Discrimination of Visual Evoked Potentials Using Image Processing of Their Time-scale Representations

Procedia Technology 17 ( 2014 ) 359 – 367

ScienceDirect

2212-0173 © 2014 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).

Peer-review under responsibility of ISEL – Instituto Superior de Engenharia de Lisboa, Lisbon, PORTUGAL. doi: 10.1016/j.protcy.2014.10.180

Conference on Electronics, Telecommunications and Computers – CETC 2013

Discrimination of Visual Evoked Potentials Using Image Processing

of Their Time-Scale Representations

El-Mehdi Hamzaoui

* Fakhita Regragui

a _{LIMIARF Laboratory, University Mohammed V-Agdal, Faculty of Sciences of Rabat, Morocco}

Abstract

In this article, the continuous wavelet transform, based on the Mexican Hat function, is used to achieve an automatic processing and classification of visual evoked potentials. Indeed, the representations of the modulus of the wavelet coefficients in the time-scale plan, allowed us to define qualitative criterion for the discrimination of normal and pathological cases. In order to enhance the classification rate, we developed a new method that considers the visual evoked potential's scalogram as an image which is segmented, according to the maximal level of energy density, into many regions of significant interest. The analysis of the resulting segmented image permits to extract a vector of most significant features that will be used to classify normal and pathological signals through SVM classifier. The obtained results show that this technique not only, allows a more reliable distinction between normal and pathological cases, with a very high classification rate (93%) in comparison with the one of the conventional latency measurement (67%), but also, suggests that indications on the progression of the pathology can be provided. © 2014 The Authors. Published by Elsevier Ltd.

Selection and peer-review under responsibility of ISEL – Instituto Superior de Engenharia de Lisboa.

Keywords: Classification; Mexican Hat; Support Vector Machines; Visual Evoked Potential; Continuous Wavelet Transform.

1. Introduction

The advent of technology and the non-invasiveness of the visual evoked potential (VEP) recordings have encouraged their clinical use for the diagnosis of certain brain diseases. These potentials are elicited by the brain when visual stimulations are successively applied to a subject. In the clinical practice, interpretation of the VEP is based on the visual reading of the latency which corresponds to the time of occurrence of the first major peak (P100)

* Corresponding author.

E-mail address: [email protected]

© 2014 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).

[1], [2], [3], [4]. These signals are very weak in amplitude and completely embedded in noise. The averaging technique is commonly used to extract the VEP meaningful signals. The concept is to average most of the signals recorded during the test until a clean plot of the VEP is obtained. This often requires the acquisition of several single brain responses which yield tiredness and discomfort in the subject [3], [4].

Seeking more accuracy and efficiency of the VEP as a clinical tool, researchers tried to develop other processing and analysis approaches. In terms of extraction, a number of authors proposed linear and nonlinear adaptive filtering as alternative to the averaging technique. Indeed, Orfanidis, S.J. et al., used a linear predictor, based on least square error algorithm, for modeling VEP signals [5]. Drissi, H. et al., showed that good plots of VEP are obtained while applying a linear adaptive filter to a few VEP signals [6]. Davila, C.E. et al., proposed a matched subspace filter for detecting multi-harmonic VEP's [7]. In most recent works, Hamzaoui, E. et al. proposed a multilayer perceptron as nonlinear adaptive filter, which permits a significant improvement of the signal to noise ratio (SNR) [8]. All these developed methods and others [9], [10],[11], are very time saving and very convenient for children, elders or people that cannot stay still. For VEP data analysis and classification purposes, spectral analysis methods were used and various types of feature vectors were proposed to model the VEP such as those obtained with linear predictive and mathematical methods [3], [12], or those based on parametric and non-parametric spectral methods [13], [14]. Recently, researchs focus on the use of time-frequency techniques for VEP spectral decomposition and filtering [15], [16], [17], [18], [19] and the artificial neural networks for VEP classification [20], [21], [22].

In this paper, we consider the time-scale representation of the VEP as an image which exhibits regions of significant interest. The segmentation of this image permits us to extract significant parameters that should enhance discrimination rate, between normal and pathological recordings, with respect to the one obtained with the latency method.

