Simulating PRIMA in age-related macular degeneration

In geographic atrophy, central vision is lost and peripheral vision remains. The subretinal photovoltaic implant PRIMAPivotal is placed inside the lesion and produces a coarse grayscale percept there.

This example covers scene, scotoma, prosthetic percept, gaze, device preprocessing, and both eyes. The objects are documented under Visual Input and Simulated Vision.

import matplotlib.pyplot as plt

from pulse2percept.implants.retina import PRIMAPivotal
from pulse2percept.models.retina import ScoreboardModel
from pulse2percept.stimuli import samples
from pulse2percept.units import dva
from pulse2percept.vision import BinocularScene, Scene, Scotoma

Residual vision

The lesion and the implant share an eccentric center (dva).

center = (6, -2) * dva
logo = samples.logo_bvl(resize=(240, 300))
scotoma = Scotoma.ellipse(5 * dva, 4 * dva, center=center)

scotoma_fill sets what is shown inside the loss. Simulations often use black or gray, but people with AMD rarely report a black spot and are often unaware of the loss. 'inpaint' fills the scotoma from the surrounding vision (skimage.restoration.inpaint_biharmonic()). This is a frame-local image operation, not a model of perceptual filling-in.

background=1 shows the logo’s transparent pixels as white. rings=True draws 5 dva eccentricity rings.

filled_in = Scene(logo, fov=40 * dva, scotoma=scotoma,
                  scotoma_fill='inpaint', background=1)

filled_in.plot(gaze=(0, 0) * dva, rings=True)
plt.title('Native vision alone, with filling-in')

Text(0.5, 1.0, 'Native vision alone, with filling-in')

'inpaint' is not combined with prosthetic percepts, so the rest of the example uses a numeric fill.

scene = Scene(logo, fov=40 * dva, scotoma=scotoma, scotoma_fill=0,
              background=1)

Prosthetic vision inside the loss

PRIMA is driven by pulsed 880 nm light. Its default PRIMAEncoder converts gray levels to projector ON durations. implant_position places the array at the lesion center, and the model grid covers the field around it.

implant = PRIMAPivotal()
model = ScoreboardModel(implant=implant, implant_position=center, rho=50,
                        xrange=(0, 12), yrange=(-8, 4), step=0.05)

percept = model.predict_percept(scene, gaze=(0, 0) * dva)

scene.plot(percept=percept, gaze=(0, 0) * dva, vmax=2)
plt.title('Native vision with a PRIMA percept in the scotoma')

Text(0.5, 1.0, 'Native vision with a PRIMA percept in the scotoma')

The scene is drawn at its own 240 x 300 pixels and the percept on its 0.05 dva model grid; neither is resampled. scene.render(...) returns a single RGB raster when one is needed.

Gaze

The lesion and the implant move with the eye; the scene does not:

percept = model.predict_percept(scene, gaze=(8, -4) * dva)

scene.plot(percept=percept, gaze=(8, -4) * dva, vmax=2)
plt.title('Looking 8 degrees right and 4 degrees down')

Text(0.5, 1.0, 'Looking 8 degrees right and 4 degrees down')

Device preprocessing

implant.preprocess applies to the prosthetic input only. An edge filter changes the percept; native vision is unchanged. Edges light fewer pixels, so vmax is lowered:

implant.preprocess = lambda stim: stim.filter('sobel')

percept = model.predict_percept(scene, gaze=(0, 0) * dva)

scene.plot(percept=percept, gaze=(0, 0) * dva, vmax=0.3)
plt.title('Edge-filtered device input, intact vision around it')

Text(0.5, 1.0, 'Edge-filtered device input, intact vision around it')

Both eyes

A Scene is one eye, so the fellow eye gets its own scene, here with a smaller, partial loss. A BinocularScene holds the pair. aperture='round' draws each field as an ellipse (display only).

implant.preprocess = False

worse_eye = Scene(logo, fov=40 * dva, scotoma=scotoma, scotoma_fill=0,
                  background=1, aperture='round')
fellow_eye = Scene(logo, fov=40 * dva, background=1, aperture='round',
                   scotoma=Scotoma.circle(3 * dva, center=center),
                   scotoma_fill=0.4)

binocular = BinocularScene(left=worse_eye, right=fellow_eye)

Models are monocular: the percept is predicted for the implanted eye and drawn on that side. The fellow eye shows only its own scene.

prosthetic_input = Scene(logo, fov=40 * dva, background=1)
percept = model.predict_percept(prosthetic_input, gaze=(0, 0) * dva)

binocular.plot(left_percept=percept, vmax=2, rings=True)

(<Axes: title={'center': 'left eye'}, xlabel='x (degrees of visual angle)', ylabel='y (degrees of visual angle)'>, <Axes: title={'center': 'right eye'}, xlabel='x (degrees of visual angle)', ylabel='y (degrees of visual angle)'>)

What this does not establish

• ScoreboardModel shows where the light lands, normalized to a fully lit pixel. It does not model photovoltaic transduction or retinal activation (Ho2018Model does).

• rho = 50 um is an illustrative spread, not a measured PRIMA point-spread function.

• Combining prosthetic and residual vision inside the scotoma is a display convention; their perceptual interaction is not modeled.

• Residual acuity, binocular combination, disparity, and depth are not modeled.