Mapping Subcortical Connectivity Related to Cortical ... - IEEE Xplore

0 downloads 0 Views 458KB Size Report
tellmore@ccny.cuny.edu. Kathrin Tertel, Nadeeka R. Dias, and Nitin Tandon. Department of Neurosurgery. The University of Texas Medical School at Houston.
Mapping Subcortical Connectivity Related to Cortical Gamma and Theta Oscillations A Combined ECoG-DTI Study

Timothy M. Ellmore

Kathrin Tertel, Nadeeka R. Dias, and Nitin Tandon

Department of Psychology The City College of the City University of New York New York, NY USA [email protected]

Department of Neurosurgery The University of Texas Medical School at Houston Houston, TX USA

Abstract— An open problem in signal processing and bioimaging involves the development of techniques to localize the circuits that subserve neural oscillations in the living human brain. This is an important challenge for basic and translational neuroscience since frequency-specific rhythms are hypothesized to code for sensorimotor and cognitive processes and aberrations in these rhythms develop as early markers of pathology in disorders such as epilepsy and Parkinson’s disease. Methods to reconstruct the subcortical white matter networks related to different frequency bands have been limited by the dual challenges of recording electrical activity directly from human cortex with high spatial resolution and the inability, until a decade ago, to map white matter connectivity in vivo. In the present study, a multi-modal fusion of electrocorticography (ECoG) and diffusion tensor imaging (DTI) tractography data collected from 9 neurosurgical patients performing a working memory task known to elicit robust distributed cortical gamma and theta oscillations is presented. A similarity analysis of connectivity underlying 629 intracranial electrodes revealed distinct patterns of white matter terminating near cortex that exhibited gamma, theta, and conjoint gamma and theta oscillatory power. Future applications of this method for parcellating neural circuitry in normal and pathological states are discussed. Keywords—working memory, oscillations, gamma, intracranial EEG, diffusion tensor imaging, white matter

I.

theta,

INTRODUCTION

The neuroanatomical localization of networks related to cortical oscillations in the living human brain is a challenging, but important, topic. It is motivated by basic questions about whether there are separate circuits that underlie the frequencyspecific rhythms evoked by different sensorimotor and cognitive tasks. For example, the maintenance of information in working memory often elicits gamma and theta oscillations from widespread regions of cortex. Some areas of cortex exhibit predominately one component (e.g., theta but not gamma, or vice versa) while others exhibit both (e.g., theta and gamma), leading to questions of how these different rhythms support information representation and whether the different frequency components are subserved by separate subcortical circuitry. Neuroanatomical localization is also motivated by

more direct clinical needs. For example, can the very high gamma activity recorded from cortex of epilepsy patients be used to localize a more extensive epileptic network by tracking the subcortical white matter that terminates near the cortex exhibiting the pathological rhythm? In the present study, we test the hypothesis that when oscillations occur, for example in the gamma or theta band, then the subcortical axonal connections linking the cortical nodes that exhibit the oscillations will be spatially dissimilar and have minimal overlap. In other words, a cortical node with oscillations in a specific frequency band is predicted to be structurally connected to other nodes oscillating in a similar range of frequencies. It has been methodologically difficult to test this hypothesis in humans due first to challenges inherent in recording brain electrical activity directly and then merging this information with information about structural connectivity. To test the prediction of spatially dissimilar networks underlying frequency-specific oscillation, a technique for measuring the electrical activity with both high temporal and spatial information is needed. Additionally, a method to reconstruct the axonal fiber bundles that connect these cortical areas is also required. Functional MRI (fMRI) using the endogenous blood oxygen level dependent (BOLD) contrast mechanism has adequate spatial resolution, but is an indirect hemodynamic measure of brain function that has low temporal resolution (~1 sec). Magnetoencephalography (MEG) and scalp electroencephalography (EEG) have high temporal resolution and provide more direct measures of neuronal activity than hemodynamic imaging, but spatial localization of functional brain activity with these methods is complicated by the inverse problem and attenuation of high frequency activity (which is especially of interest in cognitive processes) by the skull and scalp. In this study, we took advantage of the rare opportunity to record electrical activity directly from cortex with high spatial (~5 mm) and temporal (~1 ms) resolution using grids of epi-pial electrodes implanted below the dura and distributed across the cortical hemispheres in neurosurgical patients being evaluated for treatment of pharmaco-resistant epilepsy. We then combined information about neural oscillation data acquired with the electrocorticography (ECoG) technique with structural connectivity information from diffusion tensor