2. Materials and methods

The schematic diagram of figure 1 summarizes the approach we followed in our tests. The proposed method has been implemented using the MATLAB® _{8.1 toolboxes [23]. Indeed, the recoded VEP is analysed using continuous}

wavelet transform which the graphical result is considered as an image. The processing of this image allows us to extract vectors of features that will be used to classify the signals in two or more classes according to the pathology.

2.1. VEP recording

In this work, we tested 88 averaged VEP signals recorded from 22 subjects (10 normal and 12 pathological subjects with well established multiple sclerosis). The VEP data were elicited using 4 colored reversal checkerboard pattern stimulus reversed every 512 milliseconds. The color combinations used in the test were: white, black-blue, black-red and black-yellow. The checks subtended a visual angle of 1.32° and the entire pattern 10.5° corresponding to a distance of 1m from the subject [3]. The VEP signals are measured using two electrodes placed on the scalp at the Cz an Oz positions according to the international system 10-20 [24], [25]; a third electrode connected to the earlobe is used as reference. The potentials are first amplified (x20000) and converted through an A/D converter (12 bits, sampling frequency of 1024 Hz). Each set of data were filtered using a band pass filter with a bandwidth of [0.3 Hz, 400 Hz], [3], [4]. For each subject and each stimuli, 64 signle VEPs were obtained and averaged to improve the signal to noise ratio [3].

2.2. Wavelet analysis of VEP

Spectral analysis of the VEP signals is performed by computing their continuous wavelet transform. Indeed, the wavelet transform S(α,β) of the signal s(t) is given by the equation (1) below [26]:







t

s

t

dt

S

(

,

)

1

(

_

)

.

(

)











α and β represent respectively the scaling and dilation factors, whereas ψ1 is the wavelet mother function.

The scale α is associated to the frequency by:





*



F

f

where Fc is the central frequency of the wavelet (in Hz) and Δ is the sampling period.

More recent works proposed the use of wavelet transform for the analysis of diferent types of evoked potentials. Most of them use the first and second derivative of the Gaussian function as the mother wavelet, since these functions have similar profile to a typical VEP waveform [17], [18]. In our study, ψ1 is chosen to be the Mexican Hat

function, which is calculated as the second-order derivative of the Gaussian function [26]. It is expressed by the following equation (2) and illustrated by the figure 2 hereafter.

)

²

2

exp(

)

1

²

(

3

2

)

(

_

_







t



t





t

Figure 2: Plot of the Mexican Hat Wavelet.

The wavelet transform defines a local time-frequency energy density of the analysed signal s(t). Indeed, the modulus of the wavelet transform coefficients measures the energy of s(t) in the Heisenburg box of each wavelet ψα,β

centered at (α,f/β), where f is the frequency. The projection of this energy values on the time-scale plan is called: scalogram [26]. In this graphical representation, the ordinate axis represents log2(α) and the abscissas one denotes

the time. A colour map is commonly used to quantify the energy density levels. In this work, we used a hot colormap in which colours vary smoothly, according to the values of energy density, from black (low energy density) to white through shades of red, orange, and yellow. The white colour corresponds to the maximal level of the energy [23]. 2.3. Classification of VEPs

The time-scale representation of the VEP signals is considered as an image which we transform into a binary one using the Otsu method. This segmentation method requires prior calculation of the histogram of the image to define the best value of the threshold. This operation results into a black and white image. The white colour corresponds to the maximal value of the signal's energy [27], [28]. From the segmented image, we extract, for each normal VEP, a number of geometrical features such as the number of detected regions, their areas, height and width, the coordinates of their centres according to scale and time axis, the Euclidian distance between two adjacent centres, the coordinates of the midpoint of the segment formed by two nearest centres (see figure 3), … These parameters will form the feature vector to be used as the classifier input in the discrimination step.

To assess the relevance of the extracted parameters for VEP discrimination, we used a support vector machine classifier (SVM) which is widely used to solve discrimination and regression problems [29], [30]. In the proposed approach, we used, a linear kernel to map the training set and the sequential minimal optimization method to detect the optimal separating hyper plane[29], [30]. The training set includes the features extracted from the analysed VEPs belonging to the normal population.

Figure 3: Example of extracted features used to discriminate VEP signals.

