Cells in the early parts of the visual pathway have small receptive fields that are well tuned for orientation, direction of motion, and disparity. The mechanisms by which these local estimates are combined to extract the features of extended objects and surfaces have been the subject of intense investigation over the last 20 years. One of the most fruitful avenues for studying the integration of local estimates of orientation has been the study of contours defined by Gabor patterns (Field, Hayes,& Hess, 1993; Kovacs & Julesz, 1993). Gabor-defined contours are visible only against a dense noise background if the elements of the contour lie nearly collinearly along the spine of the contour (Bex, Simmers, & Dakin, 2001; Field et al., 1993). The preference for collinearity may be driven by a similar bias that is present in images of natural scenes (Elder & Goldberg, 2002; Geisler, Perry, Super, & Gallogly, 2001; Kruger, 1998; Kruger & Worgotter, 2002; Sigman, Cecchi, Gilbert, & Magnasco, 2001).

Performance in the contour-in-noise task is disrupted in amblyopia, presumably as a result of abnormal visual experience during visual development (Chandna, Pennefather, Kovacs, & Norcia, 2001; Hess & Demanins, 1998; Hess, McIlhagga, & Field, 1997b; Kiorpes & Bassin, 2003; Kovacs et al., 2000; Kozma & Kiorpes, 2003). There is some debate as to whether all types of amblyopia show contour integration losses. Hess and coworkers have reported that contour integration was abnormal in the amblyopic eyes of patients with strabismus, but not in anisometropic amblyopia (Hess & Demanins, 1998; Hess et al., 1997b). However, Chandna, Pennefather, Kovacs, and Norcia (2001) found that untreated anisometropic amblyopes also had deficits in their amblyopic eyes, and Kozma and Kiorpes (2003) have found deficits in macaques with both types of amblyopia. Deficits have also been found in both eyes of nonamblyopic patients with strabismus (Kovacs, Polat, Pennefather, Chandna, & Norcia, 2000) and in some fellow eyes of amblyopic macaques that had normal acuity and contrast sensitivity (Kozma & Kiorpes, 2003).

The presence of deficits in contour integration in eyes with normal visual acuity suggests that abnormal binocular interaction may disrupt the elaboration of contour integration mechanisms. Although most studies of contour integration have used two-dimensional tasks, it is clear that contours in the natural environment exist in three dimensions. Consistent with this, observers with normal stereopsis are better able to segregate collinear contours from noise backgrounds if the contour and noise patches lie in different depth planes (Hess, Hayes, & Kingdom, 1997a), and they can integrate contours whose elements lie in different depth planes (Hess & Field, 1995).

Here we use event-related potentials and multivariate statistical analyses to confirm that contour integration deficits can exist in the absence of a deficit in acuity. We show this dissociation between visual resolution and global integration in the dominant eyes of patients with strabismic amblyopia and in both eyes of stereoblind patients who have strabismus, but no amblyopia. We have also identified distinct patterns of loss in patients with different histories of binocular interaction and an electrophysiological correlate of sighting dominance in normal observers.

A total of 59 adults participated. All normal adult observers (n = 24) had Snellen acuity correctable to 6/6 or better in each eye and no prior history of strabismus or amblyopia. Participants with a history of either strabismic amblyopia (n = 10), anisometropic amblyopia (n = 13), or strabismus without amblyopia (n = 11) also participated. The research protocol was approved by the Institutional Review Board of the California Pacific Medical Center and conformed to the tenets of the Declaration of Helsinki. Written informed consent was obtained after the visual evoked potential (VEP) recording procedure was explained. Clinical details of the patients are presented in Table 1 (for information on refractive error, see Table in Appendix). Sighting dominance was established by asking each observer to raise a card with a small hole in it (1 cm) with both hands, so they could see a distant target (6 m) through the hole. One eye was then occluded and the eye that retained sight of the target was designated as the dominant eye. Stereo acuity was measured with the Randot and Titmus stereotests. Amblyopia was considered to be present if the acuity difference between eyes exceeded 0.2 LogMAR. The subjects were classified according to the way they presented at the time of the experiment. As seen from Table 1, most of the individuals with strabismic amblyopia had anisometropia. Of the 13 individuals who had anisometropic amblyopia, only one is known to have a history of corrective surgery for strabismus. Six of 11 individuals in the nonamblyopic strabismic group were diagnosed and treated for amblyopia in their childhood.

