You are on page 1of 5

2009 International Conference on Advances in Computing, Control, and Telecommunication Technologies

Automated Detection of Brain Tumor in EEG Signals Using Artificial Neural Networks
M. Murugesan
Assistant Professor, Department of EEE Syed Ammal Engineering College, Ramanathapuram Anna University, Trichy mmurugesanphd@gmail.com
Abstract Electroencephalograms (EEGs) are progressively emerging as a significant measure of brain activity and they possess immense potential for the diagnosis and treatment of mental and brain diseases and abnormalities. EEGs are up-andcoming as a vital methodology to suit the increasing global demand for more affordable and effectual clinical and healthcare services, with fitting interpretation methods. This research paper presents an automated system for efficient detection of brain tumors in EEG signals using artificial neural networks (ANNs). The ANN employed in the proposed system is feedforward backpropagation neural network. Generally, the EEG signals are bound to contain an assortment of artifacts from both subject and equipment interferences along with essential information regarding abnormalities and brain activity (responses to certain stimuli). Initially, adaptive filtering is applied to remove the artifacts present in the EEG signal. Subsequently, generic features present in the EEG signal are extracted using spectral estimation. Specifically, spectral analysis is achieved by using Fast Fourier Transform that extracts the signal features buried in a wide band of noise. The clean EEG data thus obtained is used as training input to the feedforward backpropagation neural network. The trained feedforward backpropagation neural network when fed with a test EEG signal, effectively detects the presence of brain tumor in the EEG signal. The experimental results demonstrate the effectiveness of the proposed system in artifacts removal and brain tumor detection. Keywords Adaptive filtering, Artifacts, Brain, Brain tumor, Electroencephalogram (EEG), Electrooculogram (EOG), Fast Fourier Transform (FFT), Spectral analysis, Artificial Neural Networks (ANNs), Feedforward backpropagation neural network.

Dr. (Mrs.) R.Sukanesh


Professor Department of ECE, Thiagarajar College of Engineering Madurai. principal and most prevalent artifact in EEG analysis, generated by eye movements and/or blinks [3]. In recent times, it has become tremendously simple to proficiently eradicate the eye-blink artifacts without affecting the underlying brain activity [4]. Moreover, the foremost significance for physicians and researchers to exemplify the cerebral activity are, the potentials arising from independent neurons inside the brain and not their superposition. Thus, a better solution would be to determine the signals of interest from the EEG recorded from the scalp using suitable signal processing techniques [5]. The pattern of electrical activity obtained from the EEG signal is chiefly valuable as well utilized to examine other conditions that might affect brain function, like head injuries, brain tumors or bleeding on the brain (hemorrhage). One important application of EEG signals is the detection of brain tumors. Brain tumor is the uncharacteristic growth of cells inside the brain or the skull that can either be cancerous or non-cancerous (benign) [6]. A very important aspect in clinical practice is the early detection and classification of brain tumors. Several researchers have proposed assorted techniques for the classification of brain tumors based on varied sources of information [7, 8, and 9]. Diagnosis of brain tumor in most cases necessitates experienced physicians and is a nontrivial, time-consuming task, irrespective of the techniques established thus far for the extraction of information, analysis and visualization of EEG. The chief reasons are [10]: Virtually very little symptoms: EEG is a propagation of internal electrical activity onto the scalp at a macroscopic level and might mislay accurate information and localization of a tumor. Moreover, features of globally connected brain activity differ from case to case. Variations in functionality and anatomy of brain tissues justify the reason behind variability of symptoms among subjects. Changing recording conditions, the existence of external stimuli and intricacy of brain activity lead to non stationary EEG recordings over different sessions.

