Journal of Neurophysiology

Electrical Stimulation of Mammalian Retinal Ganglion Cells With Multielectrode Arrays

Chris Sekirnjak, Pawel Hottowy, Alexander Sher, Wladyslaw Dabrowski, A. M. Litke, E. J. Chichilnisky


Existing epiretinal implants for the blind are designed to electrically stimulate large groups of surviving retinal neurons using a small number of electrodes with diameters of several hundred micrometers. To increase the spatial resolution of artificial sight, electrodes much smaller than those currently in use are desirable. In this study, we stimulated and recorded ganglion cells in isolated pieces of rat, guinea pig, and monkey retina. We used microfabricated hexagonal arrays of 61 platinum disk electrodes with diameters between 6 and 25 μm, spaced 60 μm apart. Charge-balanced current pulses evoked one or two spikes at latencies as short as 0.2 ms, and typically only one or a few recorded ganglion cells were stimulated. Application of several synaptic blockers did not abolish the evoked responses, implying direct activation of ganglion cells. Threshold charge densities were typically <0.1 mC/cm2 for a pulse duration of 100 μs, corresponding to charge thresholds of <100 pC. Stimulation remained effective after several hours and at high frequencies. To show that closely spaced electrodes can elicit independent ganglion cell responses, we used the multielectrode array to stimulate several nearby ganglion cells simultaneously. From these data, we conclude that electrical stimulation of mammalian retina with small-diameter electrode arrays is achievable and can provide high temporal and spatial precision at low charge densities. We review previous epiretinal stimulation studies and discuss our results in the context of 32 other publications, comparing threshold parameters and safety limits.


Recent attempts to restore vision in the blind have met with extraordinary success. Electrical stimulation of retinas in people with neurodegenerative diseases has shown the potential for direct excitation of neurons as a means of re-establishing sight. Long-term retinal implants in several profoundly blind people were shown to produce perceptions of light and allowed for the detection of motion and discrimination of very simple shapes (Humayun 2003; Humayun et al. 2003).

Such achievement brings hope to the millions of people worldwide who suffer from photoreceptor loss because of advanced retinitis pigmentosa or age-related macular degeneration (Heckenlively et al. 1988; Klein et al. 1997). It is expected that 10 years from now, macular degeneration will become the single leading cause of legal blindness, with an incidence as high as 5.5% in people over 65 (Klein et al. 1997). While degenerative diseases result in severe damage to photoreceptors, inner retinal neurons survive at fairly high rates (Kim et al. 2002; Santos et al. 1997; Stone et al. 1992) and may be electrically excitable. The fundamental concept underlying retinal neuroprosthetic devices is to electrically activate those residual neurons by bypassing the damaged photoreceptors, thus achieving artificial vision in otherwise blind patients. Of several prosthetics designs, epiretinal implants specifically target ganglion cells by positioning electrodes in close proximity to the inner surface of the retina.

Despite recent successes, the current implants are but a first step toward restoring sight. To create useful vision, stimulating electrodes must be arranged in two-dimensional arrays that generate a visual image made up of a matrix of discrete perceptions of light. Psychophysical studies suggest that foveal implants may provide the user with an acceptable level of mobility if they contain a minimum of ∼600 electrodes (Cha et al. 1992a,b). To achieve this number or greater, electrodes must be tightly packed, necessitating small stimulation sites. At present, a typical epiretinal implant contains tens of electrodes with diameters of a few hundred micrometers, spaced several hundred micrometers apart (Humayun 2003). Considering that such electrodes are much larger than the cells they stimulate, the need for implants with hundreds or thousands of much smaller electrodes is apparent. To match the intrinsic resolution of the visual system, an advanced implant would devote one electrode to every ganglion cell. This requires that each electrode be similar in size to a ganglion cell (∼5–20 μm). Instead of affecting hundreds or thousands of cells, each electrode would evoke a few spikes in a few retinal ganglion cells.

The success of the next generation of implantable devices will be tied to our understanding of how to activate neurons with extracellular electric stimuli applied to the retinal surface through electrodes that approach cellular dimensions. Little is known about the parameters that would permit reliable retinal stimulation with small electrodes. When the electrode surface area is reduced, current density and charge density increase rapidly, and high charge densities are known to cause tissue damage by electrochemical reactions (Brummer et al. 1983; Pollen 1977; Tehovnik 1996). A detailed in vitro analysis of small electrode stimulation is thus a prerequisite for developing such implants for use in human patients.

A comprehensive literature review reveals that the feasibility of stimulation with arrays of small electrodes in mammalian tissue has not been adequately tested. The majority of studies involving retinal stimulation have used needle-shaped probes with one or two conductors at the end of an insulated rod, such as platinum wires or concentric microelectrodes. In its simplest form, such stimulating probes are made of metal wires several hundred micrometers in diameter, exposed at the tip and insulated elsewhere (Doty and Grimm 1962; Humayun et al. 1994; Nadig 1999; Suzuki et al. 2004; Weiland et al. 1999).

Others have attempted to use stimulating microprobes with tip diameters of 25 μm or smaller (Dawson and Radtke 1977; Jensen et al. 2003; Rizzo et al. 1997; Wyatt et al. 1994). However, the geometry of such probes differs greatly from the planar disk electrode design developed for current epiretinal implants. Stimulation, furthermore, is always limited to a single stimulation site, precluding the study of stimulation using multiple electrodes and their interaction effects.

The use of multielectrode arrays for retinal stimulation has been mainly limited to large electrodes with diameters between 100 and 1,500 μm (Greenberg 1998; Hesse et al. 2000; Humayun et al. 1999, 2003; Rizzo et al. 2003a; Walter and Heimann 2000). Multielectrode arrays with smaller electrodes (∼10 μm diam) have been used to stimulate the retina in the subretinal space (Stett et al. 2000; Zrenner et al. 1999). Grumet has used an array to selectively stimulate the axons of retinal ganglion cells, using a separate distant array to record somatic spikes (Grumet 1999; Grumet et al. 2000). No study has targeted mammalian ganglion cell bodies for direct epiretinal stimulation using planar electrodes with surface areas <200 μm2.

In this study, we established thresholds for stimulation of ganglion cells in rat, guinea pig, and primate retina using electrodes with surface areas of 30–500 μm2 (diameters of 6–25 μm). We used these parameters to further study frequency dependence, pharmacology, and spatial interaction effects of stimulation. Our arrays use planar disk microelectrodes very similar to those used in present epiretinal prosthetics, but smaller by an order or two of magnitude. We conclude our analysis by discussing the results in the context of the pertinent literature.

Early and preliminary portions of this work have been presented elsewhere (Sekirnjak et al. 2006).


Retinal preparation

This study used retinal tissue from 28 adult rats, 4 guinea pigs, and 1 macaque monkey. The average body weight was 289 ± 8 g for rats (Long-Evans), 386 ± 52 g for guinea pigs, and 4 kg for the macaque monkey (Macaca radiata).

Rodent eyes were enucleated after decapitation of animals deeply anesthetized with 10 mg/kg xylazine and 50 mg/kg ketamine HCl. Primate eyes were obtained from terminally anesthetized macaque monkeys used by other experimenters, in accordance with institutional guidelines for the care and use of animals. Immediately after enucleation, the anterior portion of the eye and vitreous were removed in room light and the eye cup placed in bicarbonate-buffered Ames' solution. Vitreous removal in rats was aided by a homemade extractor that allowed for rapid but gentle separation of retina and vitreous gel. The success rate for vitrectomies performed in this manner was 92%.

