Impaired face identity discrimination in individuals with cerebral visual impairment: a pilot study

Introduction

Background

Cerebral visual impairment (CVI) is the leading cause of pediatric visual impairment (1) and is characterized by a broad range of visual deficits, including difficulties with higher-order visual processing. CVI is commonly associated with early brain injury or maldevelopment, such as hypoxic-ischemic encephalopathy, traumatic brain injury, neonatal hypoglycemia, seizure disorder, as well as genetic and metabolic disorders (2-4). Visual impairments in CVI are not explained by ocular disease or issues alone, but rather are due to damage or malformation of retrogeniculate pathways and higher-order visual processing areas (5). As a result, individuals with CVI often have trouble processing higher-level features of visual stimuli, even when visual acuity is within the normal range (6-8). For example, individuals with CVI often report difficulties recognizing people in large groups or at a distance, have difficulties discriminating faces and meaning conveyed by facial expression, and report experiencing discomfort or anxiety in crowds (9-12). Difficulties in recognizing dynamic facial expressions may be due to impaired motion processing abilities and have been explained in the context of dorsal stream dysfunction (4,13). Low vision, characterized by reduced visual acuity and contrast sensitivity, could also lessen a person’s ability to discriminate faces or meaning conveyed by facial expression (14).

This broad profile of visual impairments in CVI also has social consequences. The ability to recognize faces is important in everyday life [for review see (15,16)], and consequently, the inability to read facial expressions limits social interaction (8). Atypical social responses from a lack of expression understanding can also lead to social isolation from peer groups and can be mistaken for a lack of empathy or social interest (4,14), and those with difficulties identifying faces report struggling with forming and maintaining healthy social relationships (17,18).

Rationale and knowledge gap

Many tests exist to measure one’s ability to discriminate faces. Currently, the most widely utilized assessments are the Cambridge Face Memory Test (19) and the Cambridge Face Perception Test (20). These tests demonstrate an increase in reliability from previously invalidated tests [e.g., (21,22)] where stimuli contained clothing and hair (19,23,24). However, these measures do not provide a quantitative assessment of discrimination ability and they remain limited in their sensitivity to the severity of the deficit and their ability to detect gradual changes associated with the progression or improvement of prosopagnosia, such as in conditions such as autism (25-28), Turner’s syndrome (29), schizophrenia (30,31), Alzheimer’s disease (32,33) or Parkinson’s disease (34). While current face discrimination tests are sufficient for assessing one’s ability to recognize and distinguish new faces from old ones, there is currently no agreed-upon method for the quantification of this ability.

Objective

In this study, we used the Foraging Interactive D-prime (FInD) method for assessing face discrimination ability remotely. FInD has been utilized previously in measuring thresholds for color (35), contrast sensitivity (36), stereoacuity (37), motion (13,38), and face discrimination (39). With this test, we aimed to provide a reliable quantitative measure of impaired face discrimination ability in the CVI population. There are several features of this method that afford benefits over other tasks. Primarily, the faces for each trial are unique, so the test can be repeated without confounds of stimulus learning or familiarity. Furthermore, the test avoids memory effects and recognition of known individuals, and the sensitive design allows for the uncovering of group differences despite small sample sizes. Additionally, this task is simple, rapid, self-administered and could be completed remotely, so it could complement current methods that assess face perception.

Many adaptive algorithms can easily be configured to update stimulus test levels based on forced choice responses. However, in Yes/No applications, these approaches do not explicitly estimate d’ so are vulnerable to criterion and biases (40). FInD differs from these approaches by including a random number of Target and Null stimuli, so that d’ can be estimated based on the estimated probability of a hit/miss/false alarm/correct rejection response as a function of stimulus intensity, thereby correcting for bias and criteria artifacts. Unlike QUEST (41,42) and some other algorithms, FInD does not assume a fixed slope but estimates the slope directly, since it could provide additional information. In research that involves naive participants, there are several practical advantages of FInD over other single-trial algorithms, including the parallel presentation of a set of stimuli that span difficult (d’=0.1) and easy-to-see (d’=4.5) stimuli. This format reduces the frustration and memory burden on participants, that are a problem with standard paradigms in which trials are intentionally presented near threshold, which can be demotivating, and runs the risk that the participant forgets the pattern they are trying to detect. The continuous presentation of FInD and the option to change a response also avoids confounding errors based on inattention or unpreparedness for a trial, with a deficit in sensitivity.

We generated face stimuli using the Basel Face Model (43) which is a 3D morphable model comprised of 199 structural and 199 textural parameters that can be modified to generate unique faces. The database was created using 200 real adult faces where, through a principal component analysis (44-46), a model was constructed to define a relative “face space”. Within this space, a single point represents a face whose physical and textural attributes are uniquely defined in the space (47-49). In this way, faces near one another in the face space are structurally more similar. Parameter values are measured in terms of standard deviations along the dimensions of each parameter. The advantage of the Basel Face Model is that it allows for the manipulation of ground truth characteristics within faces (50-53) and for specificity in face discrimination testing that is unavailable with databases comprised of real-life faces. Disadvantages include the artificial appearance of computer-generated images and inaccurate representations of facial features or affect [for review, see (54)].

We aimed to measure face discrimination ability in individuals with CVI compared to controls. We administered the FInD Faces test (39) in controls and CVI participants using a remote desktop connection, facilitating testing for CVI participants, particularly those with mobility issues. Participants were shown a grid of nine face pairs (see Figure 1) that included a random subset (0.2 to 0.4) of pairs that were the same point in face space, and the remaining face pairs were different distances in face space. Participants were instructed to click on pairs where the two faces appeared to belong to different identities. Based on previous anecdotal reports and evidence from the literature (10-12), we hypothesized that individuals with CVI would have significantly higher threshold distances in face space for face discrimination compared to controls. We also aimed to test performance when faces were tilted along the yaw axis, removing the benefit of pointwise comparisons. The addition of tilt has been shown to reduce discrimination sensitivity in controls (39,55), and allows us to make an important distinction between performance in the two conditions. Specifically, if the CVI group performed similarly to controls on the forward-facing condition but showed diminished sensitivity in the tilt condition, this would suggest that the CVI group can make point-wise comparisons to identify faces but cannot form comprehensive, holistic representations of face identity. Contrarily, if CVI participants demonstrate reduced sensitivity in both conditions, it would suggest that they are not utilizing pointwise comparisons between faces. Ultimately, we hypothesized that CVI participants would perform similarly to controls when faces were forward-facing, as they could utilize pointwise featural comparisons between faces to discriminate them. However, we hypothesized that individuals with CVI would perform worse on tilted faces because head rotation changes the local shape properties of features more than their configuration and thus preserves holistic face processing.

Figure 1 Procedure. Example trial. (A) The hidden cell procedure is shown, such that faces are only shown in the cell hovered by the mouse cursor. (B) All faces on a given trial are shown without the hidden cells, as well as demonstrating the effect of the added tilt condition (uniform random deviation between −5° and +5°). An example of the higher end of the threshold range can be observed in the bottom left cell, while an example of the lower end of the threshold range can be observed in the middle-left cell.

Methods

Apparatus

The experiment was conducted remotely over Zoom (Zoom Video Communications Inc., San Jose, CA, USA). The program was run on the experimenter’s computer and viewed and controlled from the participant’s computer. Participants were given remote control of the experimenter’s computer and the experiment was run remotely from that computer, while being viewed on the participant’s own personal computer screen. Two desktops were used to run the experiment, the first had an AMD Ryzen 5 1600 6-core processor (Advanced Micro Devices, Inc., Santa Clara, CA, USA), NVIDIA GeForce GTX 1050 Ti graphics card (NVIDIA, Santa Clara, CA, USA), and 16GB of RAM with a 23" Dell monitor (Dell Technologies Inc., Round Rock, TX, USA) (60 Hz, 1,920×1,080 resolution). The second desktop had an Intel i5 processor (Intel Corporation, Santa Clara, CA, USA), NVIDIA GTX 1060 graphics card, and 32 GB of RAM with a 27" BenQ monitor (BenQ Corporation, Taipei) (144 Hz, 1,920×1,080 resolution). The two desktops were representative of two different testing sites, one for control subjects and one for CVI subjects. Control subjects were run at Northeastern University and CVI subjects were run at Massachusetts Eye and Ear Infirmary. This facilitated participant recruitment and satisfied Institutional Review Board (IRB) protections regarding the testing of clinical/vulnerable populations. The experiment was programmed, analyzed, and boxplots were created using MATLAB (The MathWorks, Inc., Natick, MA, USA) with Psychtoolbox (56) and the Statistics and Machine Learning toolbox.

Participants

Eight control participants (2 male) between the ages of 18 and 20 years [mean ± standard deviation (SD): 19.125±0.991] were recruited from Northeastern University’s undergraduate population and received course credit for their participation. Based on a preliminary study showing high variability within controls, the following eligibility parameters were used: participants must have normal or corrected to normal vision, be fluent in English, not have a diagnosis related to a language disorder, and not have a diagnosis of Autism Spectrum Disorder. We balanced the control pool with the CVI pool for statistical testing.

We recruited 10 CVI participants in total. However, two CVI participants were unavailable to return to complete the tilt version of the experiment and thus were excluded from analysis to allow for a pairwise statistical comparison. The remaining 8 CVI participants (3 male) between the ages of 19 and 27 (mean ± SD: 23.62±3.54) years were recruited from the Laboratory for Visual Neuroplasticity (Massachusetts Eye and Ear) and received monetary compensation for their participation. Prior to their enrollment in the study, CVI participants were previously diagnosed by specialized and highly experienced eye care professionals. All CVI participants had visual impairments related to pre- or perinatal neurological injury and/or neurodevelopmental disorders. Diagnosis was based on a battery of assessments, including visual function (visual acuity, visual field perimetry, contrast, color, and ocular motor function), a thorough refractive and ocular examination, and review of medical history and neuroimaging records (2,57).

Information regarding functional visual behaviors was collected from appropriate questionnaires and inventories (58,59) to help formalize the diagnosis (60). Best corrected visual acuities in the better-seeing eye ranged from 20/15 to 20/70 Snellen (−0.07 to 0.54 logMAR equivalent) and all had intact visual field function within the area corresponding to the visual stimulus presentation. Visual acuity was measured in clinic by a trained optometrist or ophthalmologist following standard procedures for Early Treatment Diabetic Retinopathy Study (ETDRS) charts. We reviewed the ophthalmic history, visual acuity, and visual perimetry testing results of all the CVI participants. If there was evidence of a visual field deficit (e.g., inferior), it was in the far periphery and not overlapping with the area of viewing/testing.

All CVI participants completed the Manual Ability Classification System (MACS) questionnaire to assess how an individual uses their hands to handle objects in daily activities. Previous research has shown the MACS to have strong validity and reliability, as well as agreement between ratings from parents and occupational therapists (61). All CVI participants completed MACS score (regardless of whether or not they had a diagnosis of cerebral palsy). Of our 8 CVI participants, 6 rated their MACS as level 1, indicating the participant “handles objects easily and successfully. At most, limitations in the ease of performing manual tasks require speed and accuracy. However, any limitations in manual abilities do not restrict independence in daily activities restrict independence in daily activities”. The remaining two 2 CVI participants rated their MACS as level 2, indicating they “handle most objects but with somewhat reduced quality and/or speed of achievement. Certain activities may be avoided or be achieved with some difficulty; alternative ways of performance might be used but manual abilities do not usually restrict independence in daily activities”. All participants were comfortable using a computer mouse, as needed for this task (based on self-report) and without accessibility modifications. Demographic and clinical information regarding CVI participants can be found in Table 1.

In the CVI participants, Verbal IQ was assessed using subtests from the Wechsler Intelligence Scale for Children (WISC IV) and Wechsler Adult Intelligence Scale (WAIS IV), 4th Edition (specifically, the Digit Span, Similarities, and Vocabulary subtests of WISC IV and the Digit Span, Similarities, Vocabulary, and Information subtests of WAIS IV to obtain an index of verbal comprehension). Verbal IQ scores ranged from 66 to 148 (mean ± SD: 107.00±29.98). None of our participants required testing accommodations except for one, who required additional practice trials to ensure they understood the nature of the task.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional review boards of Northeastern University and the Massachusetts Eye and Ear in Boston, MA, USA (IRB #: 14-09-16 – Psychophysical Study of Visual Perception and Eye Movement Control). Written informed consent was obtained from all control and CVI participants (and/or a patent/legal guardian in the case of a minor) prior to commencing the study.

All participants were told they must be able to perform the experiment on a laptop or desktop with an adequate screen size (approximately equal or larger than a standard iPad) in a quiet, well-lit area. Adequate screen size and room conditions were verbally confirmed by the participant before initiation of the experiment. Viewing distance was selected by participants for their own level of comfort, approximately 64 cm based on for participants in this study (62). However, viewing distance was not enforced as several studies have demonstrated that face processing is robust over a large range of viewing distances/ retinal image sizes. For example, familiar face identification has been found to be invariant over an 8-fold difference in image size (2–16°) (63), and with reduced high-frequency detail (64).

Stimuli

Faces were generated using the Basel Face Model (BFM) (43) which categorizes faces across 199 unique structural and textural parameters. In a previous study (39), we determined that changes along the first 10 structural parameters led to the lowest thresholds for differentiating between faces, meaning that participants were most sensitive to changes of these parameters. Threshold is estimated from the fit of d’ function as a function of signal intensity, where d’=1. For precision and efficiency in our threshold estimation, we opted to change one structural parameter on a given trial, and only tested the 10 most sensitive parameters instead of all 199 parameters (as in our preliminary study). A representative example of the features controlled by these 10 parameters can be found in the supplementary appendix online (Figure S1). Each face pair was generated with random values [a Gaussian deviate from a distribution with mean parameter µ=0 and standard deviation parameter σ=1, using MATLAB’s randn() function] for all 199 BFM parameters. The 199 dimensions of the Basel Face database are all scaled in z score units, with a mean of zero for each dimension, which corrects any range differences across eigenvalues, and are representative of the variation in faces along a given parameter. A threshold value determined by the FInD algorithm (as described in Procedure below) was applied to the BFM parameter of interest. For example, if testing BFM parameter 3, 199 random values (normally distributed random deviates where µ=0, σ=1.0) were assigned across the parameters (creating a randomly generated face), and then BFM parameter 3 would be altered to generate two faces that differed only by a specified distance along BFM parameter 3. In this way, the faces within each cell appeared relatively similar, while the faces between cells appeared fairly different (Figure 1).

Procedure

We modified the paradigm described in Walter & Bex (39) by integrating a hidden cell feature in which only the face pair hovered by the mouse cursor was presented. This was implemented to avoid possible crowding/simultanagnosia effects [see (13)]. This reduced clutter made the task more suitable for the CVI population. In order to determine discrimination thresholds, we used the FInD paradigm (38,65,66) which can be used to estimate discrimination thresholds for a variety of psychophysical metrics such as color (35), motion (13,38), contrast sensitivity (36), stereoacuity (37), and faces (39). FInD is self-administered and can be performed remotely, making it convenient for clinical populations such as CVI, where access barriers can prevent participation in in-lab studies. When measuring test-retest reliability, we found that the differences between groups are greater than test-retest differences and are smaller than differences in online vs. lab experiments (39). The purpose of the FInD paradigm is to center presented stimuli around a participant’s individualized threshold for discriminating differences, making it highly adaptable to a variety of skill levels. This allows for large differences in skillsets (and thus adjusted difficulties) between subjects, but because the results are a measure of thresholds, it is still possible to make between group comparisons.

In this variation of the FInD paradigm, stimuli are presented across four charts each composed of a 3×3 grid of cells. Each cell subtended 6° and was surrounded by a black border, with 3° of spacing between cells. Each cell contained two faces, and a random subset (between 0.6 and 0.8) contained two different faces (signal stimuli) and the remaining contained identical faces (null stimuli). The signal level across cells was adaptively controlled by FInD to span a range of face pairs that are easy to discriminate (d’=4.5) to very difficult to discriminate (d’=0.1).

Participants were instructed to hover the mouse cursor over cells to reveal the faces, and to click on any cells that they determined to have different face identities. Once a cell was clicked, its border was highlighted green and participants could click again to unselect it, in case they changed their mind and decided that the faces were the same person. Participants were told to click as many or as few cells as they thought corresponded to different identities with unlimited time, and to click the green “next” button to move on to the next series. The responses were then classified as hits, misses, correct rejections, or false alarms to calculate d’ as a function of signal level. The probability of a “Yes” response as a function of test level was then calculated as:

p(Yes)=1−Φ(Φ−1(1−F)−d'max×(Sθ)γ((d'max)2−1)+(Sθ)2γ)[1]

where p(Yes) is the probability of a Yes response, ϕ is the normal cumulative distribution function, F is the false alarm rate, S is the test level for the target parameter, θ is threshold level for the target parameter, d’_max is the saturating value of d’ which was fixed at 5, and γ is the slope. The resulting d’ values and slope were used adaptively to select the signal levels spanning d’=0.1 to d’=4.5 for the subsequent charts. Thus, the range centered on each individual’s threshold estimate with successive charts. Half of the test level was subtracted from the target component of one face (randomly assigned to the left or right), while the other face had half of the test level added to its target component. For null face pairs, the test level was set to 0, meaning all 199 components (including the target component) were identical across both faces. There were 4 charts per BFM parameter, and 10 BFM parameters tested, resulting in 40 total trials. BFM parameters were tested in order from 1 through 10 where after every 10 trials, the subsequent chart for each BFM parameter was tested (e.g., trial 11 was chart 2 for BFM parameter 1).

Participants completed two conditions of 4 charts for each of 10 BFM parameters. The BFM parameters were tested in turn, and cycled over 4 repeats for each condition to ensure variation. In the first condition, faces were forward-facing (yaw = 0°) and in the second condition, faces were tilted by a uniform random deviation ranging between −5° and +5° [using MATLAB’s rand() function]. A value of 5° was chosen to add slight head rotation that would allow face identification but prevent point-wise comparison of face properties. The selection of this parameter was determined a priori during pilot testing and may be investigated in future studies, however 5° was ultimately chosen subjectively because it was enough to avoid point-wise comparison but not so large to create dramatic visual differences in appearance. Control participants completed both conditions in the same session. However, to minimize possible effects related to fatigue, CVI participants completed two individual sessions.

Viewing distance was not strictly controlled, and while accommodation was not formally tested, at this viewing distance and the subtended angle of the faces, accommodation was not expected to be a factor.

Statistical procedure

Statistical analyses were conducted in MATLAB. Independent samples t-tests were used to complete planned comparisons between thresholds for control and CVI groups for both forward-facing and tilted conditions. Paired t-tests were used for within-subject comparisons of tilt effects (e.g., forward vs. tilted faces) within each group. We performed both parametric [t-test and univariate analysis of variance (ANOVA)] and non-parametric tests (Wilcoxon rank sum), which had corresponding results. The majority of results had clearly significant or non-significant outcomes, and only one P value was marginal (P=0.048), and so uncorrected P values are reported alongside effect sizes to allow for interpretation. A formal power analysis was not performed, however observed effect sizes are reported to aid in interpreting the reliability and magnitude of these findings. Although formal test-retest reliability was not assessed in the present study, Bland Altman analyses in our previous study showed no significant differences between tests (39). Thresholds were estimated from all responses across multiple charts per condition (up to four), improving stability and precision of individual estimates.

Results

All control participants completed 4 charts for all 10 BFM parameters. Two CVI participants completed only 3 charts for all 10 BFM parameters due to fatigue. A threshold value can be calculated from as little as 1 chart, however, precision increases with the number of trials obtained. We also removed any individual threshold datapoints defined as being greater than 3 standard deviations outside the mean. This resulted in 2 CVI thresholds being removed for the head tilt condition. In other words: out of 160 total recorded thresholds (2 groups ×8 subjects per group ×10 thresholds per subject) we included 158 datapoints (98.75% of data).

Reaction times

Reaction time was measured as the time between the start of a trial and when the participant clicked “next”. Values are reported as mean ± SD. Controls completed charts with an average time of 33.637±11.906 s for forward-facing faces and 22.874±5.866 s for tilted faces. CVI participants completed charts in an average time of 33.464±9.156 s for forward-facing faces and 25.391±9.794 seconds for tilted faces. A univariate analysis of variance found that there was no significant difference in time between controls and CVI [F(3,1)=0.124; P=0.73, η_p²=0.004], but there was a significant effect of tilt [F(3,1)=7.975; P=0.009, η_p²=0.222] and there was no significant interaction [F(3,1)=0.163; P=0.69, η_p²=0.006].

Parametric tests

For illustrative purposes, Figure 2 shows random faces rendered with a threshold difference for the BFM parameter #5 for forward-facing and tilted faces for control and CVI participants. The average threshold for controls was 3.125±1.289 (units are BFM parameter standard deviation values) for forward-facing faces (blue data, Figure 3) and 4.652±2.329 for tilted faces (blue data, Figure 4). The average threshold for CVI was 6.851±4.301 for forward-facing faces (red data, Figure 3) and 7.880±6.113 for tilted faces (red data, Figure 4). For forward-facing faces, CVI participants had significantly higher thresholds than controls [t(14)=−3.439, P=0.004, d=01.719] (Figure 3). For tilted faces, CVI had significantly higher thresholds than controls [t(14)=−2.163, P<0.05, d=0.964] (Figure 4). For controls, thresholds for tilted faces were significantly higher than thresholds for forward-facing faces [t(7)=−3.853, P=0.006, d=1.574]. For CVI, thresholds for tilted faces were not significantly different than thresholds for forward-facing faces [t(7)=−1.355, P=0.22, d=0.466]. A univariate analysis of variance shows a significant effect of group, where CVI participants had higher average thresholds than the control group [F(3,1)=14.218, P=0.001, η_p²=0.337], but no significant effect of tilt [F(3,1)=1.920, P=0.18, η_p²=0.064], and no significant interaction [F(3,1)=0.073, P=0.79, η_p²=0.003].

Figure 2 Representative example faces transformed using average thresholds for each group and for the two tasks. The two faces within each box are separated across the Basal Face Model parameter 5 by the average threshold for each group. Control forward-facing =3.125, control tilt =4.652, CVI forward-facing =6.851, CVI tilt =7.880. Each pair of faces thus illustrates the threshold at which participants identified the two faces as belonging to different identities. CVI, cerebral visual impairment.

Figure 3 Face identity discrimination thresholds for forward-facing faces. The x-axis shows the Basal Face Model parameter that differed between members of a face pair (the remaining 198 Basal Face Model parameters shared the same value in each face pair). The y-axis shows the difference in standard deviation units that were required for the two faces to appear to have different identities. Blue data points show results for control participants; red data points show the results for CVI participants. Bars represent interquartile range, whiskers represent minima and maxima, circles with black dots represent medians, unfilled circles represent outliers, and dotted lines represent group means. CVI, cerebral visual impairment.

Figure 4 Face identity discrimination thresholds for tilted faces. Axes and box and whisker notations are identical to previous Figure 3. The only methodological difference is the data is now representative of tilted faces. Each individual member of the face pair was tilted on the yaw axis by a random value between ±5°. CVI, cerebral visual impairment.

Non-parametric tests

Control forward-facing, control tilted, and CVI forward-facing all passed Shapiro-Wilk’s test for normality, but CVI tilted did not. Additionally, control forward-facing and control tilted passed Levene’s test for equality of variances, but CVI forward-facing and CVI tilted did not. Due to this, we also ran non-parametric tests, specifically Wilcoxon rank sum tests. All of our non-parametric test results mirrored our parametric t-test results. Control forward-facing thresholds (median =3.046) were significantly lower than CVI forward-facing thresholds (median =5.918), where U=38 and P<0.001. Control tilted thresholds (median =4.770) were significantly lower than CVI tilted thresholds (median =6.233), where U=51 and P=0.08. Control forward-facing thresholds (median =3.046) were significantly lower than control tilted thresholds (median 4.770), where U=44 and P=0.01. CVI forward-facing thresholds (median =5.918) were not significantly lower than CVI tilted thresholds (median 6.233, where U=60 and P=0.44).

Interactions

We performed one-way ANOVA’s to test the interactions between individual parameters, but did not note any particularly interesting results. Specifically, in the control forward-facing condition, parameter 4 had lower thresholds than parameters 7, 8, and 10. For the control tilt condition, there were many significant interactions, mainly that parameter 4 was lower than many parameters and 7 and 10 were higher than many parameters. For the CVI forward-facing and CVI tilt conditions, there were no significant interactions.

Face discrimination ability and acuity

We wanted to examine whether the observed impairment of the CVI group in the face identity discrimination task was directly related to differences in visual acuity. We measured the correlation between visual acuity in CVI participants and performance on the face task, across both tilt and no tilt conditions (Figure 5). There was no significant correlation between acuity and identity discrimination thresholds for forward-facing (r=0.040, R²=0.002, P=0.93) or tilted faces (r=−0.100, R²=0.010, P=0.81).

Figure 5 Correlation between visual acuity and face identity discrimination thresholds for CVI participants. Each data point shows the mean face identity discrimination threshold for one CVI participant, orange data show the results for forward-facing faces (Pearson’s r=0.040), and green data show the results for tilted faces (Pearson’s r=−0.100). Acuity is measured in Snellen equivalents in feet. CVI, cerebral visual impairment.

Nevertheless, because the majority of our CVI sample had 20/20 acuity or better, we ran a secondary in-lab test to examine the effect of dioptric blur on visual acuity and face identity discrimination. Twenty-three novel control participants were recruited from Northeastern University’s undergraduate population under the same inclusion criteria as the main experiment. Digital blur was created with a disk filter using MATLAB’s fspecial() function with kernel widths of 0°, 0.1°, 0.2°, 0.4°, and 0.8°. Face stimuli were generated as in the main experiment, then convolved with the blur filter using MATLAB’s imfilter() function. Examples of face pairs blurred with different kernel widths are shown in Figure 6. The task was the same as with the main experiment, except that instead of measuring thresholds for each of the 10 BFM parameters, we measured threshold applied to all 199 parameters simultaneously. Thus, for each difference level assigned by the FInD algorithm, half the difference level was added or subtracted (at random) from the BFM parameters of one face of each pair and half the difference was added or subtracted (respectively to the pair) to the same parameters of the second face. Viewing distance was 60 cm.

Figure 6 Examples of the effects of blur width (°) on visual acuity (top) and face identity discrimination (bottom) tasks. Face discrimination performance remained possible at blur levels that rendered visual acuity tasks impossible.

The acuity test was performed on a calibrated high-resolution display under laboratory conditions that meet American National Standards Institute (ANSI) testing standards. Acuity was measured with an Angular Indication Measurement (AIM) acuity task (67), in which participants reported the orientation of a 4×4 grid of Landolt C targets whose size was adaptively updated after each chart (see Figure 6). Participants were instructed to click along the perimeter of a white circular border to indicate the direction of the gap in the C. The Landolt C optotypes were digitally filtered with the same process as the faces, and examples of blurred optotypes are shown in Figure 6. The size of the Landolt Cs on each of 3 charts spanned an adaptive range calculated from the error in orientation report as a function of optotype size.

In order to compare units, face thresholds and acuity thresholds were normalized using min-max normalization methods. Figure 7 shows the effect of digital blur on visual acuity and face discrimination thresholds. Informal inspection reveals that the effect of blur on visual acuity is much greater than for face identity discrimination. The reader can appreciate this difference by comparing subjective performance in the stimuli shown Figure 6, which illustrates that face identity discrimination is possible at blur levels that render letter orientation identification impossible. In order to compare the effects of blur on face recognition and letter acuity statistically, we calculated the slopes of the least-squared fit to each participant’s acuity and face thresholds as a function of blur kernel width and compared the results using a paired t-test. We found that slopes of acuity and face thresholds over increasing blur kernel widths were significantly different [t(22)=−11.291; P<0.001, d=4.152] (Figure 7). The lack of correlation between visual acuity and face discrimination performance in CVI (Figure 5) and the differential effects of blur on acuity and face discrimination suggest that the difference in performance between control and CVI participants cannot simply be attributed to differences in visual acuity.

Figure 7 The effect of digital blur on visual acuity (left) and face identity discrimination (right). The x-axis shows the width or the blur disk kernel, and the y-axis shows normalized discrimination thresholds for visual acuity and face discrimination. Bars represent interquartile range, whiskers represent minima and maxima, circles with black dots represent medians, unfilled circles represent outliers, and connecting lines represent means.

Discussion

As part of their complex visual profile, individuals with CVI often report difficulty recognizing familiar faces (9-12,68). In order to quantify face discrimination ability in this population, we investigated individuals with CVI compared to controls by using a morphable face model in which differences between faces can be systematically controlled. By indicating which face pairs participants recognized as belonging to different identities, we could determine thresholds for face identity discrimination across individuals. Participants viewed face pairs that were forward-facing followed by face pairs that were tilted on the yaw axis by a random deviation between −5° and +5°. Differences between forward faces could, in principle, be based on point-wise comparison between the images without necessarily integrating features into a face, whereas the introduction of tilt prevented that strategy and requires comparison of a more holistic representation of faces [see (69) for a review of these principles].

Key findings

Individuals with CVI exhibited significantly higher face discrimination thresholds compared to controls, consistent with real-world difficulties in face recognition. Notably, CVI participants showed no significant difference in discrimination thresholds between forward-facing and tilted faces, unlike controls, who performed worse with tilted faces. The time to complete the task did not significantly differ between CVI and control participants, suggesting similar levels of cognitive engagement. However, both groups took longer to complete the forward-facing condition, potentially due to the availability of both holistic and pointwise featural comparison strategies in this condition. A follow-up study confirmed that face discrimination ability is resilient to visual acuity differences, ruling out reduced acuity as the primary explanation for deficits observed in CVI participants.

Strengths and limitations

A key strength of this study is the use of a morphable face model, allowing precise, controlled manipulation of face identity differences that is not possible with photographs of real faces. Additionally, the FInD study design accounts for ability-matched task engagement and cognitive load through an adaptive algorithm that ensures comparable challenge levels across both control and CVI participants. A limitation of this study is the fixed task order, where the tilt condition was always presented after the forward-facing condition, potentially influencing results due to task familiarity.

Comparison with similar research

Prior research has documented face processing deficits in CVI (9). The current study aligns with findings that individuals with CVI struggle with face recognition and extends that finding to rule out reduced visual acuity alone. Unlike studies that primarily assess face perception using photographs of real faces, this study uses a morphable face model, providing a controlled paradigm to systematically measure face discrimination thresholds. Furthermore, prior research has indicated that holistic face processing is crucial for recognition (69); our results suggest that CVI participants may rely more heavily on holistic strategies, though additional research is necessary. Similarly, there is extensive research on featural versus holistic processing in developmental prosopagnosia. Overall, the majority of research has found that individuals with developmental prosopagnosia have impaired (but not absent) holistic processing, specifically for faces (70-74). Additionally, Bauer et al. found that face recognition deficits were prevalent in CVI, as shown through self (or caregiver)-report questionnaire items and a facial recognition task that used blurred images of real faces to assess participants’ ability to recognize unfamiliar faces. Compared to controls, CVI participants had longer response times on the task and lower accuracy. These authors also stated that it still remains unclear whether individuals with CVI have a selective impairment in holistic vs. featural face recognition. However, their results raise the possibility that individuals with CVI may have impairments in holistic face processing, especially under conditions that obscure fine detail (9).

Explanations of findings

We found that there was no significant difference in the time it took CVI and control participants to complete the face discrimination task. The adaptive algorithm employed in this study created personalized charts around the threshold performance level for each participant, so all participants should be similarly challenged by the identity discrimination task. The observation that the task took a similar amount of time to complete suggests that CVI and control participants were similarly cognitively engaged in the task. However, for both groups, there was a significant effect of tilt on time to complete the task, where overall, it took longer to complete the forward-facing condition compared to the tilt condition. In the forward-facing condition, there are two sources of information, holistic and pointwise comparisons between faces. Whereas in the tilt condition, only holistic information is available. The difference with respect to time in completing the tasks could suggest that participants may use additional time to process pointwise information, and this may lead to lower face thresholds, at least for the control group. However, participants completed the tilted condition after the forward-facing condition, so they could have benefited from task learning.

CVI participants had significantly higher thresholds for both forward-facing and tilted faces compared to controls. This is consistent with the notion that individuals with CVI have more difficulty discriminating faces in the real world (9-12,14,68). It should be noted that at very high parameter values, the faces may appear disfigured or distorted (see, for example, the lower left face in Figure 1B) and may be required for some participants to discriminate between faces 0 although note that threshold level faces (see Figure 2 for examples) were not extreme. Interestingly, within groups, control participants had significantly higher thresholds for tilted faces compared to forward-facing, but CVI participants did not. A possible explanation for this difference between groups could be that control participants were able to utilize pointwise comparisons in the forward-facing condition, but not in the tilt condition, resulting in more precise performance when faces were presented at the same angle. CVI participants, however, may not have utilized pointwise comparisons, as their performance in forward-facing and tilt conditions was comparable, even though they also took longer to complete the forward-facing trials. A possible explanation could be that CVI individuals were only utilizing holistic representations of faces, and not using specific features to make judgments during face processing. However, if this is indeed the case, it is interesting that the times for CVI were slower for the forward-facing condition. One can then ask, if they are not utilizing pointwise comparisons, why are they taking longer to make these judgments? Another explanation, and perhaps representing a limitation of the study, could be that because the tilt condition was always presented after the forward-facing condition, participants had a better understanding of the task and could complete it more quickly. Overall, face processing deficits in CVI may involve both featural and configural processing, but more research is needed to disentangle the mechanisms of holistic processing in this population.

Given that it has been previously documented that low vision (i.e., reduced visual acuity) can impact face discrimination ability (14,70), we ran a follow up study to examine whether the higher face identity discrimination thresholds in CVI participants could be explained by reduced visual acuity. In our CVI group, there was no correlation between visual acuity and face discrimination thresholds (Figure 5), suggesting that visual acuity deficits alone do not explain our observed results. We did not measure accommodation in our participants, but the lack of correlation between visual acuity and face discrimination thresholds suggests that face processing impairments are not related to accommodation difficulties. To further confirm this suspicion, we also ran a second experiment with another cohort of control participants to measure how visual impairment simulated by visual blur could affect visual acuity and face discrimination performance. We found that while visual acuity decreased rapidly with increasing blur, face discrimination ability remained fairly stable and gradually decreased after kernel width exceeded ~0.2°, whereas visual acuity had decreased 7-fold at this same level of blur (Figure 7). These findings demonstrate that visual acuity and face discrimination are not directly correlated and therefore face processing deficits in the CVI participant cannot be attributed to differences in visual acuity. An explanation for the resilience of face identification discrimination in the presence of blur could be that the face task relies more on holistic features and/or that face identity discrimination depends on features at a lower spatial frequency scale than visual acuity. While visual acuity depends on the fine detail of the Landolt C’s gap, face identification depends on features at lower spatial frequencies (64,71-74). These critical features are differentially affected by blur or other spatial transformations (63,75). Consequently, face identities are defined by features (like overall head shape) at large spatial scales that can be distinguished without fine details. It is important to note that it is possible that crowding (76) or other spatial distortion effects could be at play, our blur experiment simply rules out blur as a contributing factor.

Implications and actions needed

These findings highlight the need for tailored interventions to support face recognition in individuals with CVI. Given the potential reliance on holistic processing, training programs could focus on enhancing feature integration or utilizing alternative recognition strategies. The results also underscore the importance of using specialized face processing assessments in CVI evaluations, as standard visual acuity tests fail to capture these specific deficits. Future research could explore neuroimaging approaches to investigate the neural mechanisms underlying face recognition in CVI and develop targeted rehabilitation programs to improve social and communication skills in affected individuals. Ultimately, the FInD procedure could be a realistic option for face processing assessment in groups such as CVI.

Conclusions

Overall, we found that CVI participants had higher thresholds for face identity discrimination compared to controls, and that CVI participants performed similarly in forward-facing and tilt conditions, unlike controls. We hypothesize that CVI individuals may not be utilizing pointwise comparisons when discriminating between faces, and perhaps rely more on holistic information at lower spatial scales. Additionally, we can conclude that this reduced performance is not due to reduced acuity in the CVI group, but more likely related to impaired higher level cognitive processing.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://atm.amegroups.com/article/view/10.21037/atm-25-53/dss

Peer Review File: Available at https://atm.amegroups.com/article/view/10.21037/atm-25-53/prf

Funding: This study was supported by the National Institutes of Health (R01EY029713 and R01EY030973).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://atm.amegroups.com/article/view/10.21037/atm-25-53/coif). The FInD method is protected by a provisional patent that is owned by Northeastern University and licensed to PerZeption Inc., of which author P.B. is a cofounder. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional review boards of Northeastern University and Massachusetts Eye and Ear in Boston, MA, USA (IRB #: 14-09-16 – Psychophysical Study of Visual Perception and Eye Movement Control) and informed consent was obtained from all individual participants.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Solebo AL, Teoh L, Rahi J. Epidemiology of blindness in children. Arch Dis Child 2017;102:853-7. [Crossref] [PubMed]
2. Fazzi E, Signorini SG, Bova SM, et al. Spectrum of visual disorders in children with cerebral visual impairment. J Child Neurol 2007;22:294-301. [Crossref] [PubMed]
3. Hoyt CS. Visual function in the brain-damaged child. Eye (Lond) 2003;17:369-84. [Crossref] [PubMed]
4. Philip SS, Dutton GN. Identifying and characterising cerebral visual impairment in children: a review. Clin Exp Optom 2014;97:196-208. [Crossref] [PubMed]
5. Sakki HEA, Dale NJ, Sargent J, et al. Is there consensus in defining childhood cerebral visual impairment? A systematic review of terminology and definitions. Br J Ophthalmol 2018;102:424-32. [Crossref] [PubMed]
6. Boot FH, Pel JJ, van der Steen J, et al. Cerebral Visual Impairment: which perceptive visual dysfunctions can be expected in children with brain damage? A systematic review. Res Dev Disabil 2010;31:1149-59. [Crossref] [PubMed]
7. Dutton GN. The spectrum of cerebral visual impairment as a sequel to premature birth: an overview. Doc Ophthalmol 2013;127:69-78. [Crossref] [PubMed]
8. McKillop E, Bennett D, McDaid G, et al. Problems experienced by children with cognitive visual dysfunction due to cerebral visual impairment – and the approaches which parents have adopted to deal with these problems. Br J Vis Impair 2006;24:121-7.
9. Bauer CM, Manley CE, Ravenscroft J, et al. Deficits in Face Recognition and Consequent Quality-of-Life Factors in Individuals with Cerebral Visual Impairment. Vision (Basel) 2023;7:9. [Crossref] [PubMed]
10. Houliston MJ, Taguri AH, Dutton GN, et al. Evidence of cognitive visual problems in children with hydrocephalus: a structured clinical history-taking strategy. Dev Med Child Neurol 1999;41:298-306. [Crossref] [PubMed]
11. Morse MT. Should Individuals who do Not Fit the Definition of Visual Impairment be Excluded from Visual Impairment Services? J Vis Impair Blind 2017;111:377-81.
12. Stiers P, Swillen A, De Smedt B, et al. Atypical neuropsychological profile in a boy with 22q11.2 Deletion Syndrome. Child Neuropsychol 2005;11:87-108. [Crossref] [PubMed]
13. Merabet LB, Manley CE, Pamir Z, et al. Motion and form coherence processing in individuals with cerebral visual impairment. Dev Med Child Neurol 2023;65:1379-86. [Crossref] [PubMed]
14. Zihl J, Dutton GN. Cerebral Visual Impairment in Children: Visuoperceptive and Visuocognitive Disorders. Vienna: Springer Vienna; 2015 [cited 2022 Jul 12].
15. Bate S. Face Recognition and its Disorders. Bloomsbury: Bloomsbury Publishing; 2017:252.
16. Oruc I, Balas B, Landy MS. Face perception: A brief journey through recent discoveries and current directions. Vision Res 2019;157:1-9. [Crossref] [PubMed]
17. Barton JJS, Corrow SL. The problem of being bad at faces. Neuropsychologia 2016;89:119-24. [Crossref] [PubMed]
18. Yardley L, McDermott L, Pisarski S, et al. Psychosocial consequences of developmental prosopagnosia: a problem of recognition. J Psychosom Res 2008;65:445-51. [Crossref] [PubMed]
19. Duchaine B, Nakayama K. The Cambridge Face Memory Test: results for neurologically intact individuals and an investigation of its validity using inverted face stimuli and prosopagnosic participants. Neuropsychologia 2006;44:576-85. [Crossref] [PubMed]
20. Duchaine B, Germine L, Nakayama K. Family resemblance: ten family members with prosopagnosia and within-class object agnosia. Cogn Neuropsychol 2007;24:419-30. [Crossref] [PubMed]
21. Benton AL, Van Allen MW. Impairment in facial recognition in patients with cerebral disease. Trans Am Neurol Assoc 1968;93:38-42.
22. Warrington EK. Recognition Memory Test. Windsor: NFER. Nelson Publishing; 1984.
23. Bowles DC, McKone E, Dawel A, et al. Diagnosing prosopagnosia: effects of ageing, sex, and participant-stimulus ethnic match on the Cambridge Face Memory Test and Cambridge Face Perception Test. Cogn Neuropsychol 2009;26:423-55. [Crossref] [PubMed]
24. Duchaine BC, Nakayama K. Developmental prosopagnosia and the Benton Facial Recognition Test. Neurology 2004;62:1219-20. [Crossref] [PubMed]
25. Barton JJ, Cherkasova MV, Hefter R, et al. Are patients with social developmental disorders prosopagnosic? Perceptual heterogeneity in the Asperger and socio-emotional processing disorders. Brain 2004;127:1706-16. [Crossref] [PubMed]
26. Duchaine B, Murray H, Turner M, et al. Normal social cognition in developmental prosopagnosia. Cogn Neuropsychol 2009;26:620-34. [Crossref] [PubMed]
27. Weigelt S, Koldewyn K, Kanwisher N. Face identity recognition in autism spectrum disorders: a review of behavioral studies. Neurosci Biobehav Rev 2012;36:1060-84. [Crossref] [PubMed]
28. Wilson CE, Palermo R, Schmalzl L, et al. Specificity of impaired facial identity recognition in children with suspected developmental prosopagnosia. Cogn Neuropsychol 2010;27:30-45. [Crossref] [PubMed]
29. Hong D, Scaletta Kent J, Kesler S. Cognitive profile of Turner syndrome. Dev Disabil Res Rev 2009;15:270-8. [Crossref] [PubMed]
30. Archer J, Hay DC, Young AW. Face processing in psychiatric conditions. Br J Clin Psychol 1992;31:45-61. [Crossref] [PubMed]
31. Frith CD, Stevens M, Johnstone EC, et al. Integration of schematic faces and other complex objects in schizophrenia. J Nerv Ment Dis 1983;171:34-9. [Crossref] [PubMed]
32. Bäckman L, Herlitz A. The relationship between prior knowledge and face recognition memory in normal aging and Alzheimer's disease. J Gerontol 1990;45:94-100. [Crossref] [PubMed]
33. Hodges JR, Salmon DP, Butters N. Recognition and naming of famous faces in Alzheimer's disease: a cognitive analysis. Neuropsychologia 1993;31:775-88. [Crossref] [PubMed]
34. Dewick HC, Hanley JR, Davies AD, et al. Perception and memory for faces in Parkinson's disease. Neuropsychologia 1991;29:785-802. [Crossref] [PubMed]
35. He J, Skerswetat J, Ross NC, et al. Effect of EnChroma glasses on HRR and FInD Color discrimination task in anomalous trichromats. J Vis 2022;22:4166.
36. Liu J, Bex P, Jnawali A, et al. Evaluation of the novel Foraging Interactive D-prime (FInD) CSF Test. Invest Ophthalmol Vis Sci 2021;62:525.
37. Neupane S, Bex PJ, Skerswetat J. Comparison of FInD (Foraging Interactive D-prime)-Depth with three conventional methods. J Vis 2022;22:4351.
38. Bex P, Skerswetat J. FInD - Foraging Interactive D-prime, a rapid and easy general method for visual function measurement. J Vis 2021;21:2817.
39. Walter K, Bex P. A novel, rapid, quantitative method for face discrimination. PLoS One 2024;19:e0315998. [Crossref] [PubMed]
40. Green DM, Swets JA. Signal detection theory and psychophysics. Oxford, England: Robert E. Krieger; 1974. xiii, 479 p. (Signal detection theory and psychophysics.).
41. Watson AB, Pelli DG. QUEST: a Bayesian adaptive psychometric method. Percept Psychophys 1983;33:113-20. [Crossref] [PubMed]
42. Watson AB. QUEST+: A general multidimensional Bayesian adaptive psychometric method. J Vis 2017;17:10. [Crossref] [PubMed]
43. Walker M, Schönborn S, Greifeneder R, et al. The Basel Face Database: A validated set of photographs reflecting systematic differences in Big Two and Big Five personality dimensions. PLoS One 2018;13:e0193190. [Crossref] [PubMed]
44. Cootes TF, Edwards GJ, Taylor CJ. Active appearance models. IEEE Trans Pattern Anal Mach Intell 2001;23:681-5.
45. Kirby M, Sirovich L. Application of the Karhunen-Loeve procedure for the characterization of human faces. IEEE Trans Pattern Anal Mach Intell 1990;12:103-8.
46. O'Toole AJ, Castillo CD, Parde CJ, et al. Face Space Representations in Deep Convolutional Neural Networks. Trends Cogn Sci 2018;22:794-809. [Crossref] [PubMed]
47. Burton AM, Vokey JR. The Face-Space Typicality Paradox: Understanding the Face-Space Metaphor. Q J Exp Psychol Sect A 1998;51:475-83.
48. Valentine T. A unified account of the effects of distinctiveness, inversion, and race in face recognition. Q J Exp Psychol A 1991;43:161-204. [Crossref] [PubMed]
49. Valentine T, Lewis MB, Hills PJ. Face-space: A unifying concept in face recognition research. Q J Exp Psychol (Hove) 2016;69:1996-2019. [Crossref] [PubMed]
50. Daube C, Xu T, Zhan J, et al. Grounding deep neural network predictions of human categorization behavior in understandable functional features: The case of face identity. Patterns (N Y) 2021;2:100348. [Crossref] [PubMed]
51. Egger B, Smith WAP, Tewari A, et al. 3D Morphable Face Models—Past, Present, and Future. ACM Trans Graph 2020;39:1-38.
52. Paysan P, Knothe R, Amberg B, Romdhani S, Vetter T. A 3D Face Model for Pose and Illumination Invariant Face Recognition. In: 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance. 2009;296-301. doi: 10.1109/AVSS.2009.58.
53. Yildirim I, Belledonne M, Freiwald W, et al. Efficient inverse graphics in biological face processing. Sci Adv 2020;6:eaax5979. [Crossref] [PubMed]
54. Miller EJ, Foo YZ, Mewton P, et al. How do people respond to computer-generated versus human faces? A systematic review and meta-analyses. Comput Hum Behav Rep 2023;10:100283.
55. Or CC, Goh BK, Lee ALF. The roles of gaze and head orientation in face categorization during rapid serial visual presentation. Vision Res 2021;188:65-73. [Crossref] [PubMed]
56. Brainard DH. The Psychophysics Toolbox. Spat Vis 1997;10:433-6.
57. Chandna A, Ghahghaei S, Foster S, et al. Higher Visual Function Deficits in Children With Cerebral Visual Impairment and Good Visual Acuity. Front Hum Neurosci 2021;15:711873. [Crossref] [PubMed]
58. McCulloch DL, Mackie RT, Dutton GN, et al. A visual skills inventory for children with neurological impairments. Dev Med Child Neurol 2007;49:757-63. [Crossref] [PubMed]
59. Ortibus E, Laenen A, Verhoeven J, et al. Screening for cerebral visual impairment: value of a CVI questionnaire. Neuropediatrics 2011;42:138-47. [Crossref] [PubMed]
60. Kran BS, Lawrence L, Mayer DL, et al. Cerebral/Cortical Visual Impairment: A Need to Reassess Current Definitions of Visual Impairment and Blindness. Semin Pediatr Neurol 2019;31:25-9. [Crossref] [PubMed]
61. Eliasson AC, Krumlinde-Sundholm L, Rösblad B, et al. The Manual Ability Classification System (MACS) for children with cerebral palsy: scale development and evidence of validity and reliability. Dev Med Child Neurol 2006;48:549-54. [Crossref] [PubMed]
62. Charness N, Dijkstra K, Jastrzembski T, et al. Monitor Viewing Distance for Younger and Older Workers. Proc Hum FACTORS Ergon Soc 52nd Annu Meet. 2008;1614-7.
63. Dear M, Harrison WJ. The influence of visual distortion on face recognition. Cortex 2022;146:238-49. [Crossref] [PubMed]
64. Näsänen R. Spatial frequency bandwidth used in the recognition of facial images. Vision Res 1999;39:3824-33. [Crossref] [PubMed]
65. He J, Bex PJ, Skerswetat J. Rapid measurement and machine learning classification of colour vision deficiency. Ophthalmic Physiol Opt 2023;43:1379-90. [Crossref] [PubMed]
66. Neupane S, Skerswetat J, Bex PJ. Comparison of Foraging Interactive D-prime and Angular Indication Measurement Stereo with different methods to assess stereopsis. PLoS One 2024;19:e0305036. [Crossref] [PubMed]
67. Skerswetat J, He J, Shah JB, et al. A new, adaptive, self-administered, and generalizable method used to measure visual acuity. Optom Vis Sci 2024;101:451-63. [Crossref] [PubMed]
68. Dutton G, Bax M. Visual Impairment in Children Due to Damage to the Brain: Clinics in Developmental Medicine. Vol. 186. London, UK: Mac Keith Press; 2010.
69. Piepers DW, Robbins RA. A Review and Clarification of the Terms "holistic," "configural," and "relational" in the Face Perception Literature. Front Psychol 2012;3:559. [Crossref] [PubMed]
70. Wallis TS, Taylor CP, Wallis J, et al. Characterization of field loss based on microperimetry is predictive of face recognition difficulties. Invest Ophthalmol Vis Sci 2014;55:142-53. [Crossref] [PubMed]
71. Dakin SC, Watt RJ. Biological "bar codes" in human faces. J Vis 2009;9:2.1-10.
72. Goffaux V, Hault B, Michel C, et al. The respective role of low and high spatial frequencies in supporting configural and featural processing of faces. Perception 2005;34:77-86. [Crossref] [PubMed]
73. Goffaux V, Rossion B. Faces are "spatial"--holistic face perception is supported by low spatial frequencies. J Exp Psychol Hum Percept Perform 2006;32:1023-39. [Crossref] [PubMed]
74. Gold J, Bennett PJ, Sekuler AB. Identification of band-pass filtered letters and faces by human and ideal observers. Vision Res 1999;39:3537-60. [Crossref] [PubMed]
75. Wiecek E, Dakin SC, Bex P. Metamorphopsia and letter recognition. J Vis 2014;14:1. [Crossref] [PubMed]
76. Martelli M, Majaj NJ, Pelli DG. Are faces processed like words? A diagnostic test for recognition by parts. J Vis 2005;5:58-70. [Crossref] [PubMed]