Relative blindsight in normal observers and the neural correlate of visual consciousness

We studied both the performance and the subjective report for conscious perception in a metacontrast masking paradigm (Fig. 1) to behaviorally demonstrate the phenomenon of relative blindsight. Volunteers had to make forced-choice judgments as to whether the target was a square or a diamond. We assessed the level of consciousness immediately after the participants made the forced-choice discrimination by asking them to press keys to indicate whether they actually saw the identity of the target or simply guessed what it was. Fig. 2 Upper shows the data for the eight participants. For the SOA points at 33 and 100 ms, the performance levels (i.e., accuracy rates for the square vs. diamond discrimination) were very similar. However, the subjective judgment of consciousness differed (P = 0.036, two-tailed Wilcoxon signed-rank test) in that at the earlier SOA point volunteers were more likely to claim to have just guessed the answers. Therefore, we confirmed our hypothesis: The subjective level of consciousness can differ in the absence of a difference of performance levels.

Fourteen volunteers participated in this experiment. The data for one participant were discarded because of unsatisfactory behavioral performance. The task in the fMRI study was the same as in the behavioral experiment (Fig. 1). Inspection of the data in that experiment suggested that the temporal profile of the masking function differed across subjects. Therefore, to maximize power in the fMRI study, we recreated the conditions according to the individual temporal profile of each participant. Every participant was tested with both a short SOA condition for which the SOA was at 33 ms and a long SOA condition for which the SOA points were computed online to ensure that the performance level in this condition was similar to that of the short SOA condition. To achieve this matching, we used an adaptive procedure to search for a long SOA time point for each participant such that performance level in both the short and long SOA conditions would be as similar as possible. This approach bypassed the problem of the variability of the temporal profile of the masking function mentioned above. The group average for the resulting long SOA value was 104 ms (SD, 40 ms).

The behavioral results are shown in Fig. 2 Lower. The participants were on average 68.27% correct (SD, 4.54%) in the short SOA condition, and 70.22% (SD, 5.24%) correct in the long SOA condition. The two measures did not differ significantly (P = 0.213, two-tailed Wilcoxon signed-rank test). We also used the criterion-free measure, d′, to assess performance and found that they did not differ significantly across the conditions either (P = 0.26, two-tailed paired t test): They were 0.99 and 1.12 for the short and long SOA conditions, respectively (SD, 0.25 and 0.31, respectively). In addition, the reaction times for visual discrimination in both conditions did not differ significantly (P = 0.858, two-tailed paired t test): They were 723 and 717 ms for the short and long SOA conditions respectively (SD, 180 and 192 ms, respectively). The reaction times for correct and incorrect trials were considered together so that the difference between all trials in the two task conditions were reflected, given that they were all analyzed in the fMRI analysis. These results suggest that not only was the performance of identification in the two conditions matched, but also that the task difficulty as measured by reaction times was very similar.

However, as predicted, the levels of subjective consciousness differed significantly (P = 0.007, two-tailed Wilcoxon signed-rank test). The “seen” rates were 49.17% and 60.13% for the short and long SOA conditions, respectively (SD, 23.17% and 19.79%, respectively). All but one participant showed the effect in the same direction, and 7 of the 13 participants showed a significant effect when their data were analyzed individually (P < 0.05, χ² test).

In the long SOA condition there was a significantly higher level of subjective awareness as compared with the short SOA condition, which was associated with a significant increase in activity in the left mid-dorsolateral prefrontal cortex (mid-DLPFC) (coordinates = −46, 48, 14). This analysis included both the correct and incorrect trials. The significance level of this activation was assessed by directly estimating the probability that the area under curve for the hemodynamic time course is bigger in the long SOA than in the short SOA condition. This Bayesian approach has the advantage of bypassing the problem of multiple statistical comparisons. The probability of activation was 0.97. The location of this activation and the time courses of activity for the two conditions are shown in Fig. 3. This activation lies specifically in Brodmann's area 46, the macaque monkey homologue of which lies around the principal sulcus (8).

We also performed a conventional analysis with standard parameters by fitting the canonical hemodynamic response function to the data, and we assessed the significance by using classical frequentist statistics (the t test). This approach is less sensitive because the canonical hemodynamic response function peaks at 5–6 s, and the present data do not support this assumption (Fig. 3). Nonetheless, the activation in the left mid-DLPFC survived a threshold of P < 0.001 (uncorrected), which suggests that the activation was not due to any idiosyncratic features of the initial analysis.

