Experimental Studies and Computer Models of the Retina for Visual Prostheses

J. George, G.T. Kenyon, A. Yamauchi, B. Perry, X.-C. Yao, B. Barrows (P-21), B. Travis (EES-6)
Excerpted from LA-14202-PR

Introduction

Several common forms of adult-onset blindness are characterized by a massive loss of photoreceptors but a relative sparing of fibers in the optic nerve. In principle, patients suffering from such visual impairments could benefit from a prosthetic system capable of acquiring images, processing and properly encoding the information, and electrically (or magnetically) stimulating remaining retinal neurons. Preliminary studies using a crude prototype of such a device have yielded encouraging results¹, but a number of biomedical, electronic, and neuroscience issues must be addressed before the potential of this technology can be fully realized.

A consortium of DOE laboratories is working on the difficult technical problems that must be solved to achieve a satisfactory “artificial retina.” Los Alamos researchers are developing techniques for imaging patterns of activation within the retina, based on fast optical signals as well as microelectrode arrays (MEAs). These methods will be used to characterize the efficacy of electrical stimulation of the retina and evaluate alternative strategies for stimulation. In order to optimize the processing and encoding of visual information to drive a retinal prosthetic implant, we hope to discover and characterize important aspects of information processing performed by the visual system by coupling dynamic functional imaging techniques with detailed computational simulations of networks. Understanding how the retina encodes visual information is critical for achieving maximum benefit from a prosthetic device and may suggest new image-processing strategies for computer vision systems.

Retina Physiology

Electrophysiological measurements are the gold standard for characterization of neural function, allowing resolution of individual action potentials (spikes) from identified neurons. Multielectrode techniques increasingly allow studies of encoding strategies employed by entire populations of cells², but the measurements are limited by coarse sampling and crosstalk due to tissue conductivity and are disrupted (at least transiently) by electrical stimulation. We have developed functional optical imaging techniques based on measurements of intrinsic optical responses of neural tissue that are closely coupled to electrophysiological activity. However, the signals are tiny; a great deal of work has gone into enhancing size and quality of the responses.

We have undertaken studies of responses to both light and electrical stimulation in isolated amphibian retina (frog or tiger salamander). Electrophysiological recordings have been obtained with single microelectrodes and, more routinely, with MEAs fabricated on a glass substrate that serves as the bottom of the recording chamber. These MEAs consist of 60 metal electrodes in a rectangular array covering several square millimeters of area. The electrodes are connected via insulated conductors to a multichannel amplifier and data-acquisition system. In some arrays, the electrodes are flat pads. In other commercially available arrays, the electrodes are 40 mm cones rising 70 mm off the array substrate [Figure 1(a)]. We have obtained patterns of electrophysiological responses reflecting the spatial distribution of the stimulating light using planar MEAs [Figure 1(b)]. With improved experimental technique and using the three-dimensional electrode array, we have been able to regularly record the electroretinogram (the integrated electrophysiological response of the entire retina) as well as local field potential responses with embedded multiunit spike activity (Figure 2). Spikes are extracted from other signal components using digital signal-processing techniques. In the local field potentials (and in single-unit data), we observed a strong oscillatory response (about 30 Hz in frog retina) that was often apparent in single-pass data. These responses are most strongly associated with off responses, i.e., responses to a decrement in illumination, and are most prominent with wide-field stimuli. Responses appear somewhat phase locked to the stimulus, i.e., peaks in the oscillating response waveform appear at about the same point in time relative to the onset or offset of the light stimulus. Signal averaging produced small reductions in the amplitude of the oscillations but did not eliminate them. The reductions were most prominent in the first few cycles of the oscillation, suggesting that the stimulus might reset some ongoing oscillatory driver.

