Interactive segmentation of EEG synchrony data in time-frequency space by means of region-growing and Bayesian regularization. Alfonso Alba and Edgar Arce Departamento de Electr ´ onica, Facultad de Ciencias, Universidad Aut´ onoma de San Luis Potos´ ı, Diagonal Sur S/N, Zona Universitaria, C.P. 78240, San Luis Potos´ ı, S. L. P., M´ exico fac@galia.fc.uaslp.mx, arce@fciencias.uaslp.mx Abstract In this paper we present a new methodology for the inter- active visualization and segmentation of electroencephalo- graphic (EEG) scalp synchrony data. Synchrony mea- surements are estimated for all electrode pairs and clas- sified as higher, lower, or equal than the baseline aver- age. The classified values are then displayed in the form of Time-Frequency-Topography (TFT) maps, which can be segmented using a seeded region growing algorithm and a Bayesian regularization technique. Finally, we present the synchronization maps that result from the analysis of real EEG data from a figure categorization experiment. 1. Introduction It is thought that during the execution of a relatively com- plex task, specialized (and possibly distant) areas of the brain interact together by means of reciprocal connections, forming what is called a neural assembly [2]. One of the most plausible mechanisms for this integration is the forma- tion of dynamical links which are reflected as some form of synchronization of the EEG signals over different frequency bands [13]. Coupling between neural assemblies may occur at multi- ple spatial and temporal scales [13]. The finest scale would be at single neuron level, where, in the timescale of cogni- tive events (hundreds of milliseconds), a single neuron may fire only a few spikes, which may not be enough to trigger a target neuron. However, if these spikes coincide with those fired by other neurons, the target neuron may be activated. On a coarser level, a scalp electrode records the average synaptic action of a large number of neurons (between 10 7 and 10 9 ) [9]. If a significant number of these neurons be- come synchronized, one may observe a power increase in the EEG signal at some frequency band with respect to the baseline (i.e., the power measured during a neutral state). Similarly, one may associate a desynchronization episode with a power decrease in the corresponding EEG signal; these episodes are known in the literature as event-related synchronization (ERS) and event-related desynchronization (ERD) [11]. Finally, one may also observe couplings be- tween two relatively distant areas (e.g., between two elec- trode signals), often called long-range synchronization. Ac- cording to David et al., brain areas are massively and re- ciprocally connected, and such connectivity induces corre- lations in the brain activity of those areas [2]. It is worth noting that long-range synchrony data is usually of higher dimensionality than ERS/ERD data, since it involves inter- actions between two or more sites. This usually implies a visualization problem, which most works on the field avoid by averaging across a large time window (e.g., [8]), and/or by limiting the analysis to only a few frequency bands or electrode pairs (e.g., [5]). These solutions are far from ideal, since they do not give an overall picture of the synchroniza- tion dynamics corresponding to a given experiment. In 2004, Marroquin et al. [6] proposed a methodology to detect and display ERS and ERD episodes in cognitive ex- periments, using a novel time-frequency-topography (TFT) visualization technique which provides detailed maps of the EEG dynamics with high spatial, temporal, and frequency resolutions. The methodology also allows for a segmenta- tion of the time-frequency (TF) plane with respect to spatial activation patterns, which may be related to specific cogni- tive tasks. Here we describe a similar visualization system designed for the interactive exploration of long-range syn- chronization patterns in TFT space, and an interactive pro- cedure for the segmentation of the TF plane with respect to specific synchronization patterns.