Although the performance levels did not differ significantly between the two conditions, the averaged value was higher in the long SOA condition than in the short SOA condition by ≈2%. To ensure that the activation in the mid-DLPFC was not due to this difference in performance, we performed another analysis, excluding the three participants who showed the biggest magnitude of superior performance in the long SOA condition. This exclusion resulted in a reversal of the trend for performance: The averaged performances in the short SOA condition and in the long SOA condition in the remaining 10 participants were 69.58% correct and 69.33% correct, respectively (they did not differ significantly, P = 0.890). Similarly, the trend reversed for the averaged d′ values for the short and long SOA conditions, which were now 1.07 and 1.06, respectively (they did not differ significantly, P = 0.940). However, the difference in subjective level of consciousness remained: The seen rates for the short and long SOA conditions were 55.04% and 62.79%, respectively (difference significant, P = 0.034). The activation in the left mid-DLPFC also remained significant (probability of >0.95).

We also tried to test whether parietal activity differed in the two SOA conditions, perhaps by a degree that was only marginally below threshold. We lowered the threshold to a probability of >0.9 for the comparison between the two SOA conditions and found no parietal activity. We also tested this hypothesis by using the canonical hemodynamic response function approach. At a liberal threshold of P < 0.01 uncorrected, we did not find any parietal activity.

We performed another analysis to test whether the activity in the left mid-DLPFC merely reflected the physical difference in the stimuli (the SOA) or whether it genuinely reflected a difference in the subjective level of consciousness. We took the activity in the mid-DLPFC at the peak of the activation for each subject for the short SOA condition, because the SOA in this condition was the same for every subject. The variability (standard deviations) in the peak location for the subjects were 6, 10, and 10 mm in the x, y, and z directions, respectively. We tested whether the level of activation correlated with the level of subjective consciousness across subjects. We found a correlation indicating that the higher the activity in the left mid-DLPFC, the higher the percentage of trials the subjects declared that they were conscious of the identity of the stimuli (r = 0.512). Because we predicted the direction of the correlation based on the results of the comparison between the short and long SOA conditions, a one-tail statistical test is justified. Under this test the correlation was significant (P = 0.037; Fig. 4).

To identify the brain areas where activity reflected discrimination performance in general, we also analyzed the fMRI results by comparing correct and incorrect trials, combining short SOA and long SOA trials. We found widespread activations in the ventral prefrontal, premotor, parietal, and temporal cortices. There also were activations in the basal ganglia and cerebellum that may relate to the fact that the reaction times for correct trials (683 ms) were quicker than for incorrect trials (815 ms) (P = 0.0001, two-tailed t test). The cortical activations were similar to those previously identified as parts of the NCC (9), as shown in Fig. 5. These activations include those in the right ventrolateral prefrontal cortex (coordinates = 36, 40, 2; probability of activation, 0.98), right intraparietal sulcus (coordinates = 42, −58, 38; probability of activation, 0.99), and middle temporal gyrus bilaterally (coordinates = 50, −36, −8 and −48, −36, −6; probability of activation, 1.00 and 0.99, respectively). However, at these coordinates, the probabilities that the activity reflected the difference in the subjective level of consciousness (i.e., long SOA > short SOA) were, in general, low: 0.45 (right ventrolateral prefrontal cortex), 0.59 (right intraparietal sulcus), 0.35 (left middle temporal gyrus), and 0.48 (right middle temporal gyrus).

Eight people participated in this study. At the beginning of the experiment, a fixation cross (0.5°) was presented at the center of a computer screen (20-inch liquid crystal display monitor, 60-Hz refresh rate), at which the participants were instructed to fixate their gaze whenever the cross was present. The disappearance of the cross marked the beginning of a trial and served as a ready signal. After 500 ms, in most trials (where SOA > 0; see the next paragraph), either a square (1°) or a diamond (i.e., a square of the same size rotated by 45°) was presented at the center of the screen for 33 ms. This presentation was followed by a blank period with a variable duration time and then a 50-ms metacontrast mask (2°) (Fig. 1) that overlapped with part of the contour of the target but not the target itself. Then, the screen went blank until 850 ms after the beginning of the trial. In the other trials (where SOA < 0), the order of presentation of the mask and the target (square or diamond) was reversed.

After the variable blank period, matched for all SOA conditions, a question appeared at 3° above the center of the screen (in Helvetica font with a size set at maximum height = 1°) that read either “Diamond or Square?” or “Square or Diamond?” (counterbalanced across participants). In the former case, the participant pressed the left key if a diamond was thought to be presented and the right key if a square was thought to be presented; the opposite was required in the case of “Square or Diamond?” After the response, a second question appeared at 3° degree below the center of the screen (same font and font size), which read “Seen or Guessed?” The participant pressed the left key if the target was actually consciously perceived and the right key if it was not and the previous response was based on intuition. After the response, the fixation cross reappeared. The trial ended at 3,650 ms after the beginning. If the participants failed to respond to both questions after this time, the question disappeared, the fixation cross reappeared, and the trial was discarded. There was an additional intertrial interval of 350 ms such that, including its following interval, each trial took exactly 4,000 ms in total. The participants were told that accuracy, but not speed, was important, so long as they were fast enough to avoid failing to respond within the 2,800-ms limit in a significant amount of trials.

The SOA values, or the differences between the onset of the target and the mask, were −50, −33, −17, 0, 17, 33, 50, 67, 83, 100, 117, and 133 ms. These time points were selected to accommodate for the refresh rate limitation of the projector used in the fMRI scanner for the main experiment. The projector operates at 60 Hz. There were 80 trials for each time point, and they were presented in a randomized order. The participants were given a chance to take a break for as long as they wished after every 40 trials.

Fourteen right-handed people (nine females) participated in this study (mean age ± SD, 24.2 ± 2.4 years), one of whom was discarded from the analysis (see below). Each volunteer gave informed consent and was screened for MRI safety before participating in the experiment. The task was similar to the one used in the pilot study, but backward masks (SOA > 0) were used so that the mask always followed the target. The participants were given a short training session (≈3 min) outside of the scanner before the fMRI experiment began to make sure that they understood and were comfortable with the tasks.

In the scanner, the participants viewed the stimuli on a projector screen through inverting mirrors, and they responded by using an MRI-compatible button box. An adaptive staircasing procedure (28) was used to ensure that every participant gave similar performance levels in the two experimental conditions. In the short SOA condition, the SOA was set at 33 ms. The SOA of the long SOA condition was 100 ms at the beginning of the experiment and was continually updated throughout the experiment according to a two-down–one-up rule: The SOA decreased by one step (16.6 ms) after two successive correct responses and increased by one step after an incorrect response. This procedure ensures that the target performance level would be at ≈71%, but it requires that there is a monotonic relationship between SOA and performance. This relationship was ensured by constraining the SOA in the long SOA condition so that it could not fall to <50 ms. To ensure that the short SOA condition matched the long SOA condition, a similar staircasing procedure was used that computed the size of the stimuli online throughout the experiment. Varying size instead of SOA for the short SOA condition has the advantage of fixing the value for the short SOA at 33 ms so that it would never be too close to that of the long SOA condition. In addition, by making sure that the performance would be at ≈71% at SOA = 33 ms, we ensured that the stimuli were sufficiently small so that the masking function would not be flat at the ceiling level for all SOA values. The size of the target started at 1° wide, as in the pilot study, but it decreased by one step (0.05°) after two successive correct responses and increased by one step after an incorrect response; the size of the mask was scaled accordingly. To ensure that the sizes of the stimuli for the two conditions were the same, a long SOA trial always used the same value obtained by the staircasing procedure from the previous short SOA trial. This procedure similarly ensures that the target performance level for the short SOA condition would be at ≈71%. The minimum target size was set at 0.25°, and the maximum was at 3°. One participant remained at the ceiling level for most of the experiment, and her overall performance was therefore far from 71% in the short SOA condition. The data for this participant were therefore discarded from the analysis.

To maximize the efficiency for detecting hemodynamic differences between the two conditions, the trials were organized such that six trials of the same condition would be presented in succession after the presentation of a series of six trials of the other condition. (The fMRI data analysis was, however, event-related.) In the train of six short SOA trials, the size of stimuli could vary, and in the train of six long SOA trials, the SOA could vary. These variations were based on the staircasing procedure described above and could last throughout the experiment to ensure that performance levels were constant across conditions and time. However, in practice, the parameters stabilized quickly in the first few minutes. Participants in total performed 240 trials for each condition (short SOA and long SOA). These were divided into four runs, between which the participant remained inside the scanner but could communicate with the experimenter. Within each run, there was also a compulsory break of 26 s given to the participant after every 40 trials, during which the participant was instructed to relax. The entire fMRI experiment lasted for ≈45 min for each participant, including the breaks between experimental runs.

Further details for the procedures of imaging data acquisition and analysis can be found in Supporting Materials and Methods, which is published as supporting information on the PNAS web site.