Behavior rating on shape similarity and conceptual association

In this study, we selected 21 Italian words, which referred to 21 manmade objects, as our stimuli. The selection was mostly based on behavior ratings of object properties from an independent group of sighted participants who did not take part in the fMRI experiment (N = 19; see Stimuli in Materials and methods about the stimulus selection procedure and criteria). To validate the rating results from the stimulus selection stage and to verify whether the early blind population had a similar shape and conceptual knowledge as the sighted control, all participants who took part in the fMRI experiments (N = 48) also rated the object properties of the stimuli selected (Fig 1). These participants consisted of three groups: 16 early blind (EB) participants, 16 gender- and age-matched sighted control (SC) participants, and 16 independent sighted (IS) participants (see Participants in Materials and methods for details).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Stimulus information.

(A) Correlation between ratings on pairwise shape similarity and pairwise conceptual association across three participant groups (EB: early blind, SC: sighted control, IS: independent sighted). (B) Correlation among linguistic variables and ratings on other object properties across three participant groups. (C) Pairwise ratings on shape similarity (i.e., the shape similarity RDM). (D) Pairwise ratings on conceptual association (the conceptual association RDM). (E) Linguistic variables and ratings on other object properties. (F) Correlations between the first five RCs and linguistic variables and ratings on other object properties. The underlying data for this figure can be found in S1 Data. RC, rotated component; RDM, representational dissimilarity matrix.

https://doi.org/10.1371/journal.pbio.3001930.g001

Shape similarity and conceptual association were rated on a 7-point Likert scale in a pairwise manner (see Procedures in Material and methods about the rating procedure). We assessed the inter-rater reliability within each group of participants using the intraclass correlation based on a mean-rating, consistency, two-way random model (i.e., ICC(C,k)) [36]. Both shape rating (ICC(C,k): 0.953–0.973) and conceptual rating (ICC(C,k): 0.984–0.985) showed “excellent” inter-rater reliability [37] (S3 Table). We averaged the rating scores within each group and compared them across groups. Fig 1A illustrates that the rating scores on both object properties were highly reliable across three groups (r(208) on shape similarity: 0.957–0.983; on conceptual association: 0.982–0.984), and the pairwise shape similarity was orthogonal to the pairwise conceptual association (r(208): 0.103–0.132).

We then averaged the pairwise rating scores of all the participants (N = 48) and calculated the representational dissimilarity matrix (RDM) of shape similarity and conceptual association (i.e., 7 minus the mean rating score). The resulting two model RDMs had comparable variance across pairs of objects (shape similarity: variance = 2.163; conceptual association: variance = 2.498) and therefore offered equated discovery possibilities when correlated with brain RDMs in the subsequent RSA. Fig 1C and 1D show the organizational structure of the two RDMs, where 21 items were grouped according to the clusters generated by the k-means clustering algorithm [38,39], with the silhouette criterion used to decide the optimal number of clusters [40]. The shape similarity RDM fell into three clusters, corresponding to square, round, and elongated objects (Fig 1C). The conceptual association RDM fell into seven smaller clusters, corresponding to different occasions in which objects were used (Fig 1D). For example, the two biggest clusters were related to eating and writing. The conceptual rating results accorded closely with the teleological perspective, which suggests the essence of a manmade object lies in its function, not its physical properties (e.g., [41]).

Behavior rating on other object properties and confounding factors

Potential confounding factors were also considered. It has been reported that other properties of manmade objects can also modulate brain activity, like object size (big versus small; e.g., [42]), toolness (tools versus non-tool manmade objects; e.g., [43]), and contextual association (strong versus weak contextual association objects; e.g., [44]). These three variables were rated on a 7-point Likert scale (see Procedures in Material and methods about the rating procedure). S3 Table shows the inter-rater reliability within each group of participants. The inter-rater reliability reached “excellent” on object size (ICC(C,k): 0.979–0.992) and varied from “good” to “excellent” on toolness (ICC(C,k): 0.893–0.928). The inter-rater reliability on contextual association differed between sighted and early blind groups. While sighted groups had a “good” to “excellent” inter-rater reliability (SC: ICC(C,k) = 0.856; SI: ICC(C,k) = 0.919), the early blind group only had a “moderate” one (EB: ICC(C,k) = 0.613). Such heterogeneity in the early blind might result from a lack of instantaneous and global information about the environment from the visual input.

