Документ взят из кэша поисковой машины. Адрес оригинального документа : http://brain.bio.msu.ru/papers/Shishkin_Kaplan_Bakardjian_Cichocki_2005_IEPS8_CombiningExtremities_SourceLoc_proofs.pdf Дата изменения: Fri Nov 19 08:31:40 2004 Дата индексирования: Mon Oct 1 20:25:05 2012 Кодировка:

DTD 5

ARTICLE IN PRESS
International Congress Series xx (2004) xxx ­ xxx

www.ics-elsevier.com

1

Sergei L. Shishkina,*, Alexander Ya. Kaplanb, Hovagim Bakardjiana, 4 Andrzej Cichockia 5
Laboratory for Advanced Brain Signal Processing, RIKEN Brain Science Institute, 2-1 Hirosawa, Wako-shi, Saitama 351-0198, Japan b Department of Human and Animal Physiology, Faculty of Biology, Moscow State University, Russia
a

PROOF

Combining the extremities on the basis of separation: 2 a new approach to EEG/ERP source localization 3

6 7 8

UNCORRECTED

Abstract. Current methods for the localization of EEG and event-related potentials (ERP) sources assume that sources are either discrete (dipole-like) or distributed. While both types of sources are likely to contribute significantly to EEG and ERP signals, each method adopts only one of these models and thus may localize the sources of other type incorrectly or not find them at all. Recently introduced Independent Component Analysis (ICA) and more general approach, Blind Source Separation (BSS), make possible the separation of signals from various brain and extra-brain (related to artifacts) sources and can be used as preprocessing technique before applying the localizing algorithms. We suggest using this preprocessing step for combining different localization methods. A brain source, if extracted correctly, can be analyzed separately from the other sources, and thus, the most appropriate localization technique can be chosen for each source. Distributed sources are likely to be localized more precisely without detailed separation but after BSS bcleaningQ data from strong localized sources. D 2004 Published by Elsevier B.V.
Keywords: Source localization; Event-related potentials; EEG; Blind source separation; Independent component analysis

9 10 11 12 13 14 15 16 17 18 19 20 21 22
23

1. The two extreme approaches to brain source localization the most difficult numerous techniques Ref. [1] for review), types of the sources.

24 25
26 27 28 29

Spatial localization of the signal sources remains one of methodological problems in the analysis of EEG and ERP. While aimed at solving this problem have been developed up to date (see each of them, at best, can only provide correct localization of some
* Corresponding author. Tel.: +81 48 462 1111x7171; fax: +81 48 467 9694. E-mail address: shishkin@brain.riken.jp (S.L. Shishkin).

0531-5131/ D 2004 Published by Elsevier B.V. doi:10.1016/j.ics.2004.11.077

ICS-04798; No of Pages 4


ARTICLE IN PRESS
2 S.L. Shishkin et al. / International Congress Series xx (2004) xxx­xxx

PROOF

The type of the sources which are most efficiently localized by a specific method depends on the source model which is assumed. Typically, sources are supposed to be discrete (the signal is produced by a few dipoles) or distributed (in this case, thousands of points are considered as possible location of a current source). It is quite evident that the discrete source model often can be inadequate in describing widely distributed activation, especially when its spatial pattern is complicated. On the other hand, the localization results obtained with the distributed source models can be evidently biased by various strong a priori assumptions required when these models are used. For example, the Laplacian Weighted Minimum Norm Algorithm implemented in the popular LORETA software [2] generally provides a rather blurred solution, which is the result of the assumption of smoothness of the spatial distribution of the sources [1]. Intuitively, it seems that any possible type of distributed source model will lead to less exact localization of the strong localized sources than discrete models can provide for the same sources. Both discrete and distributed types of sources, however, can participate in the generation of EEG and ERP signal. 2. Blind source separation (BSS) as a tool for prelocalization separating EEG and ERP components related to different sources

30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71

UNCORRECTED

The EEG and ERP signal always include mixed activity from many brain sources, which makes the task of source localization especially difficult. It seems to be reasonable to suggest that signal components from distinct sources should exhibit relatively low correlation or statistical dependency and can be separated (although not yet localized) without any a priori assumptions by finding most uncorrelated or most independent components. Many techniques for such bblindQ source separation (BSS) have been developed recently and received growing interest not only in the analysis of EEG or MEG signals but also in many other areas of application [3]. The localization algorithm can be then applied to each of the separated sources. This approach is especially attractive for the application of discrete source model because only one or, in the case of bilateral symmetry, two dipoles should be fitted to each separated component, while the presence of noise (including activities of brain sources which do not fit the model) is minimized. To date, many simulation and real data studies demonstrated the usefulness of Principal Component Analysis (PCA) followed by rotation, BSS based on Second Order Statistics (such as implemented in SOBI algorithm), and Independent Component Analysis (ICA) for improving source localization [4­13]. All these BSS techniques, theoretically, may help to localize even very weak and strongly overlapped sources, and to locate much higher number of dipole-like sources than can be localized without such preprocessing because the number of the dipoles which can be fitted is limited, in the case that separation is performed perfectly, only by the number of separated sources, which is, in the case of ICA and many other BSS techniques, usually equal to the number of electrodes. 3. BSS as a tool for combining discrete and distributed source modeling To the best of our knowledge, no algorithm combining both discrete and distributed source modeling has been proposed yet. Moreover, it seems that no model exists which