Pieces of retina 1–2 mm in diameter (Fig. 1A) were separated from the retinal pigment epithelium and placed flat on the electrode array, with the ganglion cell layer facing the array (Fig. 1B). The tissue was held in place by weighted nylon netting positioned over the array. The preparation was mounted on a circuit board attached to an inverted microscope and continuously superfused at room temperature with Ames' solution bubbled with 95% oxygen and 5% carbon dioxide at a flow rate of 2–4 ml/min. Pharmacological agents (TTX, kynurenic acid, CNQX, AP-5, cadmium chloride) were added directly to the perfusion solution.

FIG. 1.

Experimental design. A: a piece of retina positioned over the hexagonal multielectrode array (black dots) and held in place by nylon netting, photographed during a stimulation experiment. Electrode diameter = 14 μm. B: schematic cross-section of the retina–array assembly. Preparation is submerged in solution and grounded by a distant platinum wire. The 61 platinum electrodes are used both to record spike activity and to send current pulses. C: standard stimulus pulse used to stimulate ganglion cells.

Multielectrode array

The array consisted of a planar hexagonal arrangement of 61 extracellular electrodes, ∼0.5 × 0.5 mm2 in total size (Fig. 1A). These electrodes were used both to record action potentials extracellularly from ganglion cells (Chichilnisky and Baylor 1999; Meister et al. 1994) and to apply current to the tissue for stimulation. In some experiments, different neighboring electrodes were used for stimulating and recording.

The array was microfabricated on a glass substrate, with indium tin oxide leads and silicon nitride insulation (Litke 1998; Litke et al. 2003). Each electrode was formed by microwells (holes in the silicon nitride layer), which were electroplated with platinum before an experiment (Fig. 1, A and B). This was accomplished by submersing the array in a 0.0025N HCl solution containing 1% chloroplatinic acid and 0.01% lead acetate and applying voltages of 1–5 V through 10-MΩ resistors for 10–120 s. Electrode size was determined by well diameter (5, 6, 8, 10, 12, or 14 μm) as well as the amount of platinum deposited in each well. Final electrode diameter varied between ∼6 and 25 μm, with a fixed interelectrode spacing of 60 μm. The geometric electrode area (πr2) was used to calculate current and charge densities; however, platinum tends to deposit in a granular fashion, rendering the effective electrode area significantly larger (Mathieson et al. 2004). A circular chamber glued on the glass plate allowed for perfusion of saline solution. A 4-cm-long platinum wire loop integrated into the chamber served as distant ground. All stimulations were performed using a monopolar configuration (electrode to distant ground).

Electrical stimulation and recording

Unless otherwise noted, experiments were performed on a setup allowing for simultaneous recording of all 61 electrodes and stimulation on multiple electrodes. The array was connected to a circuit board containing two custom-made readout application-specific integrated circuits (ASICs) that amplified, filtered, and multiplexed signals from the 61 electrodes and sent them to ADC cards installed in a PC. The board also contained two computer-controlled ASICs capable of sending current pulses to any configuration of electrodes (Dabrowski et al. 2005). A dim level of illumination was maintained during the entire experiment (room lights or microscope illuminator).

Recording and stimulation were controlled by interface software (Labview). Extracellular potentials were recorded from all 61 electrodes, digitized at 20 kHz (Litke 1999), and stored for off-line analysis. The available discrete stimulation pulse current amplitudes were 0.6, 0.8, 1.0, 1.2, 1.5, 1.7, 2.0, 2.3, 2.7,3.0, 3.3, 4.0, 4.7, 5.3, 6.0, 6.7, 8.0, 9.3, 10.7, 12.0, 13.3, and 16.7 μA (several threshold curves reported in Fig. 6A were determined using an earlier stimulus generator capable of delivering current amplitudes as low as 0.1 μA. This device was not used in subsequent experiments). The pulse consisted of a cathodic (negative) current pulse of amplitude A and duration d, followed immediately by an anodic (positive) pulse of amplitude A/2 and duration 2d (Fig. 1C). All pulses were individually calibrated to produce stimuli with zero net charge. Charge-balanced biphasic waveforms such as these minimize electrode corrosion by preventing charge accumulation and the associated irreversible Faradaic reactions (Merrill et al. 2005). Reported current values always refer to the negative phase amplitude A. Pulse duration was 0.05, 0.1, 0.2, 0.5, or 1 ms and always refers to the duration d of the cathodic phase. The pulse shape could be inverted in time to yield an anodic-first stimulus. Stimulation frequency was varied between 0.25 and 300 Hz.

Experimental protocol

Many ganglion cells show maintained activity under diffuse retinal illumination (Troy and Robson 1992) and fire spontaneous spikes in isolated pieces of retina. Stimulation on a particular electrode was typically attempted if spontaneous extracellular spikes could be recorded from that electrode. This approach guaranteed that the electrode was properly platinized and confirmed that ganglion cells in the vicinity of the electrode were alive. Typically, at least half of the platinized electrodes on an array showed spontaneous activity from at least one cell. Spontaneous spikes were readily distinguished from evoked spikes because they bore no temporal relationship to the stimulus pulse, whereas evoked spikes were locked to the stimulus onset.

Stimulation was typically attempted by using the lowest current settings and was then increased systematically if no response was seen. Threshold was defined as the current setting which produced a spike with nearly every stimulus pulse (≥90% of trials) while stimulating at 1–2 Hz. Latency was defined as the time between stimulus pulse onset and the first deflection of the evoked spike. Unless otherwise stated, threshold current, threshold charge, and threshold charge density always refer to the negative phase of the biphasic, charge-balanced stimulus pulse.

For pharmacological manipulations, a minimum drug perfusion time of 5–10 min was allowed before responses were recorded.

Data analysis

Multielectrode data were analyzed off-line using Labview, Matlab, and Igor Pro. Means, SE, and P values were calculated in Microsoft Excel. Images were processed in Adobe Photoshop.

Chronaxies were calculated by fitting power functions y = a/x + b and y = a/xp + b (Holsheimer et al. 2000; Lapicque 1907; Ranck 1975) or exponentials y = b/(1 − ex/a) (Greenberg 1998; Lapicque 1907; Plonsey and Barr 1988) to the strength-duration data. The asymptote (coefficient b) was defined as the rheobase; chronaxie was calculated as (a/b)1/p, a/b, or aln2 for power and exponential fit functions, respectively. Fit quality was assessed by calculation of χ2 and visual inspection of the fit curve, and noticeably poor fits were not included in the analysis. Given the small number of data points available for some cells, the quality of the fit and resulting parameters differed for the individual functions, and thus values from all three are reported.

Autocorrelations of evoked and spontaneous spiking were obtained by generating histograms of spike times and interspike intervals, respectively. On average, ∼37 spikes were used per histogram. Spontaneous histograms were aligned so that time = 0 coincided with the occurrence of the peak of the first evoked spike.

Power function fit lines to literature data and R2 values were calculated in Igor Pro by fitting linear functions to the logarithmic plots of threshold parameters.

Statistical comparisons were done by performing a Student's t-test (2-tailed, equal variance) with a significance limit of P < 0.05.

Errors and error bars reported in this study are SE, unless otherwise stated.

Threshold artifact subtraction

To reveal spikes with latencies of <2 ms, a novel digital subtraction technique was used. Spikes obscured by the stimulation artifact (which typically lasted for several milliseconds) were made visible by increasing the stimulation current until a possible spike threshold was reached. Just below threshold, the recorded traces changed shape noticeably on about half of the stimulus trials (e.g., a change in curvature or peak height), indicating that a possible spike hidden inside the artifact was elicited on some trials (Fig. 2B). Subsequently, the digital difference between two such traces was calculated. Because the artifact itself was identical in both traces, the subtraction cleanly revealed the spike inside the stimulus artifact. Typically, several traces with and without a suspected spike were averaged before subtraction to increase the signal over the noise. This method necessitated recording the spikes on an electrode immediately adjacent to the stimulation electrode, because the signal recorded at the stimulation site usually saturated the amplifier and was not suitable for subtraction. The results obtained were comparable to or better than those reported for artifact suppression by local curve fitting (Wagenaar and Potter 2002).

