Figure 1.
A subject is shown performing the experiment (left) along with a close-up of the goggles with eye tracking camera (right). In use, the head frame is adjusted to the seated height of each subject, and the monitor position adjusted to maintain a consistent distance to each subject's eyes. The base of the head frame is mechanically secured to the desk, but the upper part can be raised or lowered. The camera is mounted on a flexible arm that holds position once adjusted for a close-up view of the subject's eye. Experiment workstations and experimentor controls are not shown.
Figure 2.
Each gray rectangle represents the image shown on the computer monitor during one phase of a trial. Subject eye position is shown as a red circle. The Free View period is represented here under simulated thalamic prosthetic vision as in the Primary experiment: each dot (white, gray, black) is one simulated phosphene; black and white dots represent phosphenes that were presented to the subject while gray ones were not presented and are shown only to indicate how pattern density decreases with eccentricity and that phosphenes spanned the entire visual field (see Section 2.5). During the Validation experiment, cues presented in the Free View period were shown in the clear, rather than through a phosphene field, but the task was otherwise identical.
Figure 3.
The combinations of optotype size F1–F5, Snellen 20/100 through 20/1600 (horizontal direction) and phosphene pattern P1–P4 (vertical direction) are shown in this collection of snapshots of the center part of the subject display during the Free View phase. Parameter values were selected such that it was impossible for subjects to identify letters with the lowest resolution phosphene pattern viewing the smallest optotypes (lower left subfigure) to provide a negative control, and quite easy with the highest resolution pattern viewing the largest optotypes (upper right subfigure) to provide a positive control. Snapshots shown are for gaze positions at the center of the monitor. Only the centralmost 10° of stimuli are shown; the remainder of the screen would be uniformly gray. Stimuli, animated through the gaze-contingent mechanism, were more readily identifiable than the static images above might suggest.
Table 1.
Primary experiment optotype sizes.
Table 2.
Phosphene patterns.
Table 3.
Cues and distractors.
Figure 4.
Testing using the Snellen chart (left) proceeds top to bottom, and left to right, with the subject calling out each letter on successive rows, or declaring their inability to do so. Testing started with the 20/50 line as shown here, which is the fourth line of standard charts. Testing using our letter recognition task (right) is performed in a balanced, randomly interleaved pattern; the figure, designed to show the equivalence between tests at given acuities, uses shuffle arrows to imply the interleaving. Trial conditions are represented as a cue letter followed by the two alternatives. Red and Green highlights illustrate single letter scoring (correct/incorrect) during an example data collection session where the subject was assessed with 20/18 on the Snellen task and 20/20 on the letter recognition task.
Table 4.
Validation experiment optotype sizes.
Figure 5.
Hit rate as function of font size, grouped by pattern resolution. For each condition, symbols indicate population mean and error bars, standard deviation. Each trace represents the fitted curves of subject hit rate for each pattern resolution (mean deviation below 1.2%). Font size is shown on the horizontal axis. Filled symbols are significantly above chance. The letters A, B, and C indicate the High-Performance, Mid-Range, and Chance/Low-Performance ranges, respectively (see main text). Subjects were not significantly above chance for all phosphene patterns for the smallest optotypes (F1, Snellen 20/100), with population hit rate increasing with both pattern resolution and optotype size.
Table 5.
Visual acuity values for each phosphene pattern.
Figure 6.
(LEFT) Response time as a function of font size, grouped by pattern resolution. Population mean (symbols) and standard deviation (error bars) of subject average response time for correct trials for each condition. Filled symbols indicate conditions with statistically significant population hit rates. Annotations of A and B symbols indicate conditions with hit rates in Mid-Range and High-Performance ranges, respectively, while unannotated symbols are conditions in the Chance/Low-Performance range. High-Performance conditions all have faster population response times than Mid-Range conditions, whereas Chance conditions have intermediate values that are affected by font size but not pattern resolution, together suggesting that two different perceptual mechanisms were involved. (RIGHT) Distribution of response time broken down by hit-rate performance ranges. The three graphs depict the relative occurrences for mean response times per condition binned to 100 ms for data from Chance/Low-Performance (upper), Mid-Range (middle), and High-Performance (lower) conditions. A Kolmogorov-Smirnov test found no significant difference between the Low-Performance and Mid-Range data (p = 0.47), while a significant difference was found between the Mid-Range and High-Performance data (p<0.01). There is a strong functional difference between performing at versus above chance, guessing versus knowing the answer, despite no significant difference found between the respective distributions. In contrast, there is only a weak functional difference between the upper two performance ranges, reflecting a sliding degree of task difficulty, but the significant difference in distributions suggests the emergence of a distinct mechanism at upper performance levels.
Figure 7.
Scatter plots of population hit rate and response time in the first versus last segment of trials. (LEFT) Population hit rates split by condition (stars) and overall mean (bulls eye). Data are above the line of equality as subjects perform more accurately in the last 200 trials (10 per condition) than in the first 200 trials. (CENTER) Population response time for each condition (stars) and overall mean (bulls eye). Nearly every datum is below the line of equality as subjects have faster responses in the last 200 trials than in the first 200. (RIGHT) Combining population response times and hit rates from the two previous graphs with linkages between first (open circles) and last (filled stars) segments reveals two distinct spans of behavior, one at low hit rates that is more disorganized, and one at higher hit rates that displays strong structure. The lower hit rates correspond to the Chance regime where subjects are more likely to be guessing; the higher hit rates correspond to the Mid-Range and High-Performance regimes. The largest improvements, as shown by the longest linkages, cluster around the threshold between Chance and Mid-Range regimes.
Figure 8.
Hit rate as function of font sizes (translated to the equivalent acuity) for the standard Snellen chart test (red unfilled squares) and the multiple-optotype two alternative forced choice (MO2AFC) letter recognition task (blue filled circles). Error bars are two-sigma confidence levels. Smooth traces are sigmoid fits to each data set. Uneven spacing across the horizontal axis reflects the step sizes between lines in the standard Snellen chart. Results from the letter recognition task have been normalized to span [0–1], an equivalent range as the Snellen chart test (0%–100%), for ease of comparison. The two curves are in high agreement, although there appears to be a trend to a shallower transition with the letter recognition task, possibly due to uncorrected false negatives at the upper end of the range.
Figure 9.
Bland-Altman plot of the differences in visual acuity scores between our letter recognition task and the Snellen chart test. The vertical axis displays the difference between measurements on both tasks, and the horizontal axis displays the mean acuity score value between the tasks. The solid line is the mean difference in LogMAR acuity (−0.02), and the area within the dotted lines represents the limits of agreement between LogMAR scores in the two visual acuity tasks. Each point represents the results from a single subject (blue circles), although some are overlapped (blue circles with black centers).