In experiments with a photodiode and video detectors, we have demonstrated that fast intrinsic optical responses can be measured in response to physiological activation of neural circuits including retina (Figure 3), but the signals are very small.³ Unlike other experimental techniques that employ dyes to indicate changes in cellular ion concentration or membrane potential, our methods do not require delivery of chemical to the tissue, so that in principle such measurements can be made noninvasively in the intact human eye. Unlike intrinsic signal imaging based on changes in blood flow or tissue oxygen, we detect changes in tissue light scattering or birefringence that are tightly coupled to the electrical response dynamics of the neurons. With improvements in optical configuration, we have achieved 20-fold increases in signal-to-noise ratio, and recent work has identified other possible improvements. We recently obtained the first dynamic images of the spatial patterns of physiological activation in the retina using these fast intrinsic optical responses. However, because of the small size of the scattering response, we continue to work on improving the image quality. To demonstrate the feasibility of using optical responses to monitor the efficacy of electrical stimulation, we have recently begun a series of experiments to stimulate and record from the retina using a three-dimensional MEA, while simultaneously recording the functional scattered light response elicited by stimulation. We have obtained convincing responses in several experiments, with close agreement between the duration of the plateau phase of the response in the optical signal and the corresponding electrophysiological responses recorded across a number of electrodes in the MEA. Optical measures may prove a particularly useful tool for clinical assessment of the retina. We have recently demonstrated that functional responses from retina can be recorded with optical coherence tomography⁴, a technique used routinely to assess retinal anatomy.

The availability of high-density neurophysiological data from this project provides new impetus for linking computational models of neural networks with experimental responses. If we can validate computational methods at the level of retinal neuronal networks, we increase confidence in the feasibility of modeling the dynamics of extended networks within the brain accessible by noninvasive methods.

Modeling Stimulation by a Retinal Prosthesis

An initial objective of computer modeling is to capture the biophysics of electromagnetic stimulation of neural tissue. Neural processes are activated by “gradients” in the extracellular potential and are largely insensitive to the average magnitude, or “direct current offset.” The efficacy of stimulation depends on the design of the electrode array, the properties of the tissue, and the location and orientation of neural processes within the potential field. We have developed a simplified model of the gross anatomy of the retina, and associated structures including the vitreous, retina, and peripheral tissues, as well as the MEA itself. Prosthetic stimulation is modeled by first calculating the extracellular potentials produced by the applied currents and then computing how the resulting gradients act upon dendritic and axonal processes within the retina.

Currents passed through a single isolated electrode produce a dipole-like potential field that is nearly mirror symmetric. The addition of a large insulator (corresponding to a prosthetic device affixed to the vitreous surface) forces more of the current into the retina, enhancing the potential gradient within the tissue and thus the effects of prosthetic stimulation. Our results illustrate that the size of electrodes and the overall design of the prosthetic device itself, can have a large impact on the spatial distribution of applied currents. After computing the potential distribution within the tissue, we introduce a passive cable segment to simulate the interaction between the tissue electric field and the dendritic structure of the neuron. This allows us to evaluate effects of applied currents on cellular membrane potential and thus for stimulation. Representative results are summarized in Figure 4(a).

This sort of study suggests that our ability to stimulate a particular retinal neuron will depend strongly on the details of the complicated microgeometry of the cell, as well as the spatial distribution of ionic channels within the cell. Our network-simulation software provides an adequate framework to model these functional details. We have recently added capabilities to import geometrical descriptions of real (or virtual) cells [Figure 4(b)], in order to assess the consequences of cellular geometry on the specificity of stimulation by a given configuration of stimulating electrodes. However, the sensitivity of cells to stimulation also depends on the interaction of cells within extended circuits and networks within the retina.

Computer Model of Retinal Coding and Oscillations

The retina consists of several layers of specialized neurons at the back of the eye that collectively perform the transduction and preprocessing of visual signals. The output neurons of the retina, ganglion cells whose axons make up the optic nerve, line the innermost surface of the retina (i.e., closest to the incoming light). In the absence of stimulation, most ganglion cells fire spikes in a random fashion at a background rate much lower than their maximum firing frequency. When stimulated, the firing rate increases markedly in proportion to the local contrast. This modulation of neuronal firing rate by stimulus properties and related observations are the basis for the “rate code hypothesis,” which posits that information is transmitted by the mean number of spikes per unit time irrespective of their precise timing. Other evidence, however, that the rate code hypothesis is incomplete is that as the size of a visual feature (e.g., a light or dark spot) increases, the total number of spikes is reduced, but a distinct oscillation is often observed in the firing rate. The phase of the oscillation drifts randomly over time so that the responses evoked by separate spots will rapidly become uncorrelated. Remarkably, when a single large spot stimulates two groups of neurons, their oscillations become strongly phase locked, suggesting that the relative timing of spikes in the optic nerve may convey information about the connectedness of visual features.