FIG. 2.

Ganglion cell responses evoked by electrical stimulation. A: spontaneous spikes (arrowheads) and evoked spikes (arrows) in monkey and guinea pig retina. Stimulus pulse is shown above traces. Latency in the monkey cell was 5.6 ms, whereas the evoked spike in the guinea pig trace was obscured by the stimulus artifact and required digital subtraction to reveal. B: same cell as bottom of A at an expanded scale. Superposition of several stimulation trials is shown, some of which evoked a spike (arrow). Bottom trace is the digital subtraction of averaged traces from trials with and without an evoked spike. Dashed line indicates stimulation onset. Latency was 0.25 ms. Inset (top right): spontaneous spike from this cell at the same scale. C: spike latencies measured for 48 short-latency spikes (left) and 74 long-latency spikes (right) using 0.1-ms pulses.

We further verified this technique in six cells by applying TTX (1 μM) to the bath solution. The stimulus artifact recorded in TTX was subtracted from the traces containing obscured spikes. In these cells, the resulting subtracted spikes were identical to the spikes obtained using the above method. An example of this is shown in Fig. 4A.

Literature analysis

Data from epiretinal stimulation studies were compiled as follows. Threshold current, current density, charge, and charge density necessary to elicit a ganglion cell response were median or mean values as reported in each study. When a list of thresholds was reported, an average value was calculated; when a range of thresholds was reported, the minimum or the median value was used; in some cases, a “typical” value was taken from a representative example or figure. Whenever possible, a missing parameter was calculated from reported parameters, for instance, threshold current from reported threshold charge (Humayun et al. 1999; Suzuki et al. 2004), surface area from charge density (Dawson and Radtke 1977), or current from reported charge density (Nadig 1999). In a few cases it was not possible to calculate a parameter, and it was estimated from other publications by the same author or the same group (asterisks in Table 1). When such substitution was not possible, the study was not included (Benjamin et al. 1994; Crapper and Noell 1963; Kuras and Gutmaniene 1997; Narayanan et al. 1994). Several studies were represented by multiple entries when different values of parameters were reported (electrode size, duration, pulse shape) or when several drastically different results were reported for the same parameters (such as for 2 or more human subjects). The geometric surface area was calculated from the reported electrode geometry [circular or rectangular for planar electrodes: πr2 or l2, cylindrical for exposed wires: πr2h, conical for cone tips: πr(r2 + h2)1/2, spherical for ball electrodes: 4πr2]. When two or more electrodes were reported to be coupled electrically and used simultaneously, the surface area was multiplied accordingly. When a stimulus consisted of high-frequency pulse trains, the effective pulse duration was taken as the number of pulses per train times the single-pulse duration (Laube et al. 2003; Walter and Heimann 2000). Whenever possible, data from normal animals, not those with degenerated retinas, were used.

View this table:

Literature analysis

For plotting the neural injury limit, cat cortical tissue data from McCreery et al. (1990) was fit to the equation log(Q/A) = k − log(Q), where Q is the charge in nC and Q/A is the charge density in mC/cm2 (Merrill et al. 2005; Shannon 1992). The data can be fit with a coefficient k varying between 1.7 and 2.0; both values were used for the injury limit plots in Fig. 12.


We electrically stimulated pieces of isolated mammalian retina while simultaneously recording spiking activity in ganglion cells. The properties of evoked spikes are presented first, followed by strength-duration relationships, temporal properties, and the results from multielectrode stimulation.

Stimulation at individual array electrodes resulted in all-or-none spikes recorded at latencies between a few hundred microseconds and tens of milliseconds. Of the 184 successfully stimulated ganglion cells, 165 were from rats, 11 were from guinea pigs, and 8 were from the monkey. On average, each retina yielded 6 ± 1 stimulated cells. Most responses consisted of one or two spikes, although in some cells later spikes were recorded.

Response latencies

We classified spikes with latency ≥2 ms as long-latency spikes and earlier responses as short-latency. Latencies >10 ms were infrequently observed, and virtually no spikes occurred >20 ms after stimulation onset. Typically, only long-latency spikes could be readily discerned because the stimulus artifact obscured the first few milliseconds of the recording.

Figure 2A shows two spontaneously firing ganglion cells and their responses to single stimulus pulses. While the primate cell (top) responded with a distinct spike at latency 5.6 ms, the guinea pig response (bottom) was obscured by the stimulus artifact. To isolate the evoked short-latency spike, a threshold artifact subtraction method was used (see methods). Briefly, the artifact was selectively eliminated by recording several traces near threshold and subtracting those traces that did not contain evoked spikes (Fig. 2B). This method was typically used when a neighboring electrode was used for stimulation in lieu of the recording electrode, because this configuration reduced the artifact below amplifier saturation levels and allowed the artifact to be subtracted. The result for the guinea pig cell is shown at the bottom of Fig. 2B: a spike was revealed at 0.25 ms latency. For 86 spikes in rats, visible without artifact subtraction, the average latency was 7.6 ± 0.3 ms, whereas 48 artifact-subtracted spikes had a latency of 0.73 ± 0.05 ms. Nearly all short-latency spikes occurred at <1 ms; the shortest latencies recorded in this study were ∼0.2 ms. Latency histograms for both short- and long latency spikes are shown in Fig. 2C.

Evoked spikes usually resembled the recorded spontaneous spikes, but occasionally spikes from a different cell were elicited. Short-latency spikes in particular tended to be of identical shape as the spontaneous spikes. This is shown in the inset of Fig. 2B: the evoked spike resembled the spontaneous spike. Two further examples are shown in Fig. 4A and in the inset to Fig. 7B. Of 48 subtracted short-latency spikes, 42 unambiguously matched the spontaneous spike.

To elucidate the origin of long-latency spikes, the method of digital artifact subtraction was applied to recordings which contained both short- and long-latency spikes. It seemed possible that each long-latency spike was in fact the second spike of a pair response and not a solitary spike. Indeed, analysis of 20 cells revealed that the occurrence of long-latency spikes (6.4 ± 0.3 ms) was always associated with short-latency responses (0.7 ± 0.1 ms). An example from guinea pig retina is shown in Fig. 3 : whereas the raw data traces (Fig. 3A) showed only three long-latency spikes (asterisks), the artifact-subtracted traces (Fig. 3B) revealed that every long-latency spike was preceded by a short-latency spike at 0.35 ms. Furthermore, an analysis of spontaneous spiking activity showed that spike doublets spontaneously occurred in this cell. This is shown at the bottom of Fig. 3B: the autocorrelation histogram of spontaneous spikes showed a peak at a latency similar to that of the evoked long-latency spikes. Thus the evoked spikes occurred with timing expected from the spontaneous activity. A second example from rat retina is shown in Fig. 3C for a cell with long-latency spikes at 7 ms. Spike timing analyses were performed in a total of eight cells, with similar results: the spontaneous interspike intervals matched the typical intervals between short- and long-latency spikes. These results indicate that some cells responded to a single stimulus pulse with a spike pair, with the first spike obscured by the artifact, and that this tendency toward paired spiking was evident in the spontaneous activity of the cell.

FIG. 3.