The Epilepsy Foundation and the Vivian L. Smith Foundation for Neurological Research provided funding. N.T. was supported in part by a K12 grant from the Center for Clinical and Translational Sciences, funded by grant number UL1RR024048 from the National Center for Research Resources. Partial funding for the purchase of the Philips 3T scanner was provided by NIH S10 RR19186.

imaging (DTI) acquired on the same patients. DTI is a form of diffusion-weighted MRI (DW-MRI) that allows for the principal direction and magnitude of water diffusion to be estimated in white matter, and the spatial trajectory of coherently organized and myelinated fiber bundles to be reconstructed with millimeter resolution. Higher cognitive operations, particularly memory processing, have been demonstrated to elicit large (e.g., theta, 4-8 Hz) and high frequency (gamma, 30-200 Hz) oscillations. How these two rhythms, which have garnered the most attention with respect to memory, contribute differently to information processing remains a matter of debate, as is the contribution of other frequency bands, like beta, which are not investigated in the present study. Neuroanatomical evidence, mostly from animal studies, suggests these two rhythms are generated by different brain areas, with cortico-thalamic circuitry implicated in gamma generation and a septohippocampal system for theta generation. We recently completed an ECoG-fMRI comparison that demonstrated a large number of cortical areas showing either increased low or high frequency oscillatory power during the maintenance of letter strings in working memory [1]. In this study, we replicate this finding in a larger cohort, and report a novel application relating the pattern of white matter fiber tracts with cortical locations exhibiting low and high frequency ECoG activity. Then, to test whether these complex spatial patterns of white matter are different, we adapt a measure related to the kappa statistic, apply it to the group white matter patterns, and show that there are distinct (e.g., non-overlapping) white matter networks related to theta and gamma oscillations, and a separate unique network pattern related to conjoint theta and gamma oscillations. II.

METHODS

A. Subjects Nine patients (one left-handed, 3 males, mean age 35) scheduled for implantation of electrodes for evaluation of surgical options to treat their pharmaco-resistant epilepsy provided written informed consent for enrollment in an IRBapproved study. As part of the research, each patient underwent high-resolution pre-surgical magnetic resonance imaging and task-based electrocorticographic recording after electrode implant. B. Intracranial EEG Subdural circular platinum-iridium electrodes with a top hat design (4.5 mm diameter, 3mm contact with the cortex, 10 mm inter-electrode distance, embedded in a silastic sheet (PMT Corporation, Chanhaseen, MN) were implanted as clinically indicated using standard techniques [2]. Following electrode implant, ECoG data were acquired at 1 kHz using an EEG 1100 Nihon Koden Neurofax clinical acquisition system performance of a during a modified Sternberg working memory task detailed in a previous study [3]. Synchronization of individual behavioral components of each memory trial with the data acquisition was accomplished by sending transistor-transistor logic pulses to an isolated continuously

sampled recording channel. Following recording, all data were digitally re-referenced to a common average reference taken from all channels excluding those channels with noise (i.e., 60 Hz artifacts), any electrodes with ictal activity, and any electrodes adjacent to brain tissue that was eventually resected to treat the seizures. We focused here on identifying delay activity, or sustained changes in oscillatory power during the maintenance in memory of letter strings presented during an encoding period. Delay activity is hypothesized to be the neural instantiation of short term memory, and has been described extensively in single unit and local field potential studies [3-8]. ECoG signal analysis included analytical decomposition and computation of time-frequency power spectra as described previously [9, 10] and was implemented in Matlab 7.11 (R2010b, Mathworks, Natick MA). The decomposition is analogous to a Gabor wavelet analysis in the temporal domain, but is performed in the frequency domain for speed. During this process, each signal is Fourier transformed from the time to frequency domain, multiplied by overlapping Gaussians centered at multiple frequencies of interest (50 frequencies, min. freq=1.0 Hz, max. freq=200 Hz, fractional bandwidth=0.2), and inverse Fourier transformed back to the time domain. The resulting analytic signal had an equal number of time points as the original raw signal, but contained a number of orthogonal representations equal to the 50 different frequencies of interest. Power was calculated by squaring the amplitude of analytic signal, and was then used to construct time-frequency power spectra for two different epochs of the task, the delay interval and a control baseline interval. For each of 50 total trials completed by each patient, after serial presentation of alphabetic letter strings, a 2 second delay period followed during which time all visual information disappeared and the display monitor went completely black. On each electrode channel, power spectra were computed for each of these 50 delay periods, resulting in a matrix of power spectra of size 50 (trials) x 50 (frequencies) x 2000 (ms) and another 50 x 50 x 2000 matrix of power spectra were computed for a baseline period taken between the end and beginning of each trial. At each channel, the matrices of spectra were averaged to produce a mean power spectrum for the delay period, Pd, and a mean power spectrum for the baseline period, Pb. These mean spectra were subtracted and normalized by trial-to-trial variance using the following equation:

(s

( Pd − Pb )

2 d

/ nd ) + ( s 2b / nb )

In this equation, s is the variance over n = 50 trials. So at each channel a single spectrum of size 50 (frequency) x 2000 (time) was computed with each element representing a standardized difference, equivalent to a t-statistic, between the delay and baseline periods. C. Magnetic Resonance Imaging Pre-surgical imaging was obtained at 3 telsa (Philips Intera) and included a single high-resolution T1-weighted MRI (TR/TE=8.4/3.9ms;flip angle=8 degrees; matrix

size=256x256; FOV=240 mm; slice thickness=1.0 mm thick sagittal slices) followed by acquisition of a 32-direction diffusion imaging sequence (high angular resolution, overplus on, TR/TE=8500/67 ms; flip angle=90 degrees; matrix size 128x128; FOV=224 mm; 2 mm thick axial slices, max. bvalue of 800 s/mm2). Image processing, alignment and visualization were performed with AFNI [11]. A cortical surface reconstruction that included a pial layer and white/gray matter border approximations were made from the T1 MRI using Freesurfer v4.5.0 [12, 13]. Each diffusion-weighted volume was aligned to a skull-stripped version of the patient’s T1 MRI. The b-matrix of scanner gradient orientations was rotated by the angular motion parameters [14] derived from the alignment of each diffusion-weighted volume to the T1, and a diffusion tensor was computed at each voxel. Then, for each patient, a set of whole-brain fiber pathways was computed in native scanner space using a deterministic streamline tracking algorithm [15]. D. Group DTI Analysis To characterize the relationship between oscillations recorded on subdural electrodes and nearby white matter fiber tracts we applied a recently developed processing stream that allows for single subject and group inference [16]. Briefly, electrodes locations are determined from a high-resolution computed tomography scan taken immediately after electrode implant surgery. Since the brain and implanted grids of electrodes are deformed non-linearly relative to pre-surgical anatomy, a correction is applied based on the pre- vs. postimplant geometry and implant and explant photographs in order to localize properly the electrode centers with respect to each patient’s native space pial surface approximation. The goal is to isolate the white matter fiber pathways that terminate near the brain area where an electrode of interest records the local field potential. Since white matter tracts terminate at the gray/white matter border, the representation of the electrode is dilated inward to include the thickness of cortex so that the electrode model includes a region encompassing terminations of adjacent white matter pathways. For each patient, we performed this sequence of steps on two groups of electrodes, those that included sustained elevated oscillatory power in either the gamma or theta band across the 2 second delay period. Elevated and sustained were defined operationally as a greater than 2 standard deviation change integrated across the delay interval relative to a control baseline. Electrodes showing elevated power in both the gamma and theta bands simultaneously were not considered. For each of these two groups of electrodes, sets of white matter pathways terminating near the modeled electrode region were saved in native imaging space and written to disk as binary image masks. The theta and gamma tractography masks for each patient were transformed to standard Montreal Neurological Institute space (Collins N27 brain) by applying an affine transformation

derived from a spatial normalization of the T1 weighted MRI volume. Once the two sets of masks are in a standard space, between subject comparisons can be made by summing the integer mask values at each electrode to create two volumes where voxel values indicate white matter pathways related to elevated gamma and theta respectively that are shared across the set of 9 patients. These two summary volumes were created and a thresholding step was implemented to eliminate white matter outliers, defined as voxels representing white matter originating from a single subject. E. Quantification of White Matter Network Similarity The group gamma and theta image volumes were binarized. Then, voxels related to theta oscillations were assigned a value of 1 and voxels related to gamma oscillations were assigned a value of 2. The two maps were then summed so that voxels that overlapped were assigned a value of 3. A similarity index defined as the ratio of twice the overlapping or common area (value=3) to the sum of the individual areas of theta (value=1) and gamma (value=2) voxels was then computed as: (2) In this expression, θ and γ are the sets of pixels classified as white matter voxels related to theta and gamma oscillations respectively, and n{ θ } and n{ γ } are the number of elements in sets

θ

and

γ

respectively.

This expression is derived from the kappa statistic reliability measure, and has been previously described and applied to measuring the spatial similarity between images [17, 18]. In previous studies, S is shown to be sensitive to both differences in location and size. Two equally sized regions that overlap with each other with half their area gives S=1/2, and a region completely overlapping a smaller one of half it size gives S=2/3. While the absolute value of S is difficult to interpret precisely, in general S>0.7 indicates similarity and S