To study the consequences of interconnections within the retinal circuitry and to explore mechanisms of processing and encoding, we constructed a computer model (Figure 5). The model accounts for responses of certain retinal neurons to temporally modulated stimuli. The temporal modulation transfer function (tMTF) measures how strongly the output of a system is modulated as a function of the frequency of a sinusoidal input. The model retina exhibited a sharp resonance peak in its tMTFs above 60 Hz at frequencies corresponding to experimentally observed oscillatory responses (Figure 6).^5,6 The model accounts not only for the resonance frequency, but also for an associated kink in the phase response curve that plots how much the phase of the output modulation is shifted relative to the sinusoidal input. Using our computer model, we were able to show that the kink in the phase response curve obtained from retinal ganglion cells was due to entrainment. By exploiting such resonances, we may be able to selectively activate certain retinal neurons at their favored stimulation frequencies.

To assess stimulus-evoked oscillations in the retinal model, we analyzed the local field potentials directly and also considered correlations computed between spike trains from all pairs of ganglion cells activated by the stimulus. The results were combined into an averaged correlation measure. In both the cat retina and retinal model^7,8, the phases of the oscillations to small stimuli drift randomly over time so that firing activity becomes uncorrelated over sufficiently long delays. This is a fundamentally nonlinear phenomenon arising from the threshold nature of spike generation; the phase of a linear harmonic oscillator, on average, always remains fixed relative to the stimulus onset. The retinal model was able to account for the experimentally observed size dependence of retinal oscillations (Figure 7). In both experimental data and retinal model results, small stimuli evoked little or no oscillatory response, whereas large stimuli evoked very large oscillations.

Conclusion

Evolving experimental techniques allow us to characterize the mechanism of activation of retinal neurons by electrical stimulation and to explore the processing and encoding of information by retinal neural networks. By coupling these neural population measures to computational models, we can build a useful tool for engineering the neural electronic interface, exploring advanced techniques for stimulation and optimizing the electronic systems employed to encode information for processing and interpretation by the brain.

Reference

M.S. Humayun et al., “Visual perception in a blind subject with a chronic microelectronic retinal prosthesis,” Vision Research 43, 2573–2581 (2003).
B.P. Olveczky, S.A. Baccus, and M. Meister, “Segregation of object and background motion in the retina,” Nature 423, 401–408 (2003).
D.M. Rector et al., “Scattered light imaging in vivo tracks fast and slow processes of neurophysiological activation,” Neuroimage 14, 977–994 (2001).
X.-C. Yao et al., “Rapid optical coherence tomography and recording functional scattering changes from activated frog retina,” Applied Optics (In press).
G.T. Kenyon et al., “A model of high-frequency oscillatory potentials in retinal ganglion cells,” Visual Neuroscience 20, 465–480 (2003).
L.J. Frishman et al., “Spatiotemporal frequency responses of cat retinal ganglion cells,” Journal of General Physiology 89, 599–628 (1987).
S. Neuenschwander and W. Singer, “Long-range synchronization of oscillatory light responses in the cat retina and lateral geniculate nucleus,” Nature 379, 728–732 (1996).
G.T. Kenyon et al., “Stimulus-specific oscillations in a retinal model,” IEEE Transactions on Neural Networks: Special Issue on Temporal Coding for Neural Information Processing 15, 1083–1091 (2004).

Acknowledgments

This work has received previous support from LANL Laboratory-Directed Research and Development, National Institutes of Health, and is presently supported by the U.S. DOE as part of the Artificial Retina Consortium.

For further information, contact John George, 505-665-2550, jsg@lanl.gov or Garrett Kenyon, 505-667-1900, gkenyon@lanl.gov.