Stimulus generation and signal analysis were performed by in-house software running on separate Power Macintosh G3 computers. Stimuli were presented on a multi-synch video monitor (800 by 600 pixels; 72-Hz vertical refresh; 150-MHz video bandwidth; MRHB2000, Richardson Electronics, Inc., http://www.rell.com/), which was positioned at 57 cm, generating visual fields of 27° X 23°. The mean luminance was 125 cd/m².

We studied interactions between Gabor element stimuli by varying their configuration. In each configuration, the centers of each Gabor patch were tied to invisible circular “contours.” In the “circle” configuration (Figure 1, left), all Gabor patches had their carrier orientations tangent to the circle. In the “pinwheel” configuration (Figure 1, middle), all the patches had a 60° orientation offset with respect to the local tangent. The pinwheel configuration was used as a control for statistical regularity as a possible basis for difference between responses in the circle and random configurations. In the random configuration (Figure 1, right), the elements had a random orientation with respect to the implicit contour.

There were 12 Gabor patches in each contour, and there were 11 contours present on the screen. The contours were arranged on a hexagonal grid with a 8.5-deg center-to-center spacing. Individual contours were 6.2 deg in diameter. The Gabor carrier spatial frequency was 2 c/deg and the patches were spaced by 1.5 deg (3 wavelengths) along the contours. The SD of the Gabor patches was 0.18 deg, and the carrier contrast was 80%, according to the Michelson definition. We used multiple contours to increase the amplitude of the evoked response and to obviate the need for strict fixation on a single contour. The contours were presented at full contrast for 500 ms with no change in mean luminance, followed by a return to mean luminance for 500 ms (periodic pattern onset/offset presentation at a frequency of 1 Hz).

Recording sessions consisted of ten 10-s trials per condition. The trials were randomly interleaved across conditions in blocks of 5 trials. Viewing was binocular in the first experiment, but was monocular for the subsequent experiment.

Five electrodes were placed over the occipital pole at O₁, O_Z, and O₂ of the International 10-20 system plus two sites 3-cm lateral to O₁ and O₂. The reference and ground electrodes were placed at C_Z and P_Z, respectively. The EEG was amplified at a gain of 50,000 with amplitude bandpass-filter settings of 0.3 to 100 Hz at -6dB (Model 12 A5; Grass Instruments, Quincy, MA). The EEG was digitized to a nominal 16 bits accuracy at 432 Hz (PCI-MIO-16XE-50; National Instruments, www.ni.com). The horizontal synch signal from the video card was conditioned and used to clock the A/D converter (six samples per video frame). The display was updated during the vertical blanking interval, and the vertical synch signal was provided via a digital input line to the data acquisition routine for exact synchronization of the data acquisition to the display.

Conventional time-locked averages were computed over 1000-ms time epochs. In these records, the first transition was from the blank screen to the patterned screen with pattern offset occurring at 500 ms. Difference potentials were calculated, and the statistical significance of the difference at each time point was tested using permutation methods (Blair & Karniski, 1993). The permutation testing procedure accounts for the correlation between time-samples and points of significant difference are indicated by black dots on the difference potential. Primary analyses were conducted using partial least squares (PLS) as described by Lobaugh, West, and McIntosh (2001). PLS is a multivariate technique that can be used to systematically summarize differences between experimental conditions in terms of spatial (electrode location) and temporal (response latency) variables. After computing a mean waveform across subjects for each relevant stimulus condition, we subtracted the mean for each condition from the mean across all conditions. The resulting deviation waveforms for each condition were gathered into a matrix, which was then subjected to singular-value decomposition. This generates a set of component basis vector pairs that corresponded to the latent variables (LVs) of the deviation matrix. One member of the pair of vectors comprising a latent variable is a “singular waveform” that identifies electrodes and time points that covary with different conditions of the experimental design. The second member of the pair (the “singular profile”) represents the loading of each condition on this latent effect. Significance of latent variables and their localized effects in space and time were estimated using nonparametric re-sampling techniques as described in Lobaugh et al. (2001). Additional details of the PLS and permutation methods are provided as an Appendix.

