How Do the Blind 'See'? The Role of Spontaneous Brain Activity in Self-generated Perception

Avital Hahamy; Meytal Wilf; Boris Rosin; Marlene Behrmann; Rafael Malach


Brain. 2021;144(1):340-353. 

In This Article

Materials and Methods


This study reports findings from five late-onset blind/visually-impaired individuals diagnosed with CBS (age 47 ± 8.9, two females, two left handed, one ambidextrous), 11 late-onset blind individuals not experiencing visual hallucinations (blind control group, age 40.54 ± 11, four females, two left handed, one ambidextrous) and 13 sighted individuals with normal or corrected-to-normal vision (sighted control group, age 43.85 ± 7.4, 10 females, two left handed). Experimental groups did not differ in age, gender or handedness (all P-values > 0.27, Fisher's exact test). Participants had no history of psychiatric illness or cognitive impairments, and were not taking any psychoactive medications. Recruitment of participants was carried out with the assistance of a neuro-ophthalmologist, and through standard advertisements. All participants provided written informed consent before participating, in accordance with the Declaration of Helsinki, and were paid for their participation in the study. All procedures were approved by the Tel-Aviv Sourasky Medical Center, IRB ethics committee. Table 1 presents demographic, clinical and hallucination phenomenology information of the CBS participants (see also Supplementary material, 'Hallucination phenomenology' section).

Experimental Design

We aimed to study hallucinations (unprompted perceptual events) in CBS as a model for unprompted behaviours. To isolate the unprompted component of hallucinations, we compared hallucinations in CBS to cued veridical vision in sighted controls who were presented with visual simulations of these hallucinations. However, while the unprompted nature of visual hallucinations differentiates them from veridical vision, another difference is that hallucinations are internally generated while veridical vision is evoked by external stimulation. To control for this difference, we further compared hallucinations to cued visual imagery, as both of these conditions are internally generated, but only hallucinations are unprompted. Finally, to ensure that hallucination-related brain activations are not the product of mere verbal/manual report, we used a control condition consisting of verbal and manual tasks that were unrelated to the onsets of hallucinations.

The experimental procedures included the administration of questionnaires, and between one and three functional MRI sessions per participant. Session times ranged between 45 min and 120 min. In aggregate, these sessions included one resting-state scan (all groups), one to two imagery scans (all groups), an anatomical scan (all groups), a verbal-manual control condition (CBS participants/blind controls), and one visual localizer scan (sighted controls). We also conducted two to four hallucination scans reported verbally or using button-presses for each CBS participant, and for sighted controls, we conducted six simulated hallucination scans (as elaborated below). The order of scans was counterbalanced across participants with the exception that the first session always began with a resting-state run. Analyses of the visual localizer and resting state scans will not be reported here. Experimental software is detailed in the Supplementary material.


Questionnaires were used to collect demographic and clinical details. Additionally, the Vividness of Visual Imagery Questionnaire (VVIQ; Marks, 1973) was administered to assess how vividly participants imagined different scenes and situations. This questionnaire was administered twice, with participants' eyes being open/shut, and the scores across the two administrations were summed per participant.

Hallucination Report

During prescanning simulations, Participants CBS1–3 stated that they could report their hallucination as easily and promptly as they could identify visual stimuli before their vision deteriorated. Participants CBS4 and CBS5 said they were unable to report their hallucinations, and were therefore excluded from the report condition and subsequent analyses (but were included in all other conditions and analyses). The report condition consisted of several 8-min scans (number of scans depended on the availability/stamina of participants; Table 1), in which participants provided reports of their hallucinations either verbally, or via button presses. Because some CBS participants had some, albeit minimal, residual vision, participants were instructed to close their eyes during the entire scanning procedure. Participants were trained to speak without moving their heads, both outside and inside the scanner.

In-scanner verbal reports of hallucinations were recorded, and played back to the participants outside the scanner at the end of each session, asking them to give details of these hallucinatory events. Because the visual acuity of all CBS participants deteriorated at a relatively late stage of life, their description of the hallucinations was based on their prior visual experiences.

During the button-press runs, Participant CBS1 pressed a button using her index finger whenever an image appeared (note that hallucinatory images were constantly replaced by other images with no interval between them). Participants CBS2 and CBS3 pressed a button with the index finger whenever a face appeared and pressed a second button using the middle finger when the face disappeared.

Simulated Hallucinations

The temporal structure of the in-scanner verbal reports made by the CBS participants, along with the post hoc details regarding the hallucinations' content, were used to create movies simulating these hallucinatory streams. Three such movies were created, one corresponding to each of the reporting CBS participants. The simulated hallucinations of Participant CBS1 consisted of images of humans, animals, body parts, objects, houses and patterns, presented in different sizes and positions on a grey background. The simulated hallucinations of Participant CBS2 consisted of a video recording of a male face, made small enough to move around the grey screen. The simulated hallucinations of Participant CBS3 consisted of various pictures and of video recordings of faces/patterns, all presented on a grey background. To account for the possible latency between the true onset of hallucinations and the actual reports made by CBS participants, all simulated stimuli were presented 1 s prior to their real temporal position, as reported by the CBS participants (Ben-Yakov and Henson, 2018).

Sighted control participants watched these simulated hallucination streams in an order that was counterbalanced across participants. Each simulated hallucination stream was watched twice, with instructions to report the hallucinatory content verbally or using button presses, as the CBS participants did. Participants were trained to speak without moving their heads. Two sighted control participants only completed the verbal-report runs, and additional four participants had one to three of their six scans excluded from further analyses because of excessive head motion. Nevertheless, all participants had at least one valid scan for each simulated hallucination stream.

Visual Imagery

All participants were asked to imagine faces, houses, objects and patterns. Before the scan, participants were given examples of items from each category. This 8-min run comprised 12-s blocks, each beginning with an auditory cue signalling a category name. These blocks ended with the auditory instruction 'rest', which was followed by an 8-s resting period. Block order was pseudo-randomized across participants. All participants closed their eyes during this experiment. CBS and all blind control participants completed two separate runs of this experiment (except for two blind controls, who completed only one run), and sighted controls completed one run. Data from one blind control were excluded from further analyses due to excessive head motion.

At the end of each run, participants assessed their success level in imagining each visual category on an increasing success scale of 1–5. These ratings were summed per participant.

Verbal-manual Control Condition

To test whether verbal or manual reports alone evoke activity in the visual system, CBS and blind control participants performed a tone discrimination task. During this 8 min 6 s scan, participants heard a second-long tone of either 440 Hz or 460 Hz, interleaved with silent periods of 3–5 s. Stimulus order was pseudo-randomized across participants. Participants spoke/pressed a button when presented with the higher/lower frequency tone, respectively. Participants were trained to differentiate between the two tones before being scanned. Data from one blind control were excluded from further analyses due to excessive head motion.

See the Supplementary material for MRI data acquisition and preprocessing description.

Statistical Analysis

Questionnaires. Given the small sample size of participants, here and in all similar analyses, scores were compared between experimental groups using non-parametric permutation tests (Holmes et al., 1996; Nichols and Holmes, 2002). Here, each test statistic was set to the difference between the group means. Under the null hypothesis of no group difference in imagery capabilities, participants' group labels were shuffled to create two random groups of participants, and the difference between these groups' means was calculated. This procedure was repeated for all possible permutations of participants between the two groups to construct the full null distribution, which was used to derive a two-tailed P-value for the true (unshuffled) test statistic.

Whole-brain Analyses. To create task-based statistical parametric maps, we applied a voxel-based general linear model (GLM) as implemented in FSL's FEAT, using a double-gamma haemodynamic response function convolved with the experimental model, as well as the resulting regressors' temporal derivatives. The six motion parameters and their derivatives, scrubbed volumes (Power et al., 2012), and ventricle and white matter time courses for each participant (Fox et al., 2009) were used as nuisance regressors. In addition, in the simulated hallucinations data, the first/last five repetition times (TRs) of each scan were included in the GLM model as nuisance variables, to remove the contribution of arousal-related effects. See the Supplementary material for a detailed description of all GLM designs.

As the sample size of the CBS group was small, whole-brain comparisons between the CBS group and any of the control groups were carried out using non-parametric randomization tests, as implemented in FSL's randomize (Winkler et al., 2014), including threshold free cluster enhancement correction for multiple comparisons. However, because of the inherent differences between the hallucination and the simulated hallucinations conditions (as it is impossible to simulate hallucinations with full precision), we refrained from directly contrasting activation strengths between these conditions, as any effects could be equally attributed to differences between hallucinations and veridical vision, or to differences in the visual stimuli.

Parametric activation maps were projected onto a template of a flattened cortical surface using the Connectome Workbench.

Quantifying the Similarity Between Visual Activations. Our hypothesis that hallucination-related activations would be similar to activations evoked by other visual experiences was tested in the posterior part of the brain (25 876 grey matter voxels corresponding to y < 44 in MNI space; Gilaie-Dotan et al., 2013). CBS hallucination activations (group beta values) were correlated with imagery/simulated-hallucinations beta values of each sighted control and with imagery/verbal-manual control condition beta values in each blind control. This calculation was performed twice for sighted controls in the simulated-hallucination condition: once modelled using the CBS report protocol, and once using a protocol locked to the sighted controls' own manual report. In both cases, for each sighted control participant, analysis was carried out using beta-value maps resulting from an FFX analysis of all three simulated-hallucinations data (corresponding to simulations of the three CBS participants' hallucinations). Resulting correlation coefficients of the participants in each group and experimental condition were tested using a two-tailed one-sample Wilcoxon test.

Note that we refrained from statistically testing the posterior brain correlations of CBS participants across the different experimental conditions, because: (i) any significant similarities between the activations evoked by the imagery/control condition to those evoked by hallucinations may be confounded by the fact that CBS participants hallucinated during all conditions; and (ii) any absence of statistical significance could be due to the lack of statistical power in testing very small samples (specifically, the largest possible effect size in a sample of n = 5 in a Wilcoxon test would correspond with a P-value of 0.03. Any smaller effect size would be non-significant under an alpha level of 0.05). Nevertheless, we assume that since CBS participants originate from the blind population, any effects found in blind controls during the imagery/verbal-manual control conditions should be representative of similar effects in CBS participants.

See the Supplementary material for a description of a bootstrap analysis testing whether the measured correlations were driven by noise.

Evaluation of Temporal Dynamics. To assess differences in the temporal dynamics of blood oxygenation level-dependent (BOLD) activity between the experimental groups, we extracted signals from an early/intermediate visual and a fusiform face area (FFA) regions of interest. Sensorimotor lip/hand regions of interest were also used as control regions for the verbal/manual report scans, respectively (see the Supplementary material for region of interest definitions and testing of activation in early/intermediate visual regions of interest).

Single participant's signals were extracted from each region of interest, z-score normalized and subjected to an event related averaging analysis, within a time window of 3 TRs prior to stimulus onset to 7 TRs after stimulus onset. These signals were later averaged across participants of the same experimental group and for each experimental condition.

The BOLD signal typically rises shortly after stimulus presentation, but due to noise factors (e.g. slight asynchronies between scanner and experimental protocol, inconsistencies in participants' attentiveness across trials, etc.) the event-related signals for individual participants may show slight random jitter (±1 TR) around the true event timings. Here, however, we had a clear prediction that the BOLD signal in visual regions across all CBS participants should consistently precede the reported onset of hallucinations, unlike in other experimental conditions or in the control groups. To test this prediction statistically, a canonical haemodynamic response function (HRF) was fitted to the event-related data of each individual participant (this was done automatically without the possibility of adjustment). This HRF was fitted to the data five times, each time with a different lag, ranging between 3 TRs prior to the modelled neural event (hallucination/imagery/vision) to 1 TR after the modelled neural event. The HRF lag that produced the best fit between the HRF and data was identified in each participant. These 'optimal lags' of the HRF in each region of interest were then compared between the CBS group and each of the sighted/blind control groups separately, using a permutation test (Supplementary material).

All sighted controls' simulated-hallucinations data (verbal and button-press scans) were analysed using a protocol locked to the hallucination report of CBS participants, and button-press scans were further analysed using a protocol locked to the individual button presses of each sighted control participant. The analysis of the sensorimotor lips/hand regions of interest made use of scans involving verbal/button-press reports, respectively.

Quantifying BOLD Temporal Dynamics Across the Visual Hierarchy. To assess differences in the onset of the BOLD responses across regions of the visual system, all regions of interest of a probabilistic atlas (Wang et al., 2015) were ranked based on their position in the visual hierarchy (Supplementary Table 1). The optimal HRF lag was calculated for each CBS participant and region of interest in the hallucination and imagery conditions, as explained earlier. Then, Spearman's correlation was calculated between the rank of all regions of interest and the group-averaged optimal lags in these regions of interest. The resulting correlation coefficients were tested using a permutation test, under the null hypothesis of no correlation between ranks across the visual hierarchy and optimal lags. region of interest ranks across the visual hierarchy were therefore shuffled 10 000 times, and, each time, the correlation coefficient between the random ranks and optimal lags was computed. Two-tailed P-values were derived based on this null distribution.

Data Availability

Statistical data and experimental materials are available upon request.