ARTICLE IN PRESS
S.L. Shishkin et al. / International Congress Series xx (2004) xxx­xxx 3

could be in between these two extreme approaches and combine their advantages. The two models could be easily used in parallel, however, if both types of the sources which best fit each of them are correctly separated in preprocessing. After separating the sources with some of BSS algorithms, one can select the most appropriate model for each source. The performance of BSS algorithms on such complicated signals like EEG or ERP cannot be perfect. Application of distributed source model to the components separated with BSS has been studied too little yet to decide if BSS is really helpful in this case. Analysis of separated sources can be more problematic with distributed source modeling than with discrete source modeling. For example, a spatially extended source, which is formed by long-distance neuronal interactions, can be probably less homogenous and coherent than a strong local source and can be more easily split by the separating algorithm into several subsources. However, one simple solution can help in combined use of both models. Consider the procedure of cleaning ERP or EEG data from artifacts (e.g., eye blinking) using BSS, which is shown to be rather efficient [14,15]. In this procedure, the sources which can prevent correct estimation of the target activity are separated, identified, and then removed from the data; in other words, the ERP or EEG signal is reconstructed using only nonartificial components. Strong local brain sources can also significantly interfere with the modeling of distributed sources, and removing them from the data before distributed source analysis could be desirable. Therefore, for improving distributed source modeling, we can apply essentially the same procedure as in the case of removing artifacts: separate, identify, and remove strong local sources with BSS. The data bcleanedQ from such sources can be then analyzed with distributed source modeling. Unlike in the artifact removing procedure, we can also analyze these sources which are to be removed before distributed source analysis; of course, we can analyze them with discrete source modeling. One may hypothesize that spatially distributed sources are statistically more dependent or correlated than localized sources because the large neural networks may be more often driven by the input which is common with other large networks or even be directly connected to each other than distinct localized networks. If this is true, one should avoid the use of typical BSS technologies (based on decorrelation or independence criteria) for separating of distributed brain sources; the approach described above is in accordance with this understanding of the relations between brain sources. The sources separated by BSS have one attribute whose negative effect on EEG and ERP analysis is often underestimated: their spatial pattern does not change in time. However, after removing the discrete sources, the temporal variations of the data spatial pattern do not disappear, and the spatial dynamics of distributed sources can be studied as usual. Combination of BSS and discrete source modeling has already been proven to be, in general, practically efficient, but more studies are needed to choose the criteria for determining whether a component separated by BSS can be correctly modeled as a discrete source or should it be kept in the data for distributed source analysis.

72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116

UNCORRECTED

PROOF


ARTICLE IN PRESS
4 S.L. Shishkin et al. / International Congress Series xx (2004) xxx­xxx

References
[1] C.M. Michel, et al., EEG source imaging, Clin. Neurophysiol. 115 (10) (2004) 2195 ­ 2222. [2] R.D. Pascual-Marqui, C.M. Michel, D. Lehmann, Low resolution electromagnetic tomography: a new method for localizing electrical activity in the brain, Int. J. Psychophysiol. 18 (1) (1994) 49 ­ 65. [3] A. Cichocki, S. Amari, Adaptive Blind Signal and Image Processing: Learning Algorithms and Applications, John Wiley & Sons, West Sussex, 2002. [4] Z.J. Koles, J.C. Lind, A.C. Soong, Spatio-temporal decomposition of the EEG: a general approach to the isolation and localization of sources, Electroencephalogr. Clin. Neurophysiol. 95 (4) (1995) 219 ­ 230. [5] Z.J. Koles, A.C. Soong, EEG source localization: implementing the spatio-temporal decomposition approach, Electroencephalogr. Clin. Neurophysiol. 107 (5) (1998) 343 ­ 352. [6] L. Zhukov, D. Weinstein, C. Johnson, Independent component analysis for EEG source localization, IEEE Eng. Med. Biol. Mag. 19 (3) (2000) 87 ­ 96. [7] J. Cao, et al., Independent component analysis for unaveraged single-trial MEG data decomposition and single-dipole source localization, Neurocomputing 49 (2002) 255 ­ 277. [8] K. Kobayashi, I. Merlet, J. Gotman, Separation of spikes from background by independent component analysis with dipole modeling and comparison to intracranial recording, Clin. Neurophysiol. 112 (3) (2001) 405 ­ 413. [9] K. Kobayashi, et al., Systematic source estimation of spikes by a combination of independent component analysis and RAP-MUSIC: II. Preliminary clinical application, Clin. Neurophysiol. 113 (5) (2002) 725 ­ 734. [10] A.C. Tang, et al., Independent components of magnetoencephalography: localization, Neural Comput. 14 (8) (2002) 1827 ­ 1858. [11] J. Dien, K.M. Spencer, E. Donchin, Localization of the event-related potential novelty response as defined by principal components analysis, Brain Res. Cogn. Brain Res. 17 (3) (2003) 637 ­ 650. [12] S. Makeig, et al., Mining event-related brain dynamics, Trends Cogn. Sci. 8 (5) (2004) 204 ­ 210. [13] J.E. Richards, Recovering dipole sources from scalp-recorded event-related-potentials using component analysis: principal component analysis and independent component analysis, Int. J. Psychophysiol. 54 (3) (2004) 201 ­ 220. [14] T.P. Jung, et al., Removal of eye activity artifacts from visual event-related potentials in normal and clinical subjects, Clin. Neurophysiol. 111 (10) (2000) 1745 ­ 1758. [15] C.A. Joyce, I.F. Gorodnitsky, M. Kutas, Automatic removal of eye movement and blink artifacts from EEG data using blind component separation, Psychophysiology 41 (2) (2004) 313 ­ 325.

117

118
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149

UNCORRECTED

PROOF

150