Secondary visual areas of sighted adults showed stronger functional connectivity with non-visual sensory areas (primary somatosensory and motor cortex, S1/M1, and primary auditory cortex, A1) than with prefrontal cortices (PFC). By contrast, in blind adults, visual cortices showed higher functional connectivity with PFC than with non-visual sensory areas (S1/M1 and A1) (group (sighted adults, blind adults) by ROI (PFC, non-visual sensory) interaction effect: F_{(1, 78)} = 148.819, p < 0.001; post-hoc Bonferroni-corrected paired t-test, sighted adults: non-visual sensory > PFC: t ₍₄₉₎ = 9.722, p < 0.001; blind adults: non-visual sensory < PFC: t₍₂₉₎ =8.852, p < 0.001; Fig. 1).

Like in blind adults, in sighted infants, secondary visual cortices showed higher connectivity to PFC than non-visual sensory areas (S1/M1 and A1) (non-visual sensory < PFC paired t-test, t ₍₄₇₄₎ = 20.144, p < 0.001) (Fig. 1). The connectivity matrix of sighted infants was also more correlated with that of blind than sighted adults, but strongly correlated with both adult groups (secondary visual, PFC and non-visual sensory areas: sighted infants correlated to blind adults: r = 0.721, p < 0.001; to sighted adults: r = 0.524, p < 0.001; difference between correlations of infants to blind vs. to sighted adults: z = 3.77, p < 0.001; see Supplementary Figure S1 for the connectivity matrices).

These results suggests that vision is required to set up the sighted adult functional connectivity pattern, i.e., vision enhances occipital cortex connectivity to non-visual sensory-motor networks and dampens connectivity to higher-cognitive prefrontal networks.

We checked the robustness of these results in a number of ways. First, we compared the effects across the three different secondary visual regions and found that the same pattern held across all three regions (Supplementary results and Supplementary Figure S2). Next, to check the robustness of the findings in infants we randomly split the infant dataset into two halves and did split-half cross-validation. Across all comparisons the results of the two halves were highly similar, suggesting the effects are robust (see Supplementary Figure S3 to Supplementary Figure S6). We performed this validation procedure for all analyses reported below with similar results.

As for secondary visual corticies, we examined the functional connectivity of the primary visual cortex (V1) with non-visual sensory areas (S1/M1 and A1) and PFC. V1 showed the same dissociation between sighted and blind adults as secondary visual areas: in sighted adults, V1 has stronger functional connectivity with non-visual sensory areas than with PFC. By contrast, in blind adults, V1 shows stronger connectivity with PFC than with non-visual sensory areas (group (sighted adults, blind adults) by ROI (PFC, non-visual sensory) interaction: F_{(1, 78)} = 125.775, p < 0.001; post-hoc Bonferroni-corrected paired t-test, sighted adults non-visual sensory > PFC: t ₍₄₉₎ = 9.404, p < 0.001; blind adults non-visual sensory < PFC: t ₍₂₉₎ =7.128, p < 0.001; Fig. 2).

The pattern for sighted infants in V1 fell between that of sighted and blind adults. The connectivity matrix of sighted infants (V1, PFC, and non-visual sensory) was equally correlated with blind and sighted adults (infants correlated to blind adults: r = 0.654, p < 0.001; to sighted adults: r = 0.594, p < 0.001; correlation of infants with blind vs. with sighted adults: z = 0.832, p = 0.406; see Supplementary Figure S1 for the connectivity matrices). The difference in connectivity strength between V1 to PFC and V1 to non-visual sensory regions was weaker in sighted infants than in sighted or blind adults (group (sighted adults, infants) by ROI (PFC, non-visual sensory) interaction effect: F_{(1, 523)} = 92.21, p < 0.001; group (blind adults, infants) by ROI (PFC, non-visual sensory) interaction effect: F_{(1, 503)} = 57.444, p < 0.001). V1 of sighted infants showed marginally stronger connectivity to non-visual sensory regions (S1/M1 and A1) than PFC (non-visual sensory regions > PFC, paired t-test, t ₍₄₇₄₎ = 1.95, p = 0.052; Fig.2).

The dHCP cohort included both full-term neonates and preterm infants, scanned at their equivalent gestational age. Visual exposure therefore varied somewhat in duration across infants (from 0 to 19.71 weeks), with slightly longer exposure in preterm babies. This variation did not affect connectivity patterns either in V1 or secondary visual cortices (V1: r = 0.06, p = 0.192; secondary visual: r = 0.004, p = 0.923; see Supplementary Figure S8). A few weeks of vision after birth is therefore insufficient to influence connectivity.

