Research News: 2002 - 2003 - 2004

Imaging Brain Dynamics

Authors:

• Thomas Ferree, PhD
• Corby Dale, PhD, MPH
• Tracy Luks, PhD
• Morgan Hough
• Greg Simpson, PhD

The Dynamic Neuroimaging Laboratory (DNL), located on the Parnassus campus of UCSF, utilizes a wide range of imaging and computational methods, emphasizing the integration of EEG, MEG and fMRI to study spatial and temporal aspects of human brain function in health and disease.

Why Brain Dynamics ?
The human cortex is an extended system of highly interconnected neurons, which function and process information both spatially and temporally. Most conventional radiological imaging techniques, e.g., X-ray and MRI, emphasize the spatial, while EEG and MEG emphasize the temporal.

There are many viewpoints from which temporal dynamics are essential for understanding brain function:

• Populations of neurons generate net extracellular fields which act as macroscopic state variables;
• Cognitive processing may often be decomposed into a temporal sequence of computations;
• Normal brain function appears to involve the transient emergence and dissolution of large-scale cooperative networks of activity, defining a global dynamics supporting perception and cognition;
• Abnormal fluctuations reflect many pathologies, e.g., epilepsy, stroke, autism.

By imaging both spatial and temporal aspects of brain activity, we can begin to understand the functioning of distributed neural networks underlying perception and behavior. This requires not only making the proper measurements, but also developing mathematical theories of how these signals are generated by neural populations.

Multimodal Neuroimaging

Among the variety of techniques available for measuring brain activity, each has its own unique combination of spatial and temporal resolution, as shown in Figure 2. Most significantly, PET and fMRI measure hemodynamic activity, while single-unit recordings, ECoG, EEG/MEG, and neuronal current imaging (NCI) measure electrophysiological activity. Among the electrical measures, only EEG/MEG and NCI are noninvasive.

The relationship between EEG and MEG deserves some clarification. EEG has been widely used since the 1930's. When MEG was introduced in the 1980's, it was first argued that it should have higher spatial resolution, because EEG is more distorted by the poorly conducting skull. With the advent of realistic head models that account for the skull and other head tissues, it has since become appreciated that the spatial resolutions of EEG and MEG are similar.
EEG/MEG results from the transmembrane currents of neocortical pyramidal neurons, whose apical dendrites lie along the local normal to the cortical surface. Synchronous activity of relatively few nearby neurons generates the familiar current dipoles. The volumes sampled by EEG and MEG are also similar, but the geometries of the electric and magnetic fields are very different. MEG is nearly blind to radial dipoles, for example, while EEG sees both radial and tangential. At the level of the detectors, therefore, EEG and MEG are complementary. Simultaneous EEG/MEG recording and analysis are now considered state-of-the-art, and provide higher spatial resolution and accuracy than either method independently.
Because EEG/MEG have lower spatial resolution than fMRI, intuition suggests that a combined approach should be advantageous. But because EEG/MEG and fMRI measure different physiology, and their exact relationship is still poorly understood, this is not a trivial undertaking. We could argue that a first priority should be to understand better this relationship, e.g., through detailed electrophysiological and vascular models, as this would support their combined use in researching more complicated neuroscience questions.

Integrative Neuroscience

Rapid technological advances, exponential growth in scientific literature, and increased specialization have led to a data overload in biomedical sciences. Now more than ever, real progress in neuroscience depends upon a close relationship between theory and experiment. Figure 3 depicts DNL's current approach to this multifaceted problem. Integrated measurement refers to simultaneous acquisition whenever possible, since biological systems are highly variable and sequential experiments never have identical control conditions. Integrated analysis refers to the use of different data sets, collected simultaneously or sequentially, to find globally optimal solutions that appropriately weight all the information. Neuronal current imaging (NCI) and diffusion weighted imaging (DWI) are at the development stage, and preliminary results suggest great promise as additional tools for integrated research.
Experimental programs are best motivated from sound theoretical frameworks, continually refined and streamlined to better support interpretability of data. These theories may arise from various viewpoints, corresponding roughly to scientific disciplines. In neuroscience, we seek to understand brain function in terms of underlying neuronal constituents and their synaptic interactions, but realize that a purely reductionist approach might not be the most profitable. Models and analyses are all computationally intensive, and computation itself is a metaphor for brain information processing.
At the DNL, we study brain function in both health and disease. Our emphasis in cognitive neuroscience is on attention and working memory. A better understanding of these processes in healthy subjects can guide clinical studies in aging, dementia, schizophrenia, and autism.

Pathophysiology of Acute Stroke
As an example of how brain dynamics can reflect pathophysiology, we describe a recent study aimed at detecting acute cerebral ischemia. Dense-array (128-channel) EEG data were collected from 10 subjects, presenting in the emergency room with signs of recent stroke.

Historically, analysis of EEG on paper charts looked for abnormal patterns, such as oscillations of a particular frequency or epileptic spikes. Later the estimation of frequency content was automated with Fourier transform and wavelets. But the question remains: How best to characterize EEG time series? This question is of theoretical interest, because it queries the nature of the EEG signal, and therefore the dynamics of large neural populations. It is also of practical interest, because the huge data sets produced by dense sensor arrays are too labor-intensive to analyze by eye. Furthermore, the eye may miss patterns in the data, which suitable algorithms might readily detect.
The brain is obviously a complex system and, like most complex systems, it exhibits a very broadbanded power spectrum, i.e., Fourier analysis reveals all frequencies contributing significantly. While it is tempting to look for dominant peaks and interpret them as resonances of an approximately linear system, such an approach is not necessarily suitable for complex nonlinear data. An alternative approach is to quantify fluctuations across a range of time scales, and look for patterns in the data that are independent of scale. This is akin to the familiar fractal analysis, but applied to a time series rather than a static geometric object.

In a study of 10 stroke patients and 18 normal control subjects, we found that with few exceptions 10- second segments of resting EEG may reasonably be described with just two dimensionless parameters, called scaling exponents [Hwa and Ferree (2002), Physical Review E 66: 021901]. In addition to theoretical underpinnings suggesting that scaling exponents might be natural for describing the EEG, this parameterization has practical advantages for comparing across data channels and across subject groups.

Using this method, the resting EEG for each of 28 subjects was reduced to 128 pairs of scaling exponents. It seems intuitive that a brain lesion like a stroke might change the distributions of the scaling exponents over the scalp. By considering the mean, variance, and higher moments for each subject, we were able to derive a direct physiological measure of brain function, called S, which linearly separates the two subject groups (Figure 5). This is a novel and remarkable step in data reduction, anticipated only by the general observation that many complex systems exhibit scale-independent behavior of some sort.

We have half-jokingly likened our stroke measure S to the body temperature T, because it serves a similar clinical purpose. In both cases, many poorly understood physiological processes contribute to the measure, yet knowing just this one number can inform a physician about the likelihood of pathology.