Relationship between short- and long-latency spikes. A: raw traces of 9 stimulation trials in guinea pig retina near threshold. Stimulus pulse is shown above traces. Long-latency spikes (*) were evoked on several trials (latency 2–3 ms). B: top: artifact-subtracted traces revealed short-latency spikes (latency 0.35 ms), some of which were followed by long-latency responses. Dashed vertical line marks onset of stimulation. Bottom: latency histograms of spontaneous and evoked spikes. Spontaneous histogram plots interspike intervals and was offset in time by latency of early spikes. Bin width: 0.3 ms. C: latency histograms in another cell with long-latency responses at ∼7 ms (rat). Bin width: 1.2 ms. D: latency data from a cell for which artifact subtraction was not possible and thus only the long-latency response (at 5 ms) was available (rat). Inferred short-latency spikes occurred at times indicated by dashed box. Bin width: 1.2 ms.

The method of analyzing spike timing was further used to calculate the approximate latency of obscured short-latency spikes when only long-latency spikes were available. Figure 3D shows an example of a cell in which a large stimulus artifact precluded the use of the artifact subtraction method; only long-latency spikes were discernible. By aligning the peaks of the two histograms, a short-latency spike (dashed box) was inferred at times ≤1 ms. Similar results were found in three cells and suggest that short-latency responses can be deduced from the observance of long-latency spikes.

Last, we compared the spike latencies of long-latency responses evoked with stimulation electrodes of different diameters, which ranged from 6 to 25 μm in this study. No systematic difference was observed when large rather than small electrodes were used and average latencies for the smallest electrodes (6–9 μm) were similar to the largest (20–25 μm): 8.2 ± 0.7 and 7.1 ± 0.4 ms, respectively (P > 0.2; n = 32 cells).

Pharmacological manipulations

Several ion channel antagonists were applied to the perfusion solution to further study the evoked responses. To ascertain that the observed spikes were of neuronal origin, the sodium channel blocker TTX (1 μM) was added to the perfusion solution. In 3 guinea pig and 17 rat cells, all spikes (both spontaneous and evoked) disappeared within seconds of drug application, confirming their identity as neuronal action potentials. An example is shown in Fig. 4A : application of TTX eliminated evoked short-latency spikes, leaving only the pulse artifact. When this artifact was subtracted from the control responses, an evoked spike was revealed (Fig. 4A). The waveform of this spike did not differ from that derived by threshold artifact subtraction or the spontaneous spike recorded at this electrode (Fig. 4A, bottom). Similar results were found in five cells.

FIG. 4.

Pharmacological manipulations. A: top: responses evoked by a 1-μA pulse (solid traces) were eliminated by application of TTX (dotted traces). A superposition of 3 trials is shown for each condition. Bottom: spikes obtained by subtracting the TTX response from control (1st trace), by using threshold artifact subtraction (2nd trace) and a spontaneous spike (3rd trace). Arrowhead marks onset of stimulation; latency = 0.8 ms. B and C: top: responses to 10 stimulus pulses in a cell with spikes at latency 5.5 ms (B) and in a cell with spikes at 15 ms (C) in rat retina. A combination of glutamate receptor antagonists (CNQX, APV, kynurenic acid) added to the perfusion solution (bottom traces) failed to eliminate responses.

To study whether the applied current pulses acted directly on ganglion cells or involved more distant cells with synaptic connections to the recorded cell, blockers of synaptic transmission were added to the perfusion solution. A combination of the following agents was used: the broad spectrum glutamate antagonist kynurenic acid (1 mM), the N-methyl-d-asparate (NMDA) receptor blocker APV (400 μM), and the AMPA-receptor blocker CNQX (75 μM). These blockers have been used in retinal preparations by other groups at equal or lower concentrations (Jensen et al. 2005a; Stett et al. 2000; Ziv et al. 2002). Figure 4, B and C, shows examples of responses from two cells, recorded before and after addition of the blockers. Spike shapes, latencies, and response rates were unchanged, even in the cell with spikes at latency 15 ms (Fig. 4C). No systematic differences between spikes elicited in control and drug conditions were observed in any of nine cells. These findings suggest that ganglion cells were activated directly, not trans-synaptically, and further corroborate the notion that apparent long-latency spikes (such as in Fig. 4, B and C) are not solitary spikes, but part of a two-spike response.

In separate experiments, the calcium channel blocker cadmium chloride (100–250 μM) was applied to the perfusion solution to abolish synaptic transmission (data not shown). In 10 cells, evoked spikes were still observed after drug application, indicating that the observed spikes were not produced by mechanisms involving calcium-dependent synaptic transmission.

Minimal thresholds and spatial spread

Spikes were evoked in ganglion cells using currents between 0.6 and 5 μA. When stimulated with 0.1-ms pulses, the average threshold current for 78 rat cells stimulated under similar conditions was 0.81 ± 0.03 μA, corresponding to a charge of 81 ± 3 pC and a charge density of 0.073 ± 0.005 mC/cm2. Electrode diameter ranged from 6 to 25 μm, and we present an analysis of thresholds as a function of electrode size below (see Fig. 10). In many cases, the lowest current setting of our stimulator (0.6 μA) yielded a superthreshold response, indicating that the reported average thresholds may be overestimated.

Thresholds were lowest when the recording electrode, rather than a neighboring electrode, was also used for stimulation. To examine whether spikes could be elicited by stimulating at a distance from the recording site, electrodes immediately adjacent to the recording electrode were used to stimulate. Figure 5A shows average results for eight cells, stimulated with one, three, or six adjacent electrodes (electrode diameter 15 ± 2.6 μm). The goal was to elicit the same long-latency spike using the different configurations of stimulation sites shown. Thresholds for spike initiation increased several-fold, depending on the number of active electrodes. In particular, when a single neighboring electrode was used for stimulation, about 3 times more current was needed compared with stimulation at the recording electrode. This indicates that a resolution of the order of the electrode spacing or finer (≤60 μm) can be achieved with minimal threshold stimulation.

FIG. 5.

Spatial spread of stimulation. A: spike thresholds for several stimulation configurations. Region adjacent to the center recording electrode (R) on the hexagonal array is shown, with filled circles indicating active electrodes used for stimulation, and open circles denoting unused electrodes. All active electrodes were stimulated simultaneously and at the same current amplitude. In addition to the standard setup (top left), 3 alternative configurations were tested: all neighboring electrodes, 3 electrodes, and a single neighboring electrode. Threshold currents are shown as averages over 8 cells and indicate lowest value found when multiple electrode combinations were tested. Corresponding charge densities were 0.09, 0.22, 0.25, and 0.32 mC/cm2. Electrodes were 60 μm apart. B: spikes evoked by a strong stimulus (0.35 mC/cm2) delivered through the electrode marked with the large radiating symbol. After application of TTX, the artifact was subtracted for each electrode individually. Spikes from 4 cells were detected (circles, triangles, squares, and diamonds), while most electrodes showed no evoked spikes (horizontal lines through electrodes). Empty electrodes were not tested. Traces are examples of TTX-subtracted signals recorded at sites indicated. Fast early deflections are stimulus artifacts; slower deflections are spikes. Scale was the same for all traces.

The preceding results suggest that stimulation using low stimulus amplitudes (<0.1 mC/cm2) usually affected only cells in the vicinity of the stimulation electrode. To further verify this, in 35 low-amplitude stimulation experiments (average charge density 0.071 ± 0.004 mC/cm2, electrode diameter 12.7 ± 0.5 μm), all electrodes surrounding the stimulation electrode were inspected for evidence of evoked spikes which differed in latency, shape, or reliability from the ones recorded on the center electrode. Such additional spikes would indicate recruitment of neurons at nearby locations. Of 186 neighboring electrodes analyzed for long-latency spikes, only 1 showed an additional evoked spike. However, additional evoked spikes were frequently seen on surrounding electrodes when the current was increased several-fold, suggesting recruitment of cells tens of micrometers distant, consistent with the results shown in Fig. 5A. Still higher currents sometimes elicited spikes on non-neighboring electrodes, >150 μm away from the stimulation electrode.