Besides the three object properties, all participants rated on a 7-point Likert scale about the degree to which they knew each object’s typical shape and primary function. Since most stimuli selected were everyday objects, both shape and conceptual rating scores hit the ceiling and varied only slightly across objects (averaged shape familiarity score across objects: M = 6.744, SD = 0.285; averaged conceptual familiarity score across objects: M = 6.944, SD = 0.066). Participants also rated how frequently they touched each object (1: have never touched it before; 7: touch it every day), which can be considered a sensitive and common index reflecting object familiarity across sighted and early blind groups. The inter-rater reliability on touch experience within each group of participants reached “excellent” (ICC(C,k): 0.965–0.975; S3 Table).

We averaged the above rating scores within each group of participants and evaluated the reliability of the mean rating score across participant groups. Fig 1B shows that the rating scores across three groups of participants were reliable (r(19) on objects size: 0.973–0.998; on contextual association: 0.732–0.940; on toolness: 0.883–0.933; on touch experience: 0.935–0.974). From this figure, we can also spot a moderate positive correlation between object size and contextual association (r(19): 0.363–0.529) and between toolness and contextual association (r(19): 0.264–0.622), which means the bigger the size, or the more likely an object is a tool, the more likely this object is bound to a specific context. Moreover, we also added two linguistic measures—word frequency (i.e., the Zipf value of the word occurrence in film and television subtitles; http://crr.ugent.be/subtlex-it/) and word duration. There was a moderate positive correlation between word frequency and touch experience (r(19): 0.419–0.446) and a moderate negative correlation between word frequency and word duration (r(19) = −0.577).

We then averaged the rating scores across all participants (N = 48) to get a mean rating score vector for each rating item. Fig 1E illustrates the Z-scores of all the ratings across objects. To orthogonalize these unidimensional variables, we conducted the principal component analysis and applied varimax rotation to improve the interpretability of the resulting principal components. Five components had eigenvalues greater than 1. Fig 1F shows the correlation of these five rotated components (RCs) with each rating item. RC1 to RC5 corresponded to object size, toolness, touch experience, word frequency, and word duration, respectively (r(19): 0.915–0.981). The RCs corresponding to object size and toolness also had moderate correlations with the contextual association (r(19): 0.656 and 0.584). These RC scores were used in the subsequent parametric modulation analysis.

Performance on shape and conceptual tasks during scanning

During the scanning, participants performed two tasks on the same set of auditorily presented words. In the shape verification task, participants thought carefully about the typical shape of each object and judged whether it was elongated, angular, hollow, circular, and disc-shaped. In the conceptual verification task, participants thought carefully about the primary function of each object and judged whether it was for eating, writing, sleeping, lighting, and purchasing (see S1 Fig and Procedures in Materials and methods for details).

Table 1 shows the accuracy and reaction time (RT) across participants within each group in shape and conceptual verification tasks. All groups of participants had near-ceiling accuracy on both tasks. The shape verification task took about 130 to 200 ms longer than the conceptual verification task. We built a linear mixed model to predict the RT in the correct trials with groups of participants (EB versus SC) and types of tasks (shape versus conceptual tasks) as fixed effects variables and each participant as random effects grouping factors. The analysis revealed a significant task effect (F(1,30) = 73.055; p < 0.001), whereas the group effect was found insignificant (F(1,30) = 0.732, p = 0.399), and the interaction effect between groups and tasks had only a slight trend toward significance (F(1,30) = 2.552, p = 0.123). The significant difference between shape and conceptual tasks aligns with the evidence suggesting that retrieving specific semantic features (e.g., shape knowledge) requires more time than general semantic knowledge (i.e., function knowledge [45]). The interaction effect showed a weak trend that the shape task was slightly more difficult than the conceptual task for the EB than the SC ([(EB > SC) × (shape > conceptual tasks)]; z = 1.588, p = 0.112), which might be due to a lack of visual experience.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Accuracy and RT during fMRI scanning.

https://doi.org/10.1371/journal.pbio.3001930.t001

Shape compared to conceptual tasks engaged ILOTC in both EB and SC

We first contrasted the neural activity level between the shape and conceptual tasks. To remove the domain-general RT effect, we modeled the trial-by-trial RT variability across the two tasks in the first-level general linear model (GLM) using both the variable epoch approach and the variable impulse approach [46]. Fig 2 illustrates results while the domain-general RT effect was controlled (vertex-wise p < 0.001, cluster-level family-wise error (FWE) corrected p < 0.05).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Specific brain activation in shape and conceptual tasks (vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05).

