Retinal prostheses

In retinitis pigmentosa, and in the geographic atrophy of age-related macular degeneration, photoreceptors die while much of the inner retina, which processes their signals, survives. That invites the bet that paid off in the cochlear implant: skip the dead transducers and drive the surviving neurons with current. No design to reach patients has done for sight what the cochlear implant did for hearing, and the reasons say more about the retina than about the electrodes.

Two places to put the electrodes

On the inner surface. Second Sight’s Argus II tacks a 6 × 10 grid of platinum discs, each 200 µm across, onto the ganglion cell side of the retina. Frames from a camera on glasses are shrunk to 6 × 10 and mapped onto the electrodes; a coil on the glasses sends power and data by radio to a coil and electronics case sutured to the eye, and a cable through the eye wall reaches the array. In the 30-person trial behind its approval, the share doing significantly better with the system on than off was 96% for locating an object, 57% for the direction of motion and 23% for judging the orientation of gratings; the best acuity was 20/1260 (Humayun and colleagues, 2012). The FDA approved it in February 2013 under a humanitarian device exemption, which asks for probable benefit rather than proof of effectiveness.

Underneath it. Daniel Palanker’s group at Stanford went to the other side of the retina and did away with the wires (Mathieson and colleagues, 2012). In its clinical form, PRIMA, the implant sits where the photoreceptors were, against the bipolar cells they used to drive — a synapse before the ganglion cells, which its developers argue keeps some of the inner retina’s processing in the loop. Each pixel is a small solar cell: glasses project the camera’s image into the eye in near-infrared light at 880 nm, and the pixel turns the light landing on it into pulses of current between a central electrode and a return electrode around its edge. The 2 mm square chip carries 378 pixels, each 100 µm across.

Why sixty electrodes buy less than twenty

A cochlear implant has 12 to 22 electrodes and, by Wilson and Dorman’s 2008 review, no more than four to eight effective channels: broadly overlapping sectors of an auditory nerve of some 30,000 neurons. The average user still follows a predictable conversation in quiet with relative ease, because speech survives being cut down to a few places along the cochlea’s frequency map, each carrying an envelope.

Vision has no such shortcut. Sighted volunteers walking a maze through a viewer that broke the scene into dots needed a 25 × 25 grid, 625 punctate dots, for useful mobility where little pattern recognition was asked of them (Cha, Horch and Normann, 1992). Argus II has a tenth as many, and its dots are neither punctate nor regular, for three reasons.

Size. Each 200 µm disc covers the area of hundreds of photoreceptors, so it excites many kinds of cell at once, ON and OFF pathways together, where light drives such complementary pathways selectively (Beyeler and colleagues, 2019).

Axons. Ganglion cell axons sweep across the inner surface towards the optic disc, so an epiretinal electrode sits on cables from cells elsewhere. When four users of Second Sight’s implants drew what single electrodes produced, the blobs, streaks and wedges followed the nerve-fibre bundle under each electrode; in Beyeler’s model, activation decayed over 500–1,420 µm along an axon against 144–437 µm across it. Weitz and colleagues found a way round: in isolated retinas, 25 ms pulses, about a hundred times the usual length, drove the inner retina and spared the axons, and in a patient they gave focal spots (2015).

Types. A human retina has 0.7 to 1.5 million ganglion cells (Curcio and Allen, 1990). In primates they come in at least 17 types, each thought to tile the retina and carry its own cues — particular spatial and temporal frequencies, luminance or colour contrast — so every point in the visual field is reported several times over (Nassi and Callaway, 2009). E. J. Chichilnisky’s group could often fire one cell of any of the five commonest types without its neighbours, on isolated macaque retina with electrodes 9–19 µm across (Sekirnjak and colleagues, 2008; Jepson and colleagues, 2013). They could tell the cells apart because those retinas still responded to light: each cell’s spikes were sorted from the array’s recordings and its type read largely from its light responses. A blind eye would need another way.

So the retina’s output is not a picture but at least seventeen overlapping descriptions of one — the difference from a camera that the silicon retina set out to copy — with the axons carrying them laid across the surface an epiretinal array sits on. A grid of electrodes there writes one signal into every description under each contact and into the passing axons, which is what the patients drew. Speech tolerated a coarse write to a frequency map; this code does not.

Powered by light

Coil, case and cable are what Mathieson and colleagues argued scale poorly to the dense arrays that acuity needs. A photovoltaic pixel needs none of them, but its energy budget is the light on its own area, and heating caps the light: they put the limit at an average of about 5.2 mW/mm² on the retina at 905 nm, to keep the temperature rise under 1 °C. Voltage is a second budget, because across the electrode–tissue impedance it sets how much charge the electrode can inject. A single silicon photodiode gives at most about 0.5 V at safe intensities; three in series give 1.5 V and triple an iridium oxide electrode’s charge injection, from 0.5 to 1.5 mC/cm², but need three times the light, each having a third of the area — the trade any series string of solar cells makes. The clinical pixel has two.

The pivotal PRIMAvera trial, reported in October 2025, implanted 38 people with geographic atrophy (Holz and colleagues). Without zoom, mean prosthetic acuity at a year was 20/417, matching the sampling limit of a 100 µm pixel: the limit had become the pixel pitch, an engineering number. Of the 32 assessed at a year, 26 had gained at least ten letters on a chart, and with digital zoom some read print as small as 20/42.

What it cost

Argus II’s gains were real and small. Seven in ten trial participants had no serious adverse event; the commonest was erosion or opening of the conjunctiva over the parts outside the eye, and one device had to come out. As Beyeler and colleagues summarised in 2019, most users could not report a grating’s orientation, and those who could recognise letters needed more than 40 seconds to do so.

Then the maker withdrew. In May 2019 Second Sight’s board chose to concentrate on Orion, a stimulator on the surface of the visual cortex that bypasses the eye, and suspended production of Argus II. In spring 2020, citing the pandemic and a lack of financing, it laid off most of its staff, and its annual report for 2020 recorded a decision to withdraw Argus II from the market, ending post-market studies in Germany and the US and suspending technical support worldwide. Over 300 people had been implanted. Their devices had outlived their maker’s interest in them.

PRIMA’s figures carry costs of their own. The trial was open-label and single-arm, and paid for by Science Corporation. Of the 38 participants, 19 had serious adverse events, 26 in all, mostly within two months of surgery and mostly resolved within two more; the commonest was raised pressure inside the eye. Without zoom the prosthetic image is about as sharp as the peripheral vision these patients kept, and the first five patients saw it as white-yellow (Palanker and colleagues, 2020): resolution up to the pixel, but no colour.