To further investigate spatial spread of activation, a more detailed analysis was performed to detect short-latency spikes around the stimulation electrode. In four experiments, we applied TTX and subtracted the averaged stimulus artifact on each electrode individually to reveal additional short-latency spikes, as in Fig. 4A. In two such experiments with a stimulus strength of ∼0.1 mC/cm2 (electrode diameter 10 μm), no short-latency spikes were found outside the 60 μm radius around the stimulation electrode. In two further experiments stimulated at 0.21 and 0.35 mC/cm2 (electrode diameters 7 and 6 μm), spikes were detected as far away as 160 μm. One example of strong-stimulus stimulation is shown in Fig. 5B: while the majority of electrodes on the array recorded no evoked spikes, four separate responses were elicited in the vicinity of the stimulation site. The spikes from these stimulated cells were each detected on two or more electrodes and the electrode recording the largest spike amplitude can be used to infer the approximate location of the soma. Most evoked spikes (circles, squares, diamonds) were recorded within 60 μm of the stimulation electrode, but one cell was detected nearly 160 μm away (triangles). These results show that the radius of stimulated ganglion cells can be controlled by adjusting the stimulus strength.

The above results were obtained by applying cathodic-first pulses (Fig. 1C). For most cells, thresholds were slightly higher when the anodic phase was delivered first: in 18 cells stimulated with 0.05- or 0.1-ms anodic-first pulses, spike thresholds were 115 ± 5% of the thresholds measured using cathodic-first pulses.

Strength-duration relationship

The current required to elicit a spike depended strongly on pulse duration. In all three species tested, higher currents were required to evoke a spike when shorter pulses were applied. Durations were varied from 50 μs to 1 ms and several resulting strength-duration curves are shown in Fig. 6A. In the examples plotted here, electrode diameter, stimulation configuration, and spike latency differed considerably across cells, resulting in a wide spread of threshold curves. Nevertheless, the slopes of these curves were similar in monkey, guinea pig, and rat, indicating that the threshold-duration relationship was independent of the species.

FIG. 6.

Strength-duration curves. A: current required to elicit a spike is plotted as a function of stimulus pulse duration for several individual ganglion cells from monkey (solid lines), guinea pig (dotted lines), and rat (dashed lines). Several monkey and guinea pig cells shown here were stimulated using a pulse generator capable of delivering current pulses below 0.6 μA. B: threshold current and charge in rat retina. Average thresholds of 25 cells recorded under identical conditions are plotted against pulse duration. Two configurations are shown: stimulation at the recording electrode (solid line, 13 cells) and stimulation at an electrode adjacent to the recording site (dashed line, 12 cells). Inset: corresponding threshold charge for the same 2 configurations.

To characterize each strength-duration curve by a time constant and an asymptote, power functions or exponentials were fit to the data (see methods). Rheobase is defined as the asymptote of the fit curve (Loeb et al. 1983; Ranck 1975) and chronaxie, the classical measure of responsiveness of a neuron, as the duration at which the threshold current is twice the rheobase (Lapicque 1907). The average chronaxie of 34 cells such as those shown in Fig. 6A was 407 ± 45 μs for 1/x fits, 338 ± 81 μs for power fits, and 212 ± 28 μs for exponential fits. The average rheobase was 0.51 ± 0.12, 0.60 ± 0.11, and 0.76 ± 0.16 μA, respectively. Fit quality was generally highest for power fits. Seven cells were from monkey, 7 from guinea pig, and 20 from rat, and all responses were long-latency spikes. When the group of 20 rat cells was divided into those stimulated with the recording electrode and those stimulated at a neighboring site, no difference in chronaxie was found (P > 0.5).

To facilitate comparison of thresholds in a single species and to show the influence of stimulation electrode position, Fig. 6B shows averaged data from 25 ganglion cells in rats. The cells were stimulated under identical conditions using electrodes with similar diameters (average 10.6 ± 0.5 μm). The solid line plots thresholds for stimulation at the recording electrode (13 cells), while the dashed line shows results from stimulation at a neighboring electrode (12 cells). As in Fig. 5A, eliciting a spike required several-fold higher currents when the site of stimulation was at an adjacent electrode. Because the charge delivered during the cathodic phase of the pulse is often used as a measure for stimulation strength, the inset plots the corresponding charge thresholds: charges were consistently <200 pC for pulse durations up to 0.2 ms, corresponding to charge densities <0.22 mC/cm2.

To further corroborate the above notion that short- and long-latency spikes constitute doublet responses, we measured strength-duration curves of responses with latencies <2 ms. The average strength-duration relationship in 14 cells with short-latency spikes (latency 0.69 ± 0.08 ms) was similar to that of long-latency responses: chronaxies determined from fit curves (as above) were 571 ± 149, 299 ± 52, and 311 ± 113 μs; none of these values was significantly different from long-latency chronaxies (P > 0.1; n = 45 cells). These results suggest that the same neuronal element was excited in both short- and long-latency responses.

Frequency dependence

To mimic natural spike trains, a retinal implant must be capable of delivering pulses and evoke spikes at a wide range of stimulation frequencies. Furthermore, continual stimulation at higher frequencies may be a requirement for generating sustained percepts of light. We tested stimulation at pulse frequencies of up to several hundred hertz.

To examine high-frequency responses, two closely spaced pulses were applied, with the interpulse interval corresponding to frequencies of up to 200–300 Hz. High-frequency stimulation was deemed successful when spikes were evoked after the second stimulus pulse. Pulse pairs were applied for 10–20 s at intervals of 0.5 s and at stimulus strengths of about twice threshold. In nine cells tested, spikes were evoked on the second pulse on 99 ± 1% of trials at 100 Hz and 94 ± 4% at 200 Hz. Three cells were stimulated with 300-Hz pulse pairs and all responded at >90% of trials. All responses were short-latency spikes (latency 0.8 ± 0.2 ms). Data from such an experiment is shown in Fig. 7A: a superthreshold 300 Hz pulse pair reliably produced two short-latency spikes, more clearly seen in the artifact-subtracted traces shown below. A superposition of several stimulus trials is shown to demonstrate repeatability. In four cells, TTX was added to the bath solution to facilitate artifact subtraction and spike detection (data not shown).

FIG. 7.

Frequency dependence of stimulation. A: top: superthreshold responses to 300-Hz pulse pairs, applied every 500 ms in rat retina. Several superimposed trials are shown. Bottom: artifact subtraction revealed spikes at latency 0.75 ms. Arrowheads indicate stimulation onset. Spike-free traces necessary for artifact subtraction were obtained by decreasing the interval between pulse pairs until failures were observed. B: responses to sustained stimulation at 32 Hz near threshold in a different cell. A superposition of 73 trials is shown, about half of which evoked a spike (arrow). Inset: artifact subtraction of averaged traces shows evoked spike at latency 0.4 ms, along with the subtracted spike for low-frequency stimulation, and a spontaneous spike. Arrowhead indicates stimulation onset. Scale bar, 1 ms and 20 μV. C: percent evoked spikes (response rate) for spikes with latency 0.6 ± 0.1 ms (4 cells, closed symbols) and spikes with latency 9.1 ± 1.5 ms (4 cells, open symbols) for sustained stimulation at up to 100 Hz. Dashed lines are sigmoidal fits to data.

To test responses to brief periods of sustained high-frequency stimulation, nine cells were continuously stimulated for 5–20 s at frequencies of up to 100 Hz. Figure 7B shows an example of responses to over 70 stimulus pulses near spike threshold, delivered at 32 Hz. Short-latency spikes were evoked on roughly one-half of the trials (arrow) and were used to subtract the artifact (inset). Thirty-two-hertz stimulation evoked spikes indistinguishable from those produced by 2-Hz stimulation or spontaneous activity.

Response rates, defined as the number of evoked spikes in a stimulation period, were measured at sustained pulse frequencies of up to 100 Hz at stimulus strengths of about twice threshold (Fig. 7C). Short-latency spikes (closed symbols) showed a slight reduction of the response rate at 50 Hz (<20%) and a significant drop at 100 Hz. We also observed a gradual reduction in spike amplitude throughout the stimulation period at frequencies above about 32 Hz (data not shown). Strikingly, long-latency responses (open symbols) were robust only up to 5 Hz and virtually no spikes were observed above 10 Hz. This observation was corroborated in one cell with a short-latency spike (0.7 ms), which was followed by a spike at latency 5 ms: stimulation at low frequencies consistently evoked both responses, while only the short-latency spike was observed at frequencies above 8 Hz.

We conclude that short-latency spikes can be reliably evoked in ganglion cells at pulse frequencies up to about 50 Hz and that late spikes are suppressed at moderate frequencies.

Sustained stimulation

Chronic retinal implants must be capable of delivering effective stimulation pulses over a period of many hours each day. To determine whether sustained low-frequency stimulation could reliably evoke spikes, we extended our stimulation period to the longest duration that was experimentally feasible.

Two cells were continuously stimulated for 30 min and two additional cells for 4.5 h. The longest sustained stimulations were performed using 0.8-μA pulses with 0.1-ms duration, delivered at frequencies of 1–2 Hz, and corresponding to a charge density of about 0.04 mC/cm2 per pulse (electrode diameter 16 μm). Figure 8 shows an example of spikes evoked before and after a 4.5-h stimulation period: the cell showed robust responses after having been stimulated with over 16,000 pulses. A slight increase in threshold and spike latency (∼20%) was noted at the end of the stimulation period.

FIG. 8.

Continuous low-frequency stimulation. Example of spikes evoked at the beginning (top traces) and at the end of a 4.5-h stimulation period with 0.8-μA pulses (0.1 ms, 1 Hz). Four response traces are shown for each time-point.

Multielectrode stimulation

To generate artificial vision, a functional retinal implant requires independent activation of many closely spaced electrodes. To study responses to spatial stimulation patterns, the multielectrode array was used to stimulate at several electrodes simultaneously. Our goal was to show that simultaneous activation of two or more nearby electrodes did not influence each other. If that were the case, spikes elicited during multielectrode stimulation should not differ in threshold, shape, or number from individual stimulations.

We selected seven sites that clearly showed evoked long-latency spikes when stimulated individually. These evoked spikes differed in spike shape and latency, but had similar thresholds. All seven electrodes were subsequently activated simultaneously using 0.8-μA pulses (0.1-ms duration). Figure 9A shows spikes evoked at these sites and their locations on the array. Simultaneous stimulation evoked seven distinct responses on seven spatially disparate electrodes.

FIG. 9.

Multiple site stimulation. Rat retina was stimulated at 7 electrodes simultaneously with 0.8-μA pulses. A: overlay of several trials is shown for each electrode (1–7) and evoked long-latency spikes marked with an asterisk. Inset (top right): location of active electrodes on the array. Latencies ranged from 5 to 18 ms. B: traces from neighboring electrodes 1 (left) and 2 (right). For comparison, spikes are shown for individual stimulation at only that electrode (single) as well as when all 7 electrodes were active (all). Evoked spikes showed no difference. Arrowhead indicates that the large spikes seen on electrode 2 were visible on electrode 1 as small deflections.

To establish that the spikes evoked by simultaneous stimulation did not differ from those evoked by individual stimulation, traces recorded at each electrode under both conditions were compared. Two examples are shown in Fig. 9B: individually evoked spikes (single) were identical to simultaneously evoked spikes (all) for both electrodes shown here. Furthermore, Fig. 9B shows that stimulation at neighboring electrodes evoked independent responses. While this was expected given the low currents used here (see Fig. 5A), these data clearly establish that adjacent electrodes (1 and 2) did not influence each other during simultaneous stimulation. Only a small-amplitude deflection was recorded on electrode 1 at the latency of the spikes seen on electrode 2 (arrowhead), indicating that the cell stimulated by electrode 2 was probably located close enough to electrode 1 to be recorded as small spikes. To further ensure spatial precision, all 22 inactive electrodes surrounding the 7 active stimulating electrodes were inspected for spikes. While four adjacent electrodes showed small spikes that were recorded on one of the seven stimulation electrodes, none recorded new spikes.

Multielectrode stimulation was performed five times using various electrode arrays and spatial patterns, with results very similar to the data presented above. We conclude that evoking independent spikes on multiple electrodes spaced 60 μm apart is feasible with minimal cross-electrode interaction.

Electrode size

This study used arrays with electrodes that varied in diameter between 6 and 25 μm. To determine the influence of electrode size on thresholds, stimulation results were compared in a set of 86 cells for which the exact platinum disk diameters of the stimulating electrode was measured. Figure 10 shows thresholds as a function of electrode diameter, both for cells stimulated using pulse durations of 0.1 and 0.05 ms. Current, charge, current density, and charge density are plotted. All spikes were long-latency spikes (average 7.6 ± 0.3 ms).

FIG. 10.

Dependence of thresholds on electrode size. Thresholds were plotted against electrode diameter for 78 rodent cells stimulated with 0.1-ms pulses (•) and 59 cells stimulated using 0.05-ms pulses (○). Averages in bins of 2–3 μm are shown. A: current. B: charge. C: current density. D: charge density. Each is plotted for the same set of data. A single electrode was used for both stimulation and recording, and all responses were long-latency spikes. Error bars were omitted when SE was smaller than symbol size.

Threshold current and charge (Fig. 10, A and B) increased by a factor of 2–3 between the smallest and the largest diameters, indicating that with smaller electrodes, less current and charge injection was necessary to elicit spikes in ganglion cells. Conversely, current density and charge density (Fig. 10, C and D) was drastically decreased for electrodes >10–15 μm. Note that the lowest threshold values plotted here may have been overestimated because the minimal available current setting (0.6 μA) often evoked a superthreshold response.


This study used dense arrays of small-diameter electrodes to electrically stimulate rat, guinea pig, and primate retina. We described the responses of individual ganglion cells to a wide range of pulse configurations and spatial stimulation patterns and showed that effective stimulation is feasible with high temporal and spatial precision. Our findings imply that the electrode size of future epiretinal prosthetics may safely approach the cellular dimensions of retinal ganglion cells.

Evoked spikes

Long-latency spikes (>2 ms) were readily observable, whereas spikes with shorter latencies could only be observed with digital artifact subtraction. Note that this classification into “short” and “long” latencies differs somewhat from that of other researchers: Jensen et al. (2005b) defines short as 3–5 ms and long as ≥9 ms; Stett et al. (2000) classifies spikes at latencies 1–10 ms as early and spikes at 10–20 ms as delayed. The definition of Crapper and Noel (1963) is more similar to the one used in this study: immediate spikes were defined as those around 0.5-ms latency and later responses as 5–15 ms.


The earliest observed responses occurred within several hundred microseconds of stimulation onset and probably represent the immediate activation of the ganglion cell spike generator. Actual latencies are in effect shorter than the reported values by 50–100 μs because of a delay introduced by the amplifier circuitry. Furthermore, if measured from the termination of the cathodic phase, true minimal latencies observed in this study amount to 50–150 μs. While these values are lower than the latencies in many studies, submillisecond spikes have been reported by a few authors (Crapper and Noell 1963; Grumet et al. 2000).