Relative to sighted adults, blind adults showed a stronger dominance of within hemisphere connectivity between secondary and primary visual cortices and PFC. That is, in people born blind, left visual networks are more strongly connected to left PFC networks, whereas right visual networks are more strongly connected to right PFC. By contrast, there is no hemispheric bias in the sighted group (group (blind adults, sighted adults) by lateralization (within hemisphere, between hemisphere) interaction, secondary visual cortices: F_{(1, 78)} = 131.51, p < 0.001; post-hoc Bonferroni-corrected paired: t-test: sighted adults within hemisphere > across hemisphere: t ₍₄₉₎ = 5.778, p < 0.001; blind adults within hemisphere > across hemisphere: t₍₂₉₎ = 10.735, p < 0.001; V1: F_{(1, 78)} = 87.211, p < 0.001; post-hoc Bonferroni-corrected paired: t-test: sighted adults within hemisphere > between hemisphere: t ₍₄₉₎ = 3.251, p = 0.101; blind adults within hemisphere > between hemisphere: t ₍₂₉₎ = 7.019, p < 0.001).

These connectivity results are consistent with previous studies of task-based cross-modal responses in blindness. In blind adults cross-modal responses in occipital cortex and co-lateralize with fronto-parietal networks with related functions (Kanjlia et al., 2021; Lane et al., 2017). For example, language-responsive occipital areas collateralize with language responsive prefrontal areas across individuals (Lane et al., 2017).

The connectivity pattern of sighted infants resembled sighted more than blind adults (Fig. 3), suggesting blindness-driven reorganization. There was a significant difference in laterality between blind adults and sighted infants (group (blind adults, infants) by lateralization (within hemisphere, between hemisphere) interaction effect: F_{(1, 503)} = 303.04, p < 0.001). By contrast, there was no difference between sighted adults and sighted infants (group (sighted adults, infants) by lateralization (within hemisphere, across hemisphere) interaction effect: F_{(1, 523)} = 2.244, p = 0.135; see supplementary results for a detailed group comparison of within and across hemisphere differences). Similar group by laterality interaction pattern are also observed in V1 (group (blind adults, infants) by lateralization (within hemisphere, between hemisphere) interaction effect: F_{(1, 503)} = 123.608, p < 0.001; group (sighted adults, infants) by lateralization (within hemisphere, across hemisphere) interaction effect: F_{(1, 523)} = 2.827, p = 0.093). The incorporation of visual cortices into lateralized functional networks (e.g., language, response selection) in blindness, which is observed in task-based studies, may drive stronger within-hemisphere connectivity in this population (Kanjlia et al., 2021; Lane et al., 2017; Tian et al., 2022).

It is not the case that every occipital area shows equal connectivity to every prefrontal area. On the contrary, in blind adults resting state connectivity patterns are specialized and aligned with the functional specialization observed in task-based data (Bedny et al., 2011; Kanjlia et al., 2016, 2021). For example, language-responsive subregions of occipital cortex show strongest functional connectivity with language-responsive sub-regions of PFC, whereas math-responsive occipital areas show stronger connectivity with math-responsive PFC (Fig. 4) (Bedny et al., 2011; Kanjlia et al., 2016, 2021; Lane et al., 2015). Is this fronto-occipital connectivity specialization present in infancy, potentially driving task-based cross-modal specialization?

We compared connectivity preferences across three prefrontal and three occipital regions previous shown to activate preferentially in language (sentences > math), math (math > sentences) and response-conflict (no-go > go with tones) tasks respectively (Kanjlia et al., 2016, 2021; Lane et al., 2015). For ease of viewing, Fig. 4 shows results from two of the three regions, math and language. However, all statistical analyses included all three areas (See Supplementary Figure S9 for all three regions).

Sighted infants showed a less differentiated fronto-occipital connectivity pattern relative to sighted and blind adults (Group (sighted adults, blind adults, sighted infants) by occipital regions (math, language, response-conflict) by PFC regions (math, language, response-conflict) interaction F_{(8, 2208)} = 16.323, p < 0.001). Unlike in adults, in infants, all the occipital regions showed stronger correlations with math- and response-conflict related prefrontal areas than language-responsive prefrontal areas (Fig. 4 and Supplementary Figure S9). The occipital region that is sensitive to response-conflict in blind adults showed equivalent correlations with math and response-conflict PFC regions in infants. The region of occipital cortex that responds to language and shows the strongest connectivity with language responsive PFC in blind adults, showed stronger connectivity with math and response-conflict PFC areas in infants.

Although fronto-occipital connectivity was not adult-like in infants, biases in infants were somewhat consistent with future differentiation: the preferential correlation with math responsive PFC was strongest in those occipital areas that go on to develop math responses in blind adults (occipital regions (math, language, response-conflict) by PFC regions (math, language, response-conflict) interaction in infants F_{(4, 1896)} = 85.145, p < 0.001, post-hoc Bonferroni-corrected paired t-test see Supplementary Table 1). Although the occipital region that is language-responsive in blind adults showed stronger connectivity to math and response-conflict areas of PFC in infants, this biased against language areas was smaller for this occipital areas than the other two.

Note that findings regarding regional specialization need to be interpreted with caution for two reasons. First, although, prior evidence suggests that at least some of the prefrontal specialization present in adults is already present even in young babies, we do not know whether the language/number/executive function distinctions exist in prefrontal cortices of infants (Raz & Saxe, 2020). Second, these more fine-grained comparisons across occipital/frontal regions are more vulnerable to potential anatomical alignment issues between adult and infant brains. In other words, lack of specialization in infants could reflect the different location of the areas in this population.

Fifty sighted adults and thirty congenitally blind adults contributed the resting state data (sighted: n = 50; 30 females; mean age = 35.33 years, standard deviation (SD) = 14.65; mean years of education = 17.08, SD = 3.1; blind: n = 30; 19 females; mean age = 44.23 years, SD = 16.41; mean years of education = 17.08, SD = 2.11; blind vs. sighted age, t ₍₇₈₎ = 2.512, p < 0.05; blind vs. sighted years of education, t ₍₇₈₎ = 0.05, p = 0.996). Since blind participants were on average older, we also performed analyses in an age-matched subgroups of sighted controls (n = 29) and found similar results to the full sample (see Supplementary Figure S11 to Figure S14). Blind and sighted participants had no known cognitive or neurological disabilities (screened through self-report). All adult anatomical images were read by a board-certified radiologist and no gross neurological abnormalities were found. All the blind participants had at most minimal light perception from birth. Blindness was caused by pathology anterior to the optic chiasm (i.e., not due to brain damage). All participants gave written informed consent under a protocol approved by the Institutional Review Board of Johns Hopkins University.

Neonate data were from the third release of the Developing Human Connectome Project (dHCP) (n = 783) (https://www.developingconnectome.org). Ethical approval was obtained from the UK Health Research Authority (Research Ethics Committee reference number: 14/LO/1169). After quality control procedures (described below), 475 subjects were included in data analysis, with one scan per subject. The average age from birth at scan = 2.79 weeks (SD = 3.77); average gestational age at scan = 41.23 weeks (SD = 1.77). We only included infants who were full-term or scanned at term-equivalent age if preterm, while not being flagged by the dHCP project team as not passing quality control for functional MRI (fMRI) images (n = 634). Infants with more than 160 motion outliers were exclude (n = 116 dropped). Motion-outlier volumes were defined as DVARS (the root mean square intensity difference between successive volumes) higher than 1.5 interquartile range above the 75th centile, after motion and distortion correction. Infants with signal drop-out in regions of interest (ROI) were also excluded (n = 43 dropped). To identify signal dropout, we first averaged blood oxygen level-dependent (BOLD) signal intensity for all time point, for each subject, in each of 100 parcel defined by Schaefer’s atlas (Schaefer et al., 2018). For each ROI (n = 18 ROIs) in the current study, signal dropout was then identified as BOLD intensity lower than −3 standard deviations, where the mean and standard deviations were identified across all 100 cortical parcels. Participants were excluded if any of the ROIs showed a signal dropout.

Resting state data were preprocessed using FSL version 5.0.9 (Smith et al., 2004), DPABI version 6.1 (Yan et al., 2016), FreeSurfer (Dale et al., 1999), and in-house code (https://github.com/NPDL/Resting-state_dHCP). The functional data for all groups were linearly detrended and low-pass filtered (0.08 Hz).

For adults, functional images were registered to the T1-weighted structural images, motion corrected using MCFLIRT (Jenkinson et al., 2002), and temporally high-pass filtering (150 s). No subject had excessive head movement (> 2mm) or rotation (> 2°) at any timepoint. Resting state data are known to include artifacts related to physiological fluctuations such as cardiac pulsations and respiratory-induced modulation of the main magnetic field. A component-based method, CompCor (Behzadi et al., 2007), was therefore used to control for these artifacts. Particularly, following the procedure described in Whitfield-Gabrieli et al., nuisance signals were extracted from 2-voxel eroded masks of spinal fluid (CSF) and white matter (WM), and the first 5 principal components analysis (PCA) components derived from these signals was regressed out from the processed BOLD time series (Whitfield-Gabrieli & Nieto-Castanon, 2012). In addition, a scrubbing procedure was applied to further reduce the effect of motion on functional connectivity measures (Power et al., 2012, 2014). Frames with root mean square intensity difference exceeding 1.5 interquartile range above the 75th centile, after motion and distortion correction, were censored as outliers.

The infants resting state functional data were pre-processed by the dHCP group using the project’s in-house pipeline (Fitzgibbon et al., 2020), This pipeline uses a spatial independent component analysis (ICA) denoising step to minimize artifact due to multi-band artefact, residual head-movement, arteries, sagittal sinus, CSF pulsation. For infants, ICA denoising is preferable to using CSF/WM regressors. Because it is challenging to accurately define anatomical boundaries of CSF/WM due to the low imaging resolution comparing with the brain size and the severe partial-volume effect in the neonate (Fitzgibbon et al., 2020). Like in the adults, frames with root mean square intensity difference exceeding 1.5 interquartile range above the 75th centile, after motion and distortion correction, were considered as motion outliers. Out from the 2300 frames, a subset of continuous 1600 with minimum number of motion outliers was kept for each subject. Motion outliers were censored from the subset of continuous 1600, and a subject was excluded from further analyses when the number of outlier exceeded 160 (10% of the continues subset) (Hu et al., 2022).

For both groups of adult and infants, we performed a temporal low-pass filter (0.08 Hz low-pass cutoff) and a linear detrending. ROI-to-ROI connectivity was calculated using Pearson’s correlation between ROI-averaged BOLD timeseries (ROI definition see below). The All t-tests and F-tests are two-sided. The comparison of correlation coefficients was done using cocor software package and Pearson and Filon’s z (Diedenhofen & Musch, 2015; Pearson & Filon, 1898).

ROIs in the frontal and occipital cortices were defined from separate task-based fMRI experiments with blind and sighted adults (Kanjlia et al., 2016, 2021; Lane et al., 2015). Three separate experiments were conducted with the same group of blind and sighted subjects (sighted n=18; blind n=23). The language ROIs in the occipital and frontal cortices were identified by sentence > nonwords contrast in an auditory language comprehension task (Lane et al., 2015). The math ROIs were identified by math > sentence contrast in an auditory task where participants judged equivalence of pairs of math equations and pairs of sentences (Kanjlia et al., 2016). The executive function ROIs were identified by no-go > frequent go contrast in an auditory go/no-go task with non-verbal sounds (Kanjlia et al., 2021). The occipital ROIs were defined based on group comparisons blind > sighted in a whole-cortex analysis. For example, the occipital language ROI were defined as the cluster that responded more to auditory sentence than auditory nonwords conditions in blind, relative to sighted, in a whole-cortex analysis. All three occipital ROIs were defined in the right hemisphere. Left-hemisphere occipital ROIs were created by flipping the right-hemisphere ROIs to the left hemisphere. The frontal ROIs were defined based on a whole-cortex analysis which combined all blind and sighted adult data. For example, the frontal language ROI was defined as responded more auditory sentence than auditory nonwords conditions across all blind and sighted subjects, constrained to the prefrontal cortex. For frontal ROIs, the language ROI was defined in the left, and the math and executive function ROI were defined in the right hemisphere, then flip to the other hemisphere.

The V1 was defined from a previously published anatomical surface-based atlas (PALS-B12) (Van Essen, 2005). The primary somatosensory and motor cortex (S1/M1) ROI was selected as the area that responds more to the go than no-go trials in the auditory go/no-go task across both blind and sighted groups, constrained to the hand area in S1/M1 search-space from neurosynth.org (term “hand movements”) (Kanjlia et al., 2021). The primary auditory cortex (A1) ROI was defined as the transverse temporal portion of a gyral-based atlas (Desikan et al., 2006; Morosan et al., 2001).

All the ROIs were defined in standard space. For infants, the ROIs were subsequently transformed into each subject’s native space using a two-step approach. First, the ROIs were converted from the adult’s MNI space into the 40-week dHCP template (Bozek et al., 2018). ANTS, previously shown to be effective in pediatric studies (Avants et al., 2014; Cabral et al., 2022; Jain et al., 2012; Lawson et al., 2013), was utilized to estimate the deformation field between these two spaces. In this step, the infant’s scalp and cerebellum were masked, as these structures in the infant brain greatly differ from those in the adult and can introduce bias into the registration process, as outlined in a study by Cabral et al (Cabral et al., 2022). Secondly, the ROIs were further transformed from the 40-week space into individual’s native spaces, employing the deformation field provided by the dHCP group. Nearest neighbor interpolation was applied in both steps (see examples of ROI in individual infant brains in the Supplementary Figure S15). For adults, the ROIs were also transformed into each subject’s native space, employing the deformation field estimated by FreeSurfer. For both adults and infants, any overlapping voxels between ROIs were removed and not counted toward any ROIs.

Neonate data were from the second and third release of the Developing Human Connectome Project (https://www.developingconnectome.org). The de-identified blind and sighted adults’ data will be posted on Vivli.org upon publication of the current manuscript.