Dots in colors denote the location of LOtv in three representative studies. (A) Shape versus conceptual tasks across all the participants. (B) Shape versus conceptual tasks in the EB. (C) Shape versus conceptual tasks in the SC. (D) Interaction between groups (EB vs. SC) and tasks (shape vs. conceptual). The error bars indicate the standard error. **: p < 0.01. The underlying data for this figure can be found in S1 Data. EB, early blind; FWE, family-wise error; LOtv, lateral occipital tactile-visual; SC, sighted control.

https://doi.org/10.1371/journal.pbio.3001930.g002

Fig 2A shows the contrast between the shape task and the conceptual task using all participants (N = 48). The shape task and the conceptual task involved dissociable brain networks. The shape task activated bilateral brain areas, including the ILOTC (i.e., the lateral part of the Broadman area (BA) 37), the aIPS, the posterior IPS (pIPS), the vPMC, and the inferior frontal sulcus. To verify whether the ILOTC activated in the shape task was the same region as the LOtv reported in previous literature, we projected the peak coordinates of the LOtv from three representative studies (i.e., [12,13,47]) to the brain surface and found that these coordinates largely fell over the geometric gravity center of the ILOTC region. In Fig 2A, we can identify two activity epicenters in the IPS—one was anterior and the other was posterior and joined to the intraoccipital sulcus.

The conceptual task mainly activated left-lateralized brain areas, including the anterior part of the lateral temporal lobe (aLTC), the superior temporal gyrus (STG; BA 22), the angular gyrus (AG; BA 39), and the supramarginal gyrus (SMG; BA 40). These regions were in accord with the high-level linguistic network [48–50] and are considered to underly language-supported conceptual processing [51–53].

We then looked at the brain activation in EB and SC separately (Fig 2B and 2C). Both EB and SC had ILOTC activation in the shape task compared to the conceptual task. To confirm that the regions in the ILOTC found in the two groups were the same, we calculated the overlap coefficient, i.e., the area of the intersection region divided by the smaller area of the two regions. The overlap coefficient of the left ILOTC was 100%, i.e., EB’s ILOTC fell within SC’s ILOTC. The overlap coefficient of the right ILOTC was 81.5%. Consistent with the results pooling all participants (N = 48), SC also had significant activation in bilateral aIPS, pIPS, and vPMC in the contrast between shape and conceptual tasks (Fig 2C). Although these regions did not survive the multiple comparison correction at the whole-brain level in EB (Fig 2B), analyses using the significant areas in SC as regions of interest (ROIs) showed bilateral aIPS, bilateral pIPS, and the left vPMC in the EB also showed greater activation in the shape task than in the conceptual task (S2 Fig; left aIPS: t(15) = 3.486, p = 0.003; right aIPS: t(15) = 2.487, p = 0.025; left pIPS: t(15) = 2.478, p = 0.026; right pIPS: t(15) = 3.357, p = 0.004; left vPMC: t(15) = 2.632, p = 0.019; right vPMC: t(15) = 1.861, p = 0.083).

Both EB and SC activated the language network in the conceptual task. However, EB exhibited reduced left lateralization than SC. To measure the extent of lateralization, we extracted the T scores of the top 5% percentage of vertices showing the strongest activation in the contrast between the conceptual task and the shape task within the language network, which was anatomically defined in each participant’s native space by combing bilateral STG, bilateral inferior parietal cortices (i.e., the AG), and bilateral SMG in the DKT atlas [54]. The left lateralization was measured as (L − R)/(L + R), where L and R were the sums of T scores in the left and right hemispheres. While the SC had clear left lateralization (M = 0.381, SD = 0.344, t(15) = 4.440, p < 0.001), the EB’s lateralization was not evident (M = 0.143, SD = 3.397, t(15) = 1.442, p = 0.397). The paired t test showed a significant difference between the SC and the EB (paired t(15) = 2.452, p = 0.027), while no significant difference was found in handedness scores (SC: M = 76.875, SD = 20.238; EB: M = 73.750, SD = 16.279; paired t(15) = 0.543, p = 0.595). The reduced left lateralization for language processing in EB has been reported in a recent study and is still open to interpretation [55].