We showed that evoked spikes at longer latencies are preceded by short-latency spikes. Long-latency spikes can thus be used to infer short-latency spikes even when the magnitude of the pulse artifact confounds direct observations of early spikes. However, it is possible that thresholds for long-latency spikes are systematically overestimated, because short-latency responses typically occur at lower stimulation currents than doublet responses.

Multispike responses are consistent with known intrinsic firing properties of ganglion cells, in which doublets or triplets in spike trains occur with interspike intervals of several ms (Devries and Baylor 1997). In our study, interspike intervals ranged from 4 to 16 ms, with the majority of long-latency spikes occurring ∼5–7 ms after the short-latency spikes. We favor the interpretation of long-latency spikes as part of a doublet response over other suggestions such as conduction delays (Jensen et al. 2005b) or intracellular charging mechanisms for the following reasons: the small electrodes and currents used here make activation several millimeters from the recording site exceedingly unlikely; in each case tested, every long-latency spike was preceded by a short-latency spike; and earlier studies have not employed artifact subtraction methods, thus seeing only later spikes. Furthermore, long-latency spikes are only observed at stimulation frequencies <10 Hz, suggesting that higher repetition rates suppress multispike bursts.

Stimulation thresholds


An important prerequisite of implantable stimulators is their capability to deliver current that is safe, yet efficient. Unsafe stimulation can originate from two sources: electrochemical destruction of the stimulating electrode (such as corrosion) and neural tissue damage induced by toxic products near the electrode or by neuronal hyperactivity. Several electrochemical safety limits have been proposed, such as the often-stated nongassing limit of 0.3–0.4 mC/cm2 for platinum electrodes (Brummer and Turner 1977). More recently, limits as low as 0.1 mC/cm2 for cathodic stimulation with platinum electrodes have been recommended (Rose and Robblee 1990). Thresholds for tissue injury in cortex have been shown to arise from the synergistic interaction between charge and charge density: as the charge is increased, the charge density for safe stimulation decreases (McCreery et al. 1986, 1990; Merrill et al. 2005). The data of McCreery et al. show that no histologically detectable damage is produced with low-charge stimulation (<50 nC) even when the charge density is >1 mC/cm2, whereas for pulses delivering a higher charge (1 μC), the damage threshold is <0.1 mC/cm2. In the absence of detailed threshold measurements, concerns have been raised regarding the feasibility of using small-diameter electrodes in human patients, because they have been suggested to require much higher charge densities for threshold stimulation than large electrodes (Brummer et al. 1983; Greenberg 1998; Loeb et al. 1983). However, we found in this study that threshold stimulus pulses are characterized by low currents (∼1 μA), low charge injection (∼100 pC), and low charge densities (∼0.1 mC/cm2) despite the small electrode size. Several cells had threshold charge densities of <0.03 mC/cm2, an order of magnitude lower than the platinum electrode safety limit. Furthermore, while we have used the geometric electrode area to calculate current and charge densities, the effective electrode area likely was significantly larger. Electroplated platinum tends to deposit in granular surface structures that greatly increase the area of metal in contact with the solution. It has been reported that the fractal-like platinum deposits can increase the surface area by up to 100 times (Kim and Oh 1996; Mathieson et al. 2004). Thus all density values reported here should be considered upper limits, further reducing the likelihood of electrochemical electrode damage.

Our results complement data recently reported for small-diameter needle electrodes, which have described threshold charge densities between 0.15 and 0.3 mC/cm2 (Jensen et al. 2003; Rizzo et al. 1997; Wilms et al. 2003; Wyatt et al. 1994).


One factor contributing to the low thresholds in this study is the tight contact between electrodes and tissue. This was a requirement in our experiments because extracellular spikes cannot be recorded without close juxtaposition of the retina to the array. Novel techniques to minimize the gap between retina and epiretinal implant are being developed (Johnson et al. 2004; Schanze et al. 2002) and may ensure close contact in future prosthetic devices.


We observed lower threshold current and charge for the smaller electrodes in this study than for the larger ones (see Fig. 10). However, the resulting charge density is increased for smaller electrodes. As electrode diameter drops below about 10 μm, the decrease in surface area outweighs the current decrease. It has been suggested that for electrodes smaller than the cellular size (∼10 μm), the electric field is concentrated in too small an area for effective stimulation (Palanker et al. 2004). Thus electrode diameters around 10–15 μm may be the optimal size for selective single cell stimulation and might be an ideal compromise between excellent spatial resolution and high charge density. This size range would also have less stringent requirements on the distance between electrode and cells, since stimulation with <10 μm electrodes is disproportionately more sensitive to this distance (Palanker et al. 2004). Clearly, this issue will need to be readdressed once technical advances in retinal prosthetics call for even smaller electrodes as the ratio of electrodes to ganglion cells approaches 1.


One consequence of the low required stimulation strengths was the exceedingly localized nature of stimulation: excited cells were limited to a narrow radius around the stimulating electrode and pharmacology experiments further confirmed that ganglion cells were directly activated: spikes were not suppressed in the presence of CNQX, APV, and kynurenate, which block excitatory transmission in the retina (Fujimoto and Toyoda 1991; Stett et al. 2000). This is a much more local effect than can be achieved with larger electrodes: indirect spikes sensitive to synaptic blockers have been reported for 125-μm electrodes (Jensen et al. 2002; Ziv et al. 2002) and larger electrodes (Greenberg 1998; Shimazu et al. 1999). Our results from simultaneous stimulation using multiple electrodes further confirm that the current spread in the plane of the electrode array is small enough to allow for independent activation of cells using neighboring stimulation electrodes.

Thresholds increase with the distance between stimulating and recording electrode on the array (see Fig. 5A). The observed increases are similar to those of the cathodal stimulation map reported by Jensen et al. (2003): stimulating about 60 μm away from the center of the receptive field required 2–8 times more current to elicit a spike.

Stimulation at amplitudes significantly above threshold activates cells in a larger radius around the stimulation site (Fig. 5B). However, such high stimulus strengths are unlikely to be required during normal operation of a high-resolution implant. A large number of perceived gray levels may be achieved by varying the pulse frequency, thus eliciting spike trains of different rates in individual ganglion cells (Fig. 7C). Limiting stimulation strength to 2–3 times the threshold therefore can achieve a high spatial resolution without reducing the dynamic range of stimulation.

We conclude from our observations that retinal implants with small electrodes can achieve a high spatial resolution, because the low required currents activate single (or at most a few) ganglion cells.


The use of pulses significantly longer than chronaxie contributes little to the evoked response, stipulating pulse durations smaller than chronaxie to insure that most of the applied charge contributes to evoking a response (Tehovnik 1996). Thus from the chronaxies measured in this study (∼100–400 μs), we conclude that optimal pulse durations should not exceed this range.

The measured values are similar to those reported in other studies (Crapper and Noell 1963; Greenberg 1998; Grumet et al. 2000; Jensen et al. 2005b) and can further be used to identify the neuronal element most likely excited by electrical stimulation. Our chronaxies match those reported for activation of axons (Grumet et al. 2000; Holsheimer et al. 2000; Nowak and Bullier 1998), because cell bodies and dendrites have chronaxies of 1–10 ms (Holsheimer et al. 2000; Ranck 1975). Because the initial axon segment near somas is more excitable than cell bodies (Greenberg et al. 1999; Nowak and Bullier 1998; Porter 1963; Schiefer and Grill 2002), the juxtasomal electrode used here likely activates this initial region on the axon and action potentials subsequently backpropagate a short distance to elicit the recorded somatic spike. Computational models suggest that excitation occurs near the junction of ganglion cell soma and axon or slightly more distal on the axon (Fohlmeister and Miller 1997; McIntyre and Grill 1999; Schiefer and Grill 2002).

It is difficult to experimentally rule out the activation of passing axons, in particular because tests designed to identify antidromic responses (Fuller and Schlag 1976) would not distinguish between initial axon segment excitation and more distant axon activation. Nevertheless, activation of passing axons is deemed less likely by the fact that ganglion cell axons have high thresholds away from their initial segment (Loeb et al. 1983). Because in retinal ganglion cells of most mammals (including human and rat) the axon remains unmyelinated within the retina, sodium channels are found uniformly throughout the unmyelinated region distant from the soma (Boiko et al. 2003). At the initial axon segment, however, the density of sodium channels is exceptionally high (Wollner and Catterall 1986), with clustering of the Nav1.6 subunit in particular (Boiko et al. 2003). This difference in channel density between the initial and distant region can amount to an order of magnitude or more (Ritchie et al. 1976; S. R. Levinson, personal communication). Like the nodes of Ranvier in myelinated fibers (McIntyre and Grill 2000), the initial portion of unmyelinated axons constitutes the most likely site of electrical excitation, perhaps at the “thin segment” 10–40 μm from the cell body (Boiko et al. 2003; Fohlmeister and Miller 1997).

Comparative literature analysis

To discuss data from this study in the context of previous work, a comprehensive review of the published literature was composed. Table 1 summarizes 32 studies that have reported epiretinal stimulation thresholds. These studies span several orders of magnitude in electrode size and can thus be used to elucidate threshold trends. Several key parameters have been graphed in Fig. 11, along with best fit lines and correlation coefficients. To facilitate comparison of different electrode geometries across studies, threshold parameters were plotted against the geometric electrode surface area (see methods).

FIG. 11.

Literature analysis. Data from Table 1 plotted for current, charge, current density, and charge density at threshold. Human studies are shown as open circles and representative examples from this study as crosses. Solid symbols indicate animal studies which either recorded neuronal responses in the retina (circles) or in the visual cortex (triangles). Dashed lines are best fits to the data, with the correlation coefficient R2 and the fit line equation shown for each graph.


Several factors contribute to the relatively wide scatter of points in Fig. 11. Threshold was defined inconsistently from study to study, spanning the range of 50–90% probability of eliciting a spike, cortical recordings, and human percept reports. Furthermore, while the majority of studies used charge-balanced biphasic pulses, several reported monophasic stimulation (typically cathodal), leading to lower thresholds in some cases. Moreover, human studies (open symbols) usually involved degenerated retinas, whereas animal studies were typically performed on normal tissue. Finally, studies which measured retinal responses by monitoring cortical activity (triangles) may have overestimated spike thresholds in ganglion cells, because the concerted activity of many cells is typically required for a cortical response.


Both threshold current and charge (Fig. 11, A and B) decrease dramatically as electrode size is reduced. Correlation was highest for threshold charge, because it takes into account both current and pulse duration, which varied across studies. These trends confirm that smaller electrodes require several orders of magnitude lower currents to elicit responses. They also mirror the results found within this study over a much more narrow range of electrode sizes: current and charge thresholds were small for small stimulating electrodes and large for large electrodes (Fig. 10, A and B).

Current density also increases somewhat when electrode size is reduced (Fig. 11C), owing largely to the fact that studies using smaller electrodes also tended to apply shorter pulses (see Table 1). Such a trend is not seen in the plot of charge density thresholds (Fig. 11D). While the large variability does not permit an accurate fit to the data, there is no definitive change of charge density with electrode size, such as is seen for threshold charge over four to five orders of magnitude. In fact, charge density is virtually independent of electrode size for electrodes smaller than 104 μm2 (disk diameter about <100 μm). This trend is in contrast to the increased charge densities observed for the smallest electrodes used in this study (see Fig. 10D), which is probably caused by an effect restricted to electrodes smaller than ∼10 μm diam.

Representative data from this study (monkey and rat) have been included in Fig. 11 and fit well with the trends established by the published literature. Our data substantiate the main conclusion from this analysis: small electrodes require much less charge injection for threshold stimulation than larger electrodes, but the accompanying increase in charge density is almost negligible.


Because both charge and charge density must be considered when discussing stimulation safety (Merrill et al. 2005), Fig. 12 shows a plot of both parameters for the same set of studies. Two types of safety limit lines were included: the often used electrochemical limits for platinum (0.35 and 0.1 mC/cm2) and iridium oxide electrodes (1 and 4 mC/cm2), as well as the limits for neural injury from cortical stimulation data (k = 1.7 and 2.0; see methods). To show the spread of thresholds measured in this study, Fig. 12 includes data points for all cells stimulated using 0.05-ms pulses and electrodes with diameters between 6 and 25 μm (crosses). Most human studies (and several animal studies) fall near or outside of the safe region formed by the neural injury limits, possibly because degenerated retina requires higher currents to produce phosphenes in humans. While most data points lie close to or above the electrochemical limits for platinum, it should be emphasized that the increase in effective surface area afforded by the electroplating process raises the true safety limits by at least an order of magnitude. Furthermore, electrodes made of materials such as iridium clearly have higher electrochemical limits (Beebe and Rose 1988; Merrill et al. 2005; Weiland et al. 2002) and may be used to substantially extend the region of safe stimulation.

FIG. 12.

Threshold charge density plotted against charge. Data from Table 1 were used, along with several safety limits, indicating the region of potentially safe and unsafe stimulation parameters. The plotted electrochemical limits are for platinum (0.1 and 0.35 mC/cm2, dotted lines) and iridium oxide (1 and 4 mC/cm2, dashed lines). Neural injury limits (dash-dotted lines) are from cortical stimulation data (k = 1.7 and 2. 0). Crosses are thresholds for 0.05-ms pulses from this study spanning all electrode diameters (6–25 μm; symbol size is proportional to electrode diameter).

Figure 12 validates our claim that the small electrodes used in this study can safely stimulate mammalian retina: except for the data collected using the smallest electrodes, most thresholds are well within all safety limits.


The purpose of this study was to elucidate basic stimulation parameters to test whether a future generation of implants could incorporate a design using significantly smaller electrodes than are presently available. We used planar microelectrode arrays that closely resemble those currently in use for chronic human testing (Humayun et al. 2003) but contain much smaller electrodes at a much smaller electrode spacing. We suggest that future implants could directly activate ganglion cells instead of affecting large areas of retina by indirect stimulation, making possible a reasonable spatial resolution of artificial sight. One can envision high-resolution arrays containing thousands of stimulation sites with diameters around 10–20 μm and separation between electrodes of 20–60 μm. While years away, results from this study suggest that there is no fundamental hindrance to the feasibility of such a device. Once implanted, the stimulus parameters can be adjusted to stimulate individual or small overlapping groups of ganglion cells, depending on the desired phosphene size. By using low currents, activation of axon bundles can be avoided. As a next step toward the development of such implants, further experiments using small electrodes with degenerated retina are warranted.


This research was supported by the Salk Institute Pioneer Postdoctoral Fellowship, Second Sight Medical Products Inc., McKnight Foundation National Eye Institute Grants EY-13150 to (E. J. Chichilnisky) and EY-12893 (to Second Sight Medical Products), National Science Foundation Awards PHY-0245104 and PHY-0417175 (to A. M. Litke), Polish State Committee for Scientific Research, Project 3 T11E 011 27 (to W. Dabrowski), and NATO Collaborative Linkage Grant (to A. M. Litke and W. Dabrowski).


We thank M. McMahon for constructive comments on the manuscript and vital feedback throughout this project.


  • The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked “advertisement” in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.


View Abstract