The scalp-recorded VEP to Gabor-defined contours consisted of a multiphasic response that is typical of pattern appearance response recorded with other stimuli: an initial positive peak at 85 to 95 ms, a negative peak at 125-135 ms, followed by a positivity peaking around 220 ms (Figure 2). Each waveform shown here is the average across trials and subjects (N = 13). The response to the circle configuration is larger than the response to either the pinwheel configuration or the random configuration (compare thin and thick lines in Figure 2, left and middle) and the response differs little between the pinwheel and random configurations (Figure 2, right). PLS analysis yields a single latent variable whose time course and topography are shown in Figure 3 (panel I), along with the difference potential between the circle and pinwheel configurations. In this simple case, there is only one pattern of difference and only one LV is recovered. The waveform of the LV matches both difference potentials between circle responses and those from pinwheel or random stimuli (only one difference potential is shown for clarity). Filled symbols indicate the time points of significant difference at the two SE criterion from the PLS analysis. Significant configural effects are first apparent as early as 90 ms, corresponding to an enhancement of the initial positivity, but they are more consistently observed at 190 ms (enhancement of the second positivity) and at 250-300 ms, corresponding to a sharpening of the second positivity. Effects on the negativity at 125-130 were not significant in the PLS analysis.

As one expects from the pattern of the difference potentials in Figure 2, the condition weights (Figure 3, II) show a simple contrast pattern between the circle and the other two configurations. Note that the sign of the weights is arbitrary. The pattern of results shown in Figures 2 and 3 indicates that the enhancement and sharpening of the response is not simply due to the circle configuration being more regular than the random pattern, because both circle and pinwheel configurations are equally regular. The circular configuration is locally collinear, whereas the pinwheel configuration is not. This control experiment indicates that the circle/random comparison used in the main analysis reflects collinear interactions.

Monocular data were collected from 14 normal observers (3 of whom were in the binocular experiment) as baseline data for comparison with the monocular data from the patient groups. Figure 4 shows the average waveform across trials and subjects at Oz. The configuration effect (circular vs. random) was seen again, but it was more prominent in the dominant eye determined by sighting (Figure 4, row 1). Note the relatively fewer points of significant difference when the configuration effect is tested in the nondominant eye (NDE), compared to the dominant eye (DE) (Figure 4, row 1, left panels), and the presence of significant differences between eyes for the circle configuration (Figure 4, row 1, right panels). This is somewhat surprising, given that these observers have clinically normal vision and equal acuity in both eyes. The effects seen in Figure 4 (row 1) were tested formally for significance using PLS. The four stimulus conditions could in principle support up to three latent variables; however, only one was significant and it is plotted in the leftmost panel of Figure 5A. The condition weights are shown in Figure 5B, directly under the waveforms. The contrast is greatest between circles in the dominant eye (+0.8) versus random patterns in the nondominant eye (-0.6). The remaining conditions produce similar near zero weights, consistent with the difference potentials shown in Figure 4. These results indicate that there is a greater sensitivity to stimulus configuration in the dominant eyes of normals than in their nondominant eyes. In this experiment, significant effects were also seen in the 125-135-ms range for both difference potential (Figure 4, row 1) and PLS analyses (Figure 5), but not at the later 250-300-ms range.

This result was sufficiently provocative to motivate a follow-up experiment in which we presented the same 14 observers with full-field grating stimuli presented with the same spatial frequency, contrast, and timing parameters as used for the contour stimuli. There are only sporadic differences between the two eyes, none of which occur in the latency range of the onset response (Figure 6). There was no significant difference revealed by PLS analysis–the first (and only) latent variable had a p value of .21. The full-field response waveform is quite different from that evoked by the contour stimuli–there is no negative peak in the 125-to-135-ms range, but rather a large positivity, followed by a second positivity. We thus conclude that the effect of eye dominance in normal observers seen in Figures 4 and 5 is stimulus specific and is not due to a general superiority of the dominant eye.

Anisometropic amblyopia (amblyopia without clinically apparent strabismus) resulted in response waveforms and a pattern of eye and configuration effects that were very similar to those observed in normal observers: The circular configuration produced the largest differential response in the dominant, nonamblyopic eye, and the difference between eyes was greatest for the circular configuration (Figure 4, row 2).

PLS analysis produced a single highly significant latent variable (Figure 5, second column) that had the same weight profile as that of the normal observers. The peak-to-peak amplitude of the latent variable is larger in the nonstrabismic group, reflecting the larger overall differences between the dominant and nondominant eyes of the amblyopes. Visual deprivation from anisometropia thus appears to simply exaggerate the normal pattern of eye dominance.

In contrast to the results in the anisometropic group, observers with strabismic amblyopia show a loss of configural sensitivity in both eyes. This can be seen in the difference potentials (left panels, Figure 4, row 3) where there are few points of significant difference between circle and random conditions in either eye but there are consistent differences between eyes for both circle and random conditions (right panels, Figure 4, row 3). These effects reflect themselves in a different pattern of weights in the PLS analysis (Figure 5), which recovered a single significant latent variable (p < .005). The contrasts are now between the dominant eye (positive weights) and the amblyopic eye (negative weights) for both stimuli, rather than showing a linking of configuration and eye dominance as in the case of the normals and anisometropes. The weights are the same for both configurations in the amblyopic eye, suggesting that there is no sensitivity to configuration. The difference in weights between the two configurations in the fellow eyes is less than it is in either the normals or the anisometropes, consistent with reduced configural sensitivity in the dominant eyes. In addition to having a different pattern of eye and configuration effects, strabismic amblyopes have a different underlying response waveform. The amplitude of the positive peak at 200 ms is reduced in the normal acuity-dominant eye by about a factor of two relative to the other groups. This peak is absent in the amblyopic eyes of the patients with strabismus but is present in the patients with anisometropia. The two groups have similar levels of amblyopia (strabismics: 0.66 LogMAR vs. anisometropes: 0.54 LogMAR), so the waveform effects are not due to large between-group acuity differences.

We next asked whether amblyopia was necessary to produce defects in the dominant eyes of the strabismus patients. For this comparison, we took advantage of a subtype of strabismus patients–those whose history and clinical presentation suggest that they habitually alternated fixation during early visual development and thus avoided developing amblyopia. This group of patients was the only group to produce two significant latent variables, which are shown in the rightmost columns of Figure 5. Their difference potentials are shown in Figure 4, row 4. The first latent variable has a weight pattern that shows equal configural sensitivity in each eye. (In Figure 5, panel B, the difference in weights is the same for Circle-DE and Random-DE, as it is for the Circle-NDE and Random-NDE.) The second latent variable is eye specific, rather than configuration specific: It has the same weight pattern as that of the strabismic amblyopes, although the signs of the weights (which are arbitrary) are inverted. This group of patients has normal, high visual acuity in each eye, although this does not appear to be sufficient to make their configural sensitivity the same as that of normal observers: These patients also have a small but measurable abnormality, reflected in the second latent variable, which is similar to that seen in the patients with strabismic amblyopia. This portion of their response indicates a specific, temporally localized (ca 110 ms) loss of configural sensitivity in both eyes. The response waveforms in this group of patients are more similar to those of normals and anisometropes, with equivalent amplitudes at 200 ms and slightly reduced negative peak at 140 ms.

We have found larger amplitudes for collinear contours compared to misaligned contours, especially in dominant eyes. This increase in activity may underlie the increases in the BOLD signal for similar stimuli that been recently seen using fMRI (Altmann, Bulthoff, & Kourtzi, 2003). This difference is first apparent at approximately 100 ms under binocular viewing conditions, but is largest and most consistent at around 200 ms. Previous evoked-potential studies have found enhanced responses for collinear arrangements of Gabor patches compared to noncollinear ones (Polat & Norcia, 1996) and for Gabor patches that were elongated along the orientation axis, compared to elongations orthogonal or oblique to the orientation axis (Polat & Norcia, 1998). Oka, van Tonder, and Ejima (2001) measured evoked responses to square-shaped figures defined by Gabor patches and found waveform differences between collinear versus mixed orientation figures in the 180-to-220-ms range that were consistent with faster latencies for the collinear configuration. The normal pattern of interaction between collinear Gabor patches is disrupted in amblyopia, both psychophysically and in the VEP (Polat, Sagi, & Norcia, 1997).

In normal observers, eye dominance enhanced the differences in the contrast response to aligned versus misaligned contours as early as 90 ms, with the most prominent enhancement occurring around 200 ms. This superiority of the eye that is sighting dominant does not extend to grating stimuli that were matched on all other presentation parameters, such as spatial frequency, contrast, and temporal frequency. Eye dominance in normal observers is poorly understood. Our task, which is typical, involves aligning a proximal target (the hole in the card) with a distant target. Why dominance in a sighting task would reflect itself in contour integration is unclear. The apparent eye dominance we have observed may be an indirect effect of the mechanism that determines sighting dominance. The effects we have observed may have been due to a subtle form of rivalry induced by the occlusion needed to perform monocular testing. Under monocular viewing, the two eyes’ images do not match and observers occasionally report a blanking out of the image from the viewing eye. If this intermittent form of suppression acts more strongly on contour integration mechanisms than it does on grating contrast responses, the pattern we found in normal observers might be seen.

We also found what appears to be an exaggerated pattern of normal eye dominance in anisometropic amblyopes. The similarity of the anisometropic deficit to the normal pattern of eye dominance suggests that monocular deprivation by blur has less severe consequences for contour integration mechanisms than does strabismus, which produces a different pattern of loss and waveform abnormalities. Psychophysical deficits on contour-in-noise detection are consistently observed in strabismus (Hess & Howell, 1977; Kovacs et al., 2000; Kozma & Kiorpes, 2003), but not always in anisometropic amblyopia (Hess & Damanina, 1998; Chandna et al., 2001). Anisometropic amblyopes have essentially normal motor alignment, and most of these patients (8/13) also had demonstrable stereopsis in spite of having reduced visual acuity in their amblyopic eyes.

In contrast, patients with strabismus lose configural sensitivity in both eyes, regardless of whether visual acuity is reduced or not. While deficits in amblyopic eyes are perhaps not surprising and have been reported previously, losses in eyes with normal visual acuity require an explanation that does not depend on the reduction of high spatial frequency sensitivity that underlies acuity loss. In addition, the individuals with good visual acuity in both eyes and poor stereopsis (nonamblyopic strabismus) also have a defect in configural sensitivity that manifests as a second latent variable whose weight pattern indicated differences between dominant and nondominant eyes but no difference between configurations.

A potential common factor in patients showing configural deficits is a lack of stereopsis: 19 out of 21 patients with strabismus were stereoblind. In contrast, 8 of 13 anisometropic amblyopes had some degree of measurable stereopsis. At first glance it would seem that our task has little to do with stereopsis because the recordings were done under monocular viewing conditions. However, it appears that contour integration mechanisms are disparity selective: Noise elements that lie in different depth planes have less effect on contour visibility (Hess et al., 1997a), and placing contours in a different depth plane than the noise background produces an enhanced BOLD response in lateral occipital cortex (Altmann et al., 2003). A developing visual system without access to this mechanism for segregating contours from their background may not fully develop. A current developmental model of contour integration (Prodohl, Wurtz, & von der Malsburg, 2003) relies on common motion cues to augment an innate bias for collinear stimuli. Continuity in depth may serve a similar role in selecting connections that are optimal for segregating contours and figures from noisy backgrounds. Evolutionary pressure acting within an environment where collinearity in three dimensions is a prominent feature may have led certain cortical mechanisms to “expect” fully stereoscopic input for their development. Mechanisms that are "designed to utilize the sort of environmental information that is ubiquitous and has been so throughout much of the evolutionary history of the species" have been termed “experience expectant” (Greenough, Black, & Wallace, 1987, p. 291). This is to distinguish these mechanisms from “experience dependent” mechanisms that are unique to the experience of particular individuals.

The fMRI results (Altmann et al., 2003) suggest that lateral extrastriate cortex is a dominant player in contour integration. This area is also activated by stereoscopic stimuli (Mendola, Dale, Fischl, Liu, & Tootell, 1999; Tsao et al., 2003). The most robust effects of configuration in our data are in the 200-ms range, which is consistent with an extrastriate generator, given that human V1 is first activated at around 50 ms (Ducati, Fava, & Motti, 1988; Moradi, Liu, Cheng, Waggoner, Tanaka, & Loannides, 2003). What little is known about the development of extrastriate cortex suggests that development is completed later than in primary visual cortex, resulting in greater exposure during what may be later critical periods (Landing, Shankle, Hara, Brannock, & Fallon, 2002; Rodman, 1994; Schroder, Fries, Roelfsema, Singer, & Engel, 2002).

The analyses described hereafter were performed separately for each recording channel. For a given group of subjects (e.g., normals and anisometropic amblyopes), we calculate the mean response waveform averaged over trials for each subject and for each condition. We then compute the mean waveform averaged over subjects for each condition, and the grand mean waveform averaged over subjects. Each condition mean waveform is subtracted from the grand mean and collected into a t-by-m deviation matrix, D, where t is the number of time points in the waveform, and m is the number of conditions. The columns of D represent the deviation of each condition mean from the grand mean.

This operation generates a set of three matrices, W, S, and C, that satisfy the linear algebraic expression D = WSC', where W denotes the matrix transpose of W. C is a symmetric m-by-m matrix whose rows contain loading coefficients for each condition; W is a t-by-m matrix whose columns contain linear combinations of the deviation waveforms in D; and S is a diagonal m-by-m matrix whose diagonal elements are positive weighting coefficients called singular values. The rows of C and the columns of W are ordered so that the first row of C is paired with the first column of W, the second row of C with the second column of W, etc. Each of these pairs is mutually orthogonal and are basis vectors of the multidimensional space spanned by D. The outer product of a pair of basis vectors (one from C and one from W), weighted by the corresponding singular value from S, will generate a t-by-m matrix that represents the proportion of the deviation matrix accounted for by this basis vector pair. Taken together, a pair of basis vectors and its corresponding singular value are termed a latent variable, and are denoted LV1, LV2, etc., in descending order of singular value.

In Figures 3 and 5, the bar charts and the waveforms respectively depict the row of C and the column of W for a given latent variable. From the properties of SVD, it follows that the waveform of a given latent variable is obtained by multiplying the deviation waveform for each condition (i.e., the columns of D) by the corresponding element from the bar chart, dividing by the singular value, and summing. In other words, the LV waveform is a linear combination of the condition deviation vectors weighted by the loading coefficients for each condition. For example, the waveforms in Figure 3A were roughly 0.8A – 0.5B – 0.2C, where A, B, and C were the deviation waveforms for the three experimental conditions. Although the absolute sign for a given loading coefficient is essentially arbitrary, the signs and magnitudes of the various coefficients with respect to each other are of obvious importance.

Given a null hypothesis of no effect due to stimulus condition, the waveform responses for each condition within a given subject are exchangeable. Random permutations of condition exchanges were applied to each subject's data to generate a new deviation matrix, D, and then SVD was repeated. Repeated re-randomization of conditions and application of SVD generated a reference distribution of LV data to which the original, un-exchanged data could be compared. An LV was deemed significant if its singular value ranked above 95% of the corresponding singular values in the reference distribution. All LVs depicted in this study were significant by this criterion.

For a given condition, a new data set was obtained by sampling with replacement from the pool of individual subject averages for that condition. Then a new deviation matrix D was created from this bootstrap data set, and SVD was applied. Repeated re-sampling and application of SVD generated a reference distribution of LV waveform data, from which the SE of each time point was obtained. A value at a given time point in the original LV waveform was deemed significant when it exceeded twice the SE from the corresponding time point in the reference distribution.

As before, given a null hypothesis of no effect due to stimulus condition, the waveform responses for any two conditions from a given subject are exchangeable. Response waveforms were randomly exchanged for each individual in the pool of subjects from a given group. For this permutation sample, we calculated the mean difference potential and the T value of this difference for each time point in the response waveform. Repeatedly re-randomizing the permutation of condition exchanges for each subject allowed us to accumulate a reference distribution of T values. From each permutation sample, we noted the maximum T value over all time points in the response waveforms, and accumulated these maximum T values into a second reference distribution. The difference potential at a given time point in the original, unexchanged response data was deemed significant if its T value exceeded 95% of those in the maximum T value reference distribution.