Next, we directly contrasted the neural activity between EB and SC. As a sanity check, we first compared the brain activity level in shape and conceptual tasks to the resting state between EB and SC. As both tasks included auditory input, the occipital cortex in EB should show enhanced activation due to cross-modal neuroplasticity (e.g., [56–59]), and the results showed up as expected (S3 Fig; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). We then compared the activity level between shape and conceptual tasks between EB and SC. We found only one significant region in the left cuneus near the parieto-occipital sulcus (Fig 2D; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). ROI analysis showed that this region in EB had greater activation in the conceptual task than in the shape task (t(15) = −3.447, p = 0.004), whereas, in SC, it showed an opposite pattern (t(15) = 3.213; p = 0.006). This finding suggests that the earlier “visual” cortex in EB (i.e., the left cuneus) might repurpose itself to a similar role as what the language network played in the conceptual task (see also a recent meta-analysis [60]).

S4A Fig illustrates the RT effect across the two tasks (N = 48; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). As expected, it involved both frontoparietal and cingulo-opercular networks underlying top-down control [61]. It also involved regions in the default mode network, which could be because both the shape and the conceptual tasks require mental simulation [62]. Intriguingly, contrasting the RT effect between EB and SC revealed the lateral and ventral parts of the occipital cortex (S4B Fig; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). These regions substantially overlapped with the LOC involved in visual shape perception in the sighted population, suggesting a functional reorganization of these regions in EB. Note that these regions did not overlap with the ILOTC.

Other object properties did not modulate ILOTC activity

To investigate whether the other object properties modulated brain activity in the ILOTC, we conducted a parametric modulation analysis. The set of the parametric modulators included the task type (i.e., the shape task coded as 1 and the conceptual task coded as −1), the z-scores of the RT across all the trials in each run, the RCs corresponding to object size, toolness, touch experience, word duration, and word frequency. Fig 3 presents the significant brain areas encoding these parametric modulators (N = 48; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Neural correlates of task types and other object properties (vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05).

Dots in colors denote the location of LOtv in three representative studies. (A) Neural correlates of task types (the shape task coded as 1 and the conceptual task coded as −1). (B) Neural correlates of object size. Activations in the significant brain areas positively correlated with object size, i.e., larger objects induced higher activation. (C) Neural correlates of touch experience. Activations in the significant brain area negatively correlated with touch experience, i.e., less-touched objects induced higher activation. The underlying data for this figure can be found in S1 Data. FWE, family-wise error; LOtv, lateral occipital tactile-visual.

https://doi.org/10.1371/journal.pbio.3001930.g003

When potential confounding factors were modeled, the difference between the two task types was still preserved (Fig 3A): The shape task activated bilateral brain areas, including the ILOTC, the aIPS, the pIPS, and the vPMC. The conceptual tasks mainly activated brain areas in the left hemisphere, including the orbital part of the inferior frontal gyrus (IFG) (i.e., BA 47), the aLTC, the posterior part of the STG (pSTG), the SMG, and the AG. These regions neatly matched the language network [48,49] with the absence of the triangular part of the IFG and the 55b region in the premotor cortex [63] (see the overlap in S5 Fig), in line with previous studies suggesting these two dorsal regions play a non-semantic role in language processing (e.g., [64–66]). Since the brain clusters in Fig 3A were more discrete than those reported in the univariate contrast reported in Fig 2 (with no control for alternative object properties), we used the significant regions in Fig 3A to define the ROIs in the following analyses. No regions showed significant differences between EB and SC. We also found the same region in the cuneus when directly comparing EB and SC under a lower threshold (vertex-wise p < 0.001, uncorrected).

Fig 3B and 3C shows the brain areas sensitive to the other object properties. The object size was mainly localized to the three scene-selective regions—the transverse occipital sulcus, the parahippocampal place area, and the retrosplenial cortex (Fig 3B). It has already been reported that these areas also prefer large nonmanipulable objects (e.g., [42,67]) and objects with a strong contextual association (e.g., [44,68]). Since the object size component here had a moderate correlation with the rating scores on contextual association (Fig 1F), we cannot distinguish between these two factors in this study. Moreover, we found a region in the left ventral and medial temporal cortex (mainly in the BA 20), of which the level of activity negatively correlated to touch experience (Fig 3C), suggesting this region was sensitive to the novelty of objects. We did not find any brain areas significantly modulated by toolness, which might result from the lack of typical tools (e.g., hammers or scissors) in the stimuli. Directly comparing the effects of all these parametric modