is to shift toards Neton's method as quicly as possible. Thus, µ is decreased after each successful step (reduction in performance function) and is increased only hen a tentative step ould increase the performance function. In this ay, the performance function ill alays be reduced at each iteration of the algorithm []. 3.3 Bayesian Rule Algorithm Given the Baye s Rule as [3] D θ ) θ D) = (3) D) here p (θ ) is the prior probability of a parameterθ before having seen the data and θ D) called the lielihood ere the probability of the data D. Bayes Rule is used to determine the posterior probability of θ given the data D [3]. In general this ill provide an entire distribution over possible values of θ. This process as apply to neural netors and come up ith the probability distribution over the netor eights,, given the training data D). When finding a posterior distribution over eights, D ) ) D) = D) D ) = D ) ) d (4) In the Bayesian formalism, learning the eights means changing our belief about the eights from the prior, ), to the posterior, D) as a consequence of seeing the data as illustrated by Figure. Figure : Changing prior eights to posterior eights 4. Methodology and Data Samples As discussed earlier, this study focuses on classification of three species of malaria parasite; i.e. Plasmodium falciparum, Plasmodium vivax and Plasmodium malariae from the thin blood smears. To determine the applicability of the MLP netor as malaria parasite diagnosis technique, the MLP netor needs to go through training and testing phases. During both phases, the optimum structure and diagnosis performance of the MLP netors are determined. The performance analysis of the MLP netor is based on accuracy. Accuracy is defined as the percentage of overall correct determination of malaria species. In this study, 56 data of malaria parasite (7 Plasmodium falciparum, 94 Plasmodium vivax and 96 Plasmodium malariae) are used in classifying malaria parasite using the MLP netor. The data are taen from Hospital Universiti Sains International Journal of the Computer, the Internet and Management Vol. 6. No. (January-April, 008) pp

5 Malaysia (HUSM). Six features of the thin blood images ill be used as input data to the MLP netor. The features are size of RBC infected per size of normal RBC, shape of parasite, number of chromatin, number of parasite per RBC, texture of RBC and location chromatin of parasite. 350 data (0 Plasmodium falciparum, 0 Plasmodium vivax and 0 Plasmodium malariae) are used as training data hile the remaining data (6 Plasmodium falciparum, 74 Plasmodium vivax and 76 Plasmodium malariae) are used as testing data. The data are fed ramdomly into the MLP netor. 5. Result and Discussion The optimum number of hidden node and number of training epochs are obtained hen the MLP netor achieves the highest performance. Figure 3 (a) and (b) sho the result of obtaining the optimum training epochs and number of hidden node for bac propagation algorithm respectively. The MLP netor using bac propagation algorithm achieves the highest performance at the number of hidden nodes equal to 00 and 4 training epochs. Results for Levenberg-Marquardt training algorithm are shon in Figure 4. This training algorithm produced the best performance at 00 training epochs and 7 hidden nodes. Figure 5 shos the result for the Bayesian Rule training algorithm. This training algorithm achieved an optimal result at 000 training epochs and 3 hidden nodes. (a) Training epochs (b) Hidden node Figure 3: Performance of the MLP netor ith bac propagation algorithm. (a) Training epoch (b) Hidden node Figure 4: Performance of the MLP netor ith Levenberg-Marquardt algorithm. 50

6 Classification of Malaria Parasite Species Based on Thin Blood Smears Using Multilayer Perceptron Netor (a) Training epoch (b) Hidden node Figure 5: Performance of the MLP netor ith Bayesian rule algorithm. After obtaining the optimum structure for the netor, the performance of the MLP netor as determined. Table shos the performance comparison of the MLP netor using those three training algorithms. The result shos that bac propagation training algorithm produces the highest accuracy, 89.80% as compared to Levenberg-Marquardt and Bayesian Rule training algorithm, hich produce 87.35% and 87.76% of accuracy respectively. Table : The performance comparison of the MLP netor ith three different training algorithms. Training algorithm Accuracy Bac propagation 89.80% Levenberg-Marquardt 87.35% Bayesian Rule 87.76% 6. Conclusion The results obtained indicate that, the MLP netor hich has been trained ith the bac propagation algorithm produced the highest performance compared to the Levenberg-Marquardt and Bayesian Rule algorithms. The result also proved that the MLP netor can be implemented to classify malaria parasite into three classes (i.e. Plasmodium falciparum, Plasmodium vivax and Plasmodium malariae) based on six features (i.e. size of RBC infected per size of normal RBC, shape of parasite, number of chromatin, number of parasite per RBC, texture of RBC and location chromatin of parasite) extracted from thin blood smears. Further study by applying the different types of neural netor architectures combining ith the other learning algorithms can be done in order to find the most appropriate netor for classification malaria parasite. References [] Citing internet sources URL: ses/malaria_ disease.htm [] Citing internet sources URL: e/malaria.htm [3] Jayavanth, S., Singh, M. (003), Artificial neural netor analysis of malaria severity through aggregation and deformability parameters of erythrocytes, Clinical Hemorheology and Microcirculation, 9(3-4), [4] Gao, C.Y., Xiong, H.Y., Yi, D., Chai, G.J., Yang, X.W., Liu, L. (003), Study on meteorological factors-based neural netor model of malaria, Zhonghua liuxingbingxue zazhi, 4(9), [5] Bender, A.a, Van Dooren, G.G.b, Ralph, S.A.b, McFadden, G.I.b, International Journal of the Computer, the Internet and Management Vol. 6. No. (January-April, 008) pp

7 Schneider, G. A. (003. Properties and prediction of mitochondrial transit peptides from Plasmodium falciparum Molecular and Biochemical Parasitology,). 3(), [6] Cybeno, G. (989), Approximations by superposition of a sigmoidal function, Mathematics of Control, Signal and Systems,, [7] Funahashi, K. (989). On the approximate realisation of continuous mappings by neural netors, Neural Netors,, [8] El-Fallahi, A.A., Martí, R.A., Lasdon, L.B. (006). Path relining and GRG for artificial neural netors, European Journal of Operational Research, 69(), [9] Mashor, M.Y. (003). Modified Recursive Prediction Error Algorithm for Training Layered Neural Netor, International Journal of the Computer, the Internet and Management, (), [0] Battiti, R. (99). First and secondorder methods for learning beteen steepest descent and Neton s method, Neural Computation, 4, [] Dong Wang, Wei-Zhen Lu (006). Forecasting of ozone level in time series using MLP model ith a novel hybrid training algorithm, Atmospheric Environment, 40(5), [] Citing internet sources URL: pdes/help/toolbox/ nnet [3] Bishop, C.M. (004). Neural Netor For Pattern Recognition, Oxford University Press. 5