Main idea - Cortech Solutions
Transcription
Main idea - Cortech Solutions
Real-time, noninvasive estimation of neuroelectric activity in brain regions of interest Mark E. Pflieger Source Signal Imaging, Inc. San Diego USA http://www.sourcesignal.com g.tec & Cortech Solutions Neural Engineering Applications Seminar Arlington, VA, March 16, 2005 Main idea A 3D spatial filter that estimates regional brain activity may be designed offline, using structural MRI, previously recorded EEG, and (optionally) functional MRI. Then it may be applied in real-time via rapid matrix multiplication. To handle multiple, simultaneous ROIs, the matrices are simply concatenated. NE Applications, 3/16/2005, SSI/mep EEG vector v1 (t ) v (t ) v (t ) = 2 # vm ( t ) ROI activity vector rk 1 (t ) r (t ) rk (t ) = k 2 # rkd (t ) Estimator matrix E11 E Ek = 21 # Ed 1 E12 E22 # Ed 2 " E1m " E2 m % # " Edm Multiple, simultaneous ROIs Single ROI rk (t ) = Ek v(t ) r1 (t ) E1 r (t ) E 2 = 2 v (t ) # # rK (t ) E K Overview Review of EEG source estimation Global versus local estimators Why real-time source estimation? Noninvasive BCI Regional activity estimation (REGAE) 3D spatial filter design using MRI Illustration (offline) fMRI-selected regions of interest (ROIs) NE Applications, 3/16/2005, SSI/mep I. Brief review of EEG source estimation NE Applications, 3/16/2005, SSI/mep Local vs. Global Source Analysis Filter the data Region by region Independent design Estimate activity Then localize Model the data Whole brain Coupled design Localize sources Then estimate Examples Examples Resolution kernel LCMV beamformer REGAE ST multiple dipoles Linear distributed Normalized MN sLORETA NE Applications, 3/16/2005, SSI/mep Local estimators are particularly suitable for real-time applications Multiple dipole localization with unaveraged data seems infeasible. Source activity estimation using prescribed dipole locations is an improvement; but the fixed source model must explain the raw data, and source activities are coupled. Linear distributed global estimators can provide ROI signals via restriction, but are not generally optimal for this purpose. Local estimators are “tuned” to specific ROIs. NE Applications, 3/16/2005, SSI/mep II. Why real-time source estimation? NE Applications, 3/16/2005, SSI/mep Motivations Adaptive control of experimental sequences Conditional stimulus presentations ROI-centric studies Epilepsy monitoring Neurofeedback BCI Recovery of lost function Augmented function NE Applications, 3/16/2005, SSI/mep Improved EEG-based BCI? Leuthardt et al. (J. Neural Eng., vol. 1, pp. 63-71, 2004) ECoG signals Brief training periods (minutes) More precise closed-loop control However: Highly invasive; high risk of infection probably limits duration Can individuals control source signals more readily than scalp signals? NE Applications, 3/16/2005, SSI/mep III. Regional activity estimation (REGAE) NE Applications, 3/16/2005, SSI/mep Source domain, dipole elements, source space Source domain Underlying volumes and/or surfaces in the brain where possible generators of extracranial measurements may reside Macroscopic ionic currents; gray matter Dipole element ECD on a relatively small scale n = ~10K elements; discretization of source domain Volume: triples per location Surface: normal orientation Source space n-dimensional vector space spanned by the dipole elements in the source domain NE Applications, 3/16/2005, SSI/mep Measurement domain, etc. Measurement domain Sensors and channels m Measurement space (a.k.a. signal space) m-dimensional vector space spanned by the measurements Derived channels Derivation matrix, D, [l-by-m] l-dimensional derived measurement space NE Applications, 3/16/2005, SSI/mep Forward solution Volume conductor model (a.k.a. head model) Physical approximations + geometry + settings of bulk parameters + solver Analytic, BEM, FEM Forward matrix, F, [m-by-n] Net gain matrix, G=DF, [l-by-n] NE Applications, 3/16/2005, SSI/mep ROI and non-ROI The ROI may taper from a central location of interest (like the passband of a bandpass filter). R, diagonal, [n-by-n]. The non-ROI is the complement of the ROI over the source domain, I-R. The ROI center is specified in advance. Size (FWHM or radius) may be fixed or automatically adjusted. ROIs may be customized. NE Applications, 3/16/2005, SSI/mep Spatial bandpass filtering Signal of interest (SOI) measurement effects generated from the region of interest (ROI) Signal of no interest (non-SOI) measurement effects generated (at the same time) from the rest of the brain Noise measurement effects not of brain origin Spatial bandpass filtering Pass the signal of interest while rejecting signal of no interest and noise Use models, but do not model the data NE Applications, 3/16/2005, SSI/mep Primary filter = L2-norm matrix Norm matrix, H, [l-by-l] We seek a norm that assigns, for each y in the derived measurement space, a magnitude h(y) = |y|H = (ytHy)1/2 that reflects the signal of interest only “Quasi-norm”: H may have singularities h(y) is a nonnegative scalar that reflects total signal of interest ~ total observed ROI activity NE Applications, 3/16/2005, SSI/mep Estimator performance is gauged via signal detection performance An estimator h may be converted to a signal detector by setting a threshold ROC analysis of signal detectability Source covariance matrix, S, [n-by-n] Maximum entropy subject to some constraints Noise covariance matrix, N, [l-by-l] Simulate: yhit = Gx + ε Simulate: yfa = G(I-R)x + ε Generate ROC curve AUROC measures performance [0.5, 1.0] NE Applications, 3/16/2005, SSI/mep Tradeoff between spatial resolution and signal detectability 1 0.9 0.8 hit rate 0.7 10 20 30 40 50 0.6 0.5 0.4 0.3 0.2 0.1 1.0 0.9 0.8 0.7 0.7 0.6 0.5 0.4 0.4 0.3 0.2 0.1 0.1 0.0 0 false alarm rate NE Applications, 3/16/2005, SSI/mep Regional variability 1 0.9 0.8 hit rate 0.7 10 20 30 40 50 0.6 0.5 0.4 0.3 0.2 0.1 false alarm rate NE Applications, 3/16/2005, SSI/mep 0.9 0.9 0.8 0.7 0.6 0.6 0.5 0.4 0.4 0.3 0.2 0.1 0.1 0.0 0 KLD (quasi-)norm matrix Kullback-Leibler divergence (KLD) A quasi-distance between two densities Measures information available for discriminating the densities Is the ½ the trace of a matrix that may be used as a REGAE norm (Gaussians) C = GSGt + N CI-R = G(I-R)S(I-R)Gt + N HR = (C-CI-R)((CI-R)-1-C-1) NE Applications, 3/16/2005, SSI/mep REGAE estimator matrix A practical approximation H = Et E E is the [d-by-l] estimator matrix d = estimator dimension Typically, d<l Vector spatial filter May vary from region to region ~Current configurations x(t) = Ey(t) NE Applications, 3/16/2005, SSI/mep Derived time series NE Applications, 3/16/2005, SSI/mep IV. Illustration NE Applications, 3/16/2005, SSI/mep Dataset: Verbal working memory EEG + sMRI + fMRI Clark CR, Moores KA et al., 2001 Task “High-load” (VT) versus “low-load” (FT) visual verbal working memory task 128-channel event-related EEG Whole-head structural MRI Functional MRI SPM VT vs. FT, thresholded NE Applications, 3/16/2005, SSI/mep Summary of Processing Steps Step Description sMR I fMRI 3D sensors 1 Parcellate brain and head tissues √ 2 Make cortical mesh (brain model) √ 3 Make head model meshes (volume conductor geometry) √ 4 Register sensors to sMRI √ √ 5 Calculate forward matrix √ √ 6 Obtain measurement covariance matrix 7 Estimate source covariance matrix (maximum entropy estimate) √ 8 Select regions of interest √ √ 9 Derive REGAE estimators √ 10 Characterize signal detection properties of REGAE estimators (ROC curves) 11 Apply estimators to real-time EEG EEG presample EEG realtime √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ Step 1: Parcellate brain & head tissues NE Applications, 3/16/2005, SSI/mep Step 2: Make cortical mesh NE Applications, 3/16/2005, SSI/mep Steps 3-5: Make head model meshes, register electrodes, & calculate forward matrix NE Applications, 3/16/2005, SSI/mep Steps 6 & 7: Measure sensor covariance matrix & estimate maximum entropy source covariance matrix NE Applications, 3/16/2005, SSI/mep Step 8: Select ROIs NE Applications, 3/16/2005, SSI/mep NE Applications, 3/16/2005, SSI/mep Steps 9 & 10: Derive REGAE estimators & characterize signal detection ROC curves for ROIs AUROC ~ 0.75 1 0.9 0.8 0.7 p(Hit) 0.6 HRLFl 0.5 SubPSl 0.4 RCalcSr AGr 0.3 MFGr-IFS 0.2 IFSl-IFG IFGrf 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 p(False Alarm) NE Applications, 3/16/2005, SSI/mep Step 11: Apply estimators to real-time EEG Yet to be done Off-line “simulation” NE Applications, 3/16/2005, SSI/mep Single trial example A (high occipital alpha) NE Applications, 3/16/2005, SSI/mep Single trial example B (“Event” near Broca’s area, HRLFl) NE Applications, 3/16/2005, SSI/mep Conclusion The whole procedure is feasible Limitations: Accuracy of volume conductor model Source statistics model highly complex Tradeoff between spatial resolution and the ability to discriminate ROI signals from remaining brain activity Real-time or quasi-real-time derived signals may help to improve BCIs NE Applications, 3/16/2005, SSI/mep SSI NE Applications, 3/16/2005, SSI/mep References Leuthardt EC, Schalk G, Wolpaw JR, Ojemann JG, Moran DW (2004): “A brain-computer interface using electrocorticographic signals in humans,” J Neural Eng 1: 63-71. Greenblatt RE, Ossadtchi A, Pflieger ME (in press): “Local linear estimators for the bioelectromagnetic inverse problem,” IEEE Trans Signal Proc. Pflieger ME, Greenblatt RE, Kirkish J (in press): “Regional resolving power of combined MEG/EEG,” Neurol Clin Neurophysiol. Pflieger ME (in press): “Maximum entropy estimation of neuroelectric source covariance statistics,” IJBEM. Clark CR, Moores KA, Lewis A, Weber DL, Fitzgibbon S, Greenblatt RE, Brown R, Taylor J (2001): “Cortical network dynamics during verbal working memory function,” Int J Psychophysiol 42: 161-76. NE Applications, 3/16/2005, SSI/mep