I. INTRODUCTION Electroencephalogram (EEG) can be defined as a representative signal comprising of information corresponding to electrical activity produced by the cerebral cortex nerve cells, and has been one of the extensively employed signal for clinical assessment of brain activities and the detection of epileptiform discharges in the EEG [1]. For years now, the application of EEG in neurosurgery has been focused on the identification of brain lesions. EEG recorded by the scalp electrodes is a superposition of an enormous amount of electrical potentials produced from a number of sources (including brain cells i.e. neurons and artifacts) [2]. Among the artifacts present, the Electrooculogram (EOG) signal is the

This research paper is an enhancement of our previous work [20] that performed brain tumor detection in EEG signal
284

978-0-7695-3915-7/09 $26.00 2009 IEEE DOI 10.1109/ACT.2009.77

using Support Vector Machines (SVM). Here, we present an efficient system to perform early detection of brain tumors from EEG signals, with the aid of Artificial Neural Networks (ANNs). The proposed research makes use of the feed-forward backpropagation neural network that has proved largely to provide better performance in performing classification of the functional classes of EEG signal. Generally, EEG records bear information about abnormalities or responses to certain stimuli in the human brain. In addition, EEG signals are extremely contaminated with an assortment of artifacts. Artifacts in EEG records are caused by various factors, like line interference, EOG and ECG (electrocardiogram). These EEG artifacts augment the intricacy of EEG analysis and thus hinders in obtaining effective clinical information. The aforesaid problem necessitates a preprocessing step that removes artifacts in EEG records. To start with, the proposed research employs adaptive filtering to eliminate the artifacts present in the EEG signal. In adaptive filtering, an EOG signal is used as a reference input, which subtracts the EOG artifacts from inputted EEG signals. Subsequently, the generic features present in an EEG signal are extracted with the aid of spectral analysis. Particularly, FFT is used in spectral analysis for extracting the EEG signal features buried in a wide band of noise. The set of features extracted from the clean EEG data are fed as input to the feedforward backpropagation neural network to achieve effective brain tumor detection. With organized training and testing of the feedforward backpropagation neural network, the proposed system effectively detects the EEG signal with brain tumor. The proposed system enables the early detection of brain tumors in the EEG signals. The paper is organized as follows: In Section II, the proposed system for detection of brain tumors from EEG signal using feedforward backpropagation neural network is detailed. Section III presents the experimental results of the proposed system and section IV sums up the conclusion. II. PROPOSED SYSTEM FOR BRAIN TUMOR DETECTION IN EEG SIGNALS An extensive survey over the literature [7, 10 - 17] shows that there have not been a lot of effectual schemes for brain tumor detection from EEG signals probably because of a number of difficulties, 1) Symptom variability among diverse subjects and exhibition of varied brain electric behavior for different lesions and 2) Symptoms being most reliable under gliomas and less obvious under brain stem tumors. Recently, Neural Networks (NNs) has been extensively utilized as pattern and statistical classifiers in biomedical engineering. The proposed research devises an efficient and automated brain tumor detection system for EEG signals, by employing feedforward backpropagation neural network. The proposed research is composed of three phases namely; 1) Artifacts removal, 2) Feature extraction and 3) Brain tumor detection using feedforward backpropagation neural network. Primarily, artifacts are removed from the EEG Signals by means of adaptive filtering. The data are then convolved with

the filter coefficients to obtain the filtered output. Subsequently, features are extracted by applying Fast Fourier Transform (FFT) on the filtered data. The spectral analysis of EEG signal outputs the range of the frequency band. The clean or artifact removed EEG signal attained is fed to the feedforward backpropagation neural network for organized training and testing. Based on the degree of training, the feedforward backpropagation neural network accurately identifies those EEG signals with abnormal features, indicating the presence of brain tumor. A. Artifacts Removal The disturbances that are viable to appear in signal acquisition are termed as artifacts and can considerably affect the analysis of the signals. EEG (electroencephalogram) artifacts include line interference, EOG (electro-oculogram) and ECG (electrocardiogram). Acquiring effective clinical information from EEG analysis is largely influenced by the noise sources present in them. Hence, it is necessitated to devise specific filters to remove such artifacts in EEG records. The proposed system makes use of adaptive filtering for artifacts removal in EEG signals. The objective of adaptive filtering is to alter (adapt) the coefficients of the linear filter, and thus its frequency response, to generate a signal similar to the noise found in the signal, for filtering. The adaptive filtering process comprises of the minimization of a cost function that is utilized to determine the filter coefficients. Commonly, the adaptive filter tunes its coefficients to minimize the squared error between its output and the primary signal. In stationary conditions, the filter automatically converges to a Wiener solution. Whereas, under non-stationary conditions, the coefficients will vary with the timing of signal variation, thus, converging to an optimum filter [10]. Figure1. Exemplifies the typical block diagram of an EOG noise canceller using adaptive filtering.

Figure 1. EOG noise canceller system using adaptive filtering

s (n) is modeled as a combination of background EEG signal x (n) and EOG artifacts z (n) . The EOG signal r (n) picked up by an
The input to the proposed system electrode is fed as reference input to the system. The reference input and the noise component of EEG signal are correlated in some peculiar way. Here, h(m) represents a finite impulse response (FIR) filter of length M . The noise canceller generates an output signal by adaptively modifying the coefficients of the filter, where e( n) is an estimation of background EEG

x(n) [16].
(1)

e(n) = s(n)r (n)

285

where
r (n) =

h(m )r (n + 1 m )
m =1

(2)

B. Feature Extraction The primary confront in EEG analysis is the need for extracting the apt features or verifying what is "applicable information" from an EEG that strictly associates to pathological cases with high rate of inter as well as intra-rate agreement. Methods for spectral analysis are the principal source of solving the convolution of generic feature extraction "embedded" in an EEG signal. Spectral Analysis on EEG signals is performed presuming short time stationary and study energy content of EEG channels in Conventional frequency bands espoused by human raters physicians (Delta: below 3.5 Hz, Theta: 4 -7.5 Hz, Alpha: 8-13 Hz and Beta: 13 - 40 Hz). The proposed system employs Fast Fourier Transform (FFT) for spectral analysis. FFT's are generally used to determine the frequency components of a signal buried in a noisy time domain signal. The functions Y = FFT ( x) and

composed of two layers, 1) hidden/first layer and 2) Output/second layer. The hidden layer has a tansigmoid (tansig) activation function, and the output layer, has a linear activation function. Therefore, the first layer restricts the output to a finer range, from which the linear layer can generate all values. The output of each layer can be denoted by (4) Y = f ( W NXM X M ,1 + b N ,1 ) where Y is a vector consisting of the output from each of the N neurons in a given layer, W is a matrix consisting of the weights for each of the M inputs for all N neurons, X is a vector consisting of the inputs, b is a vector consisting of the biases and f () is the activation function [18]. The structure of a neuron in a backpropagation network is given below (Figure 2).

y = IFFT ( X ) implement the transform and inverse transform pair given for vectors of length N by:
X (k ) =

x ( j )
j =1

Figure 2. Structure of a Neuron in Backpropagation Network


( j 1 )( k 1 ) N

The training process of a backpropagation network is composed of two steps that are applied alternately [19]. The working of the backpropagation step involves,

x ( j ) = (1 / N ) X ( k ) N( j 1)( k 1) k =1

(3)

1)

Calculation of the error in the gradient descent and Backward Propagation to each neuron in the output layer and then hidden layer.

2)
th

where , is a N root of unity. The absolute band power for prominent EEG spectral bands (Delta: 1-4 Hz, Theta: 4-8 Hz, Alpha: 8-13 Hz and Beta: 13-30 Hz) is computed following the application of the Fast Fourier Transform. C. Brain Tumor Detection Using Feedforward Backpropagation Neural Network In the proposed research, effective detection of brain tumors from EEG signals is accomplished by an ANN, feedforward backpropagation neural network. The input to the feedforward backpropagation neural network is clean (artifacts removed) EEG signals. The processed EEG signals are categorized into two subdivisions namely, training data and test data. The data corresponding the training phase are used to train the feedforward backpropagation neural network for detecting possible cases of brain tumor. For testing, the input to the feed-forward backpropagation network is an artifacts removed EEG signal with/without tumor instances. The feedforward backpropagation neural network examines the EEG signals based on the trained data and if any signs of tumor are identified, then the neural network notifies the presence of tumor in the EEG signal, or else not. 1) Feedforward Backpropagation Neural Network: A feedforward backpropagation neural network is chiefly

N = e ( 2 i ) / N

In the second step, 1) Recomputation of the weights and biases and

2) Forward Propagation of the output of the activated


neurons, from the hidden layer to the output layer. III. EXPERIMENTAL RESULTS The results obtained from experimentation of the proposed system are presented in this section. The proposed system is programmed in MATLAB (Matlab 7.4). The proposed system takes as input an EEG signal recorded from the patients. EEG signals in Fig. 3 & 4 (a) represent the signals obtained with and without brain tumor respectively. Initially, adaptive filtering and FFT spectral analysis are applied as preprocessing steps to the EEG signal and the results obtained are shown in Fig. 3 (b) & (c) and Fig. 4 (b) & (c) respectively, where Fig. 3 corresponds to normal EEG signals without brain tumor and Fig.4 corresponds to abnormal EEG signals with brain tumor. Fig. 3 & 4 (d) exemplifies the feature extraction process on the filtered data using spectral analysis. Afterwards, the clean data obtained is fed as training input to the feedforward backpropagation neural network. The experimental results portray the effectiveness of the proposed system in artifacts removal and detection of brain tumor from the EEG signals.

286

(a)

(b)

(c)

(d)

Figure 3. Output representation of brain tumor detection using EEG signal (a) Normal Input EEG signal (b) Artifacts removed EEG signal (c) Spectral Analysis Using FFT (d) Features extracted using spectral analysis

(a)

(b)

(c)

(d)

Figure 4. Output representation of brain tumor detection using EEG signal (a) Abnormal Input EEG signal (b) Artifacts removed EEG signal (c) Spectral Analysis Using FFT (d) Features extracted using spectral analysis

One important aspect of any system devised for clinical diagnosis is its ability or rate of accurately detecting the presence or absence of a medical phenomenon. The proposed system was evaluated with 325 samples of EEG data recorded

from patients. Of which, 163 samples correspond to EEG data with brain tumor and the remaining 162 samples correspond to EEG data without brain tumor. The feedforward backpropagation neural network employed is trained and

287

tested with those 325 samples of EEG data. The detection rate obtained from the evaluation is tabulated in Table.1. In general, neural networks have been touted as having tremendous potential for improving diagnostic accuracy in patient specific data. The detection accuracy determined shows that feedforward backpropagation neural networks are computationally more efficient once trained.
TABLE I. DETECTION ACCURACY Detection Accuracy (%) 325 Samples of EEG Data Normal Abnormal 94.4785 98.7654

IV. CONCLUSION In this research paper, we have devised an automated system for efficient detection of brain tumor in EEG signals using artificial neural networks (ANNs). The proposed system has taken an EEG signal with artifacts as input. Firstly, the inputted EEG signal is subjected to artifacts removal by means of wiener filtering. Then, fearures of interest for brain tumor detection are extracted from the EEG signal. Subsequently, the ANN employed, feedforward backpropagation neural network, is trained with the features extracted from the EEG signal. Finally, when a EEG signal is fed as test input, the trained feedforward backpropagation neural network has detected the presence of brain tumor in the test signal. The experimental results have demonstrated the effeciency of the proposed system to perform brain tumor detection. REFERENCES
[1] K. Lehnertz, F. Mormann, T. Kreuz, R. Andrzejak, C. Rieke, P. David, and C. Elger, Seizure prediction by nonlinear EEG analysis, IEEE Engineering in Medicine and Biology Magazine, 2003. [2] Prabhakar K. Nayak, and Niranjan U. Cholayya, "Independent component analysis of Electroencephalogram", IEE Japan Papers of Technical Meeting on Medical and Biological Engineering, Vol.MBE-06, No.95115, pp.25-28, 2006. [3] S. Sanei and J. A. Chambers, EEG Signal Processing, John Wiley & Sons, New York, NY, USA, 2007. [4] Kianoush Nazarpour, Hamid R. Mohseni, Christian W. Hesse, Jonathon A. Chambers, and Saied Sanei, "A Novel Semi blind Signal Extraction Approach for the Removal of Eye-Blink Artifact from EEGs", EURASIP Journal on Advances in Signal Processing, Article ID 857459, 12 pages, 2008. [5] M. Ungureanu, C. Bigan, R. Strungaru, V. Lazarescu, "Independent Component Analysis Applied in Biomedical Signal Processing", Measurement Science Review, Volume 4, Section 2, 2004. [6] Steven Pacia, MD, "Brain Tumors and Epilepsy", Fight against childhood epilepsy & seizures. [7] Rosaria Silipo, Gustavo Dero, Helmut Bartsch, "Brain tumor classification based on EEG hidden dynamics", Intelligent Data Analysis, Vol. 3, No.4, pp. 287-306, October 1999. [8] C. Majs, M. Juli-Sap, J. Alonso, M. Serrallonga, C. Aguilera, J. Acebes, C. Ars and J. Gili, Brain Tumor Classification by Proton MR Spectroscopy: Comparison of Diagnostic Accuracy at Short and Long TE, American Journal of Neuroradiology 25:1696-1704, NovemberDecember 2004. [9] L. Lukas, A. Devos, J.A.K. Suykens, L. Vanhamme, F.A. Howe, C. Majos, A. Moreno-Torres, M. Van Der Graaf, A.R. Tate, C. Arus, and S. Van Huffel. Brain tumor classification based on long echo proton MRS signals. Artificial Intelligence in Medicine: 31, 7389, 2004.

[10] Fadi N. Karameh, Munther A. Dahleh, "Automated classification of EEG signals in brain tumor diagnostic", Proceedings of the American control conference, Chicago, Illinois, June 2000. [11] T. P. Jung, S. Makeig, M. Westerfield, J. Townsend, E. Courchesne, & T. J. Sejnowski, Removal of eye activity artifacts from visual eventrelated potential in normal and clinical subjects, Clin. Neorophysiol., vol. 11, pp 1745-1758, 2000. [12] W.T. Blume, M. Kaibara, and G.B. Young, Atlas of Adult Electroencephalography. Lippincott Williams and Wilkins, Philadelphia, 2002. [13] Shane M. Haas, Mark G. Frei, Ivan Osorio, Bozenna Pasik- Duncan, & Jeff Radel, EEG ocular artifact removal through ARMAX model system identification using extended least squares, Communications in Information and Systems, 3, (1), pp 19-40, 2003. [14] Alexander V. Kramarenko, Uner Tan, "Effects of High-Frequency Electromagnetic Fields on Human EEG: A Brain Mapping Study", International Journal of Neuroscience, Vol. 113, pp. 10071019, 2003. [15] M.A.Klados, C. Papadelis1, C.D. Lithari and P.D. Bamidis, "The Removal of Ocular Artifacts from EEG Signals: A Comparison of Performances for Different Methods", IFMBE Proceedings, Eds.): ECIFMBE 2008, Vol. 22, pp. 12591263, 2008 [16] Parisa Shooshtari, Gelareh Mohamadi, Behnam Molaee Ardekani, Mohammad Bagher Shamsollahi, "Removing Ocular Artifacts from EEG Signals using Adaptive Filtering and ARMAX Modeling", Proceedings Of World Academy Of Science, Engineering And Technology, Vol. 11, pp. 277-280, February 2006 [17] Siew Cheok Ng and P. Raveendran, "Removal of EOG Artifacts Using ICA Regression Method", In Proceedings of Biomed 2008, N.A. Abu Osman, F. Ibrahim, W.A.B. Wan Abas, H.S. Abd Rahman, H.N. Ting (Eds.):, Vol. 21, pp. 226-229, 2008. [18] J. A. Freeman and D. M. Skapura, "Neural networks: algorithms, applications and programming techniques", Addison Wesley Longman, pp. 89-105, 1991. [19] A. L. Betker, T. Szturm, Z. Moussavi, "Application of Feedforward Backpropagation Neural Network to Center of Mass Estimation for Use in a Clinical Environment", Proceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Vol. 3, pp. 2714-2717, 17-21 Sept. 2003. [20] M. Murugesan and Dr. (Mrs.).R. Sukanesh, "An Effective System for Brain Tumor Detection via Electroencephalogram Signals and Support Vector Machines", International Journal of Computer Theory and Engineering (IJCTE), Vol. 1, No. 5, December 2009 [accepted for publication].

288

You might also like