-5 -4 -3 -2 -1 0 1 2 3 4 5 -0.4 -0.2 0 0.2 0.4 0.6 0.8

1 Mexican Hat wavelet

The variable t W a ve le t c oe ffi ci e nts

3. Results and discussion

In a first time, we performed various tests to define the best values of the wavelet width δ and the central frequency Fc of the Mexican Hat wavelet. Then, we computed and plotted the energy of the signal, provided by the wavelet transform, versus δ and Fc. We found that the energy reaches its maximal value for δ=0.47 and 2πFc=5.34 Hz (see figure 4). Also, high values of the energy density are gotten when the scale number α is chosen to be 128.

Results of the application of wavelet transform to VEP signals are illustrated in the figure 5. Examination of the scalograms obtained in the normal case reveals two regions of high intensities around 100 and 150 milliseconds for high scales (low frequencies). These two regions are separated by a black vertical line that occurs through all scales in the time range [105-120 msec]. These zones of maximum densities of energy may match the two first negative and positive peaks of the normal VEP. The black line can explain the polarity changes in the signal. A third area of high energy is also found in the time range [270-320 msec] which may correspond to the P300 component of the VEP. In the pathological cases, the wavelet representations show high energy concentrations within ranges all over the time axis. These concentrations are observed for medium and high scales depicting smaller changes throughout the signal. Based on these results found by processing the entire dataset, the maximum energy localization in time and scale allowed us to define qualitative criterion for the VEPs discrimination [4], [31].

(a)

(b) (c)

Figure 4: (a) 3D representation of the maximal values of the VEP's energy density versus the wavelet width δ and the central frequency Fc. (b) Projection of the maximal values of the energy density on the central frequency Fc axis. (c) Projection of the maximal values of the energy

density on the wavelet width δ axis.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2 4 6 8 10 120 0.2 0.4 0.6 0.8 1 1.2 1.4

W avelet width delta Maximum Energy density Vs wavelet width delta and Central frequency Fc

W avelet central frequency

Ma x imu m e ne rg y d en s it y 0 2 4 6 8 10 12 0 0.2 0.4 0.6 0.8 1 1.2

1.4 Maximum Energy density Vs and Central frequency Fc

Wavelet central frequency

Ma xi mu m e ne rg y d en si ty 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.2 0.4 0.6 0.8 1 1.2

1.4 Maximum Energy density Vs wavelet width delta

W avelet width delta

Ma xi mu m e ne rg y d en si ty

Figure 5: Plot of averaged VEP and their corresponding wavelet transform. Left: Normal case. Right: Pathological case.

The second tests were achieved to choose the best value of the binarization threshold. This value is automatically calculated to reach the minimum of the intraclass variance of the black and white pixels. The results show that a value of 0.2 provides an optimal segmentation of the image with a minimum loss of information. Indeed, for the normal subjects, we were able to isolate, the three regions of interest cited above. Each region is characterized by a vector of the most stationary parameters that will be used for the SVM training phase. In particular, we found that three parameters are reproducible for all analysed VEPs:

 Coordinates of each region's centre: [Cix, Ciy];  Euclidian distance between these centres: Dij;

 Coordinates of the midpoint of segments formed by two adjacent region's centres [Mijx, Mijy].

In order to evaluate the effectiveness of the proposed method, we compared its classification rates with those obtained using the conventional measurement of latency [3], [6], [14]. Indeed, for all the used stimuli, a significant improvement of classification rate is reached as shown in the table 1. The graphical results of VEP classification using the CWT-SVM method are illustrated in the following figure 6.

Table 1. VEP classification rate obtained using the latency measurement and the CWT-SVM methods.

Color of stimuli Latency Method CWT-SVM Method

Black – White 61,56 90,92

Black – Red 63,25 91,82

Black – Blue 65,69 92,74 Black – Yellow 67,82 93,27

In particular, the use of checkerboard black-yellow, not only, allows a more reliable distinction between normal and pathological VEPs with a very high classification rate (93%), but also, permits a discrimination among pathological signals. This suggests that the technique may provide indications on the stage or disease progression. 4. Conclusion

The continuous wavelet transform of the visual evoked potentials is performed using the Mexican hat function. Examination of the scalograms obtained in the normal case, reveals three regions of high intensities around 100, 150 and 300 milliseconds for high scales (low frequencies). In the pathological cases, the scalograms show high energy concentrations within ranges all over the time axis. These results allowed us to define qualitative criterion for VEPs discrimination. 50 100 150 200 250 300 350 400 450 500 -10 -5 0 5 10 Am p lit ud e ( μ V) Time (msec) Avraged VEP (Normal case)

Sc al e Time (msec) CWT Coefficients (Moduli) 50 100 150 200 250 300 350 400 450 500 20 40 60 80 100 120 50 100 150 200 250 300 350 400 450 500 -5 0 5 A m p lit ude ( μ V) Time (msec) Avraged VEP (Pathological case)

Sc al e Time (msec) CWT Coefficients (Moduli) 50 100 150 200 250 300 350 400 450 500 20 40 60 80 100 120

Aiming to improve the VEPs classification rates according to the pathology, we considered each scalogram as an image which we transform into a binary and segmented one using the Otsu method. From the resulting image, we extract a vector of three features which were found to be the most stationary and relevent to characterize a normal VEP. Based on this vector, the classification of the signals is achieved through linear SVM classifier. The comparison of obtained classification rates with those of the conventional measurement of latency shows that a significant improvement of the classification rate is reached for all the used stimuli. In particular, using this method, we found that the black-yellow stimulus allows a very good discrimination between normal and multiple sclerosis cases with a very high rate (93%). The proposed method is revealed to be very efficient; it is based on only three main features related to the geometry of the patterns that the binarized scalogram contains. It permits also a discrimination among VEPs belonging to pathological population. We think that such results could be linked to stage of the disease.

Figure 6: VEPs discrimination using the proposed method for 4 colored stimuli: black-white, black-blue, black-red, and black-yellow.

References

[1] Chiappa K.H. Evoked Potentials in Clinical Medicine, Raven Press, NY, USA, 1983.

[2] Epstein C.M. Introduction to EEG and Evoked Potentials. Lippincott Williams & Wilkins, 1983.

[3] Regragui F., Extraction and Linear Prediction Modelling of the Visual Evoked Potential, PhD. Thesis, Rutgers University, USA, 1990. [4] Hamzaoui E-M., Identification and characterization of visual evoked potentials sources for a better clinical use, Ph.D. Thesis, Mohammed V-Agdal University, Faculty of Sciences, Rabat, Morocco, 2011.

[5] Orfanidis S.J., Regragui F., Micheli-Tzanakou E., Visual evoked potential extraction by Adaptive Filtering, in Proc. of IEEE 19th Annual Conference of the Engineering in Medicine & Biology Society, 2, 1987, p. 968-969.

[6] Drissi H., Regragui F., Essaid A., Estimation du Potentiel Evoqué visuel par la Technique du Filtrage Adaptatif, in Proc. of Colloque du Capteur à l'Instrumentation, April 1992, Rabat, Morocco.

[7] Davila C.E., Srebro R., Ghaleb I.A., Optimal Detection of Visual Evoked Potentials, IEEE Transactions on Biomedical Engineering, Vol.45, No. 6, 1998, p. 800-803.

[8] Hamzaoui E-M., Regragui F., Himmi M.M., Mghari A., Visual Evoked Potentials' Non Linear Adaptive Filtering Based on Three Layers Perceptron, Applied Mathematical Sciences, Vol. 4, 2010, no. 56, 2010, p. 2759 – 2772.

[9] Wang Z., Maier A., Leopold D.A., Logothetis N.K., Liang H., Single-trial evoked potential estimation using wavelets, Computers in Biology and Medicine, Vol. 37, No. 4, 2007, p. 463-473.

[10] Quiroga R.Q., Obtaining Single Stimulus Evoked Potentials with Wavelet Denoising, Physica D: Nonlinear Phenomena, Vol. 145, No. 3–4, 2000, p. 278-292.

[11] Friman O., Volosyak I., Gräser A., Multiple Channel Detection of Steady-State Visual Evoked Potentials for Brain-Computer Interfaces, IEEE Transactions on Biomedical Engineering, Vol. 54, No. 4, 2007, p. 742-750.

[12] Jansen B.H., Rit V.G., Electroencephalogram and Visual Evoked Potential Generation in a Mathematical Model of Coupled Cortical Columns, Biological Cybernetics, Vol. 73, No. 4, 1995, p. 357-366.

[13] Cerutti S., Baselli G., Liberati D., Pavesi G., Single Sweep Analysis of Visual Evoked Potentials Through a Model of Parametric Identification, Biological Cybernetics, Vol. 56, No 2-3, 1987. p. 111-120.

[14] Himmi M.M., Regragui F., El Khamlichi M., Ouazzani R, Résolution spectrale de la Méthode de Pisarenko dans le cas des Signaux non stationnaires, in Proc. of CSS'01 – Conférence Internationale en Communications, Signaux et Systèmes, March 2001, Rabat, Morocco.

[15] Drissi H., Regragui F., Bennouna M., Antoine J.P., Wavelet Transform analysis of Visual Evoked Potentials: Some Preliminary Results, ITBM-RBM, Vol. 21, No. 2, 2000, p. 84-91.

[16] Quiroga R.Q., Sakowitz O.W., Basar E., Schürmann M., Wavelet Transform in the analysis of the frequency composition of evoked potentials, Brain Research Protocols, Vol. 8, No. 1, 2001, p. 16-24.

[17] Borodina U.V., Aliev R.R., Wavelet spectra of visual evoked potentials: Time course of delta, theta, alpha and beta bands, Neurocomputing, Vol. 121, No. 9, 2013, p. 551-555.

[18] Thie J., Sriram P., Klistorner A., Graham S-L., Gaussian wavelet transform and classifier to reliably estimate latency of multifocal visual evoked potentials (mfVEP), Vision Research, Vol. 52, No. 1, January 2012, p.p 79-87.

[19] Bian Y., Li H., Zhao L., Yang G., Geng L., Research on Steady State Visual Evoked Potentials based on Wavelet Packet Technology for Brain-Computer Interface, Procedia Engineering, Vol. 15, 2011, p.p 2629-2633.

[20] Mghari A., Himmi M.M., Amaloud A., Regragui F., Visual evoked potentials discrimination based on adaptive zero-tracking neural network, Computers in Biology and Medicine, Vol. 36, No. 4, 2006, p. 408-418.

[21] Cecotti H., A time–frequency convolutional neural network for the offline classification of steady-state visual evoked potential responses, Pattern Recognition Letters, Vol. 32, No. 8, 2011, p. 1145-1153.

[22] Çelik, U., Arıca S., Classification of evoked potentials of familiar and unfamiliar face stimuli using multi-resolution approximation based on excitatory post-synaptic potential waveform, Computers & Electrical Engineering, Vol. 39, No. 5, 2013, p. 1571-1584.

[23] The Mathworks Inc, MATLAB® Products: User guides,available at: http://www.mathworks.com.

[24] Electrode Position Nomenclature Committee, 2006. Guideline 5: Guidelines For Standard Electrode Position Nomenclature. Journal of Clinical Neurophysiology, Vol. 11, p. 111-113.

[25] Odom, J.V., Bach, M., Brigell, M., Holder, G.E., McCulloch, D.L., Tormene, A.P., Vaegan, 2010. ISCEV Standard for Clinical Visual Evoked Potentials (2009 Update), Documenta Ophthalmologica, Vol. 120, No. 1, p.111–119.

[26] Mallat S., A Wavelet Tour of Signal Processing: The Sparse Way, 3rd Ed, Academic Press, 2008.

[27] Otsu N., A Threshold Selection Method From Gray-level Histograms, Automatica, Vol. 11, No. 285-296, 1975, p. 23-27.

[28] Liao P. S., Chen T. S., Chung P. C., A Fast Algorithm for Multilevel Thresholding. Journal of Information Science and Engineering, Vol. 17, No. 5, 2001, p. 713-727.

[29] Steinwart I., Christmann A., Support Vector Machines, Springer, 2008. [30] Wang L., Support Vector Machines: Theory and Applications. Springer, 2005.

[31] Hamzaoui E-M., Himmi M.M., Regragui F., Laraki K., Bouyakhf E.H., PEDiag II: A New Tool to Improve Visual Evoked Potentials Extraction and Interpretation, The 2nd International Symposium on Communications, Control and Signal Processing (ISCCSP 2006), 13-15 March 2006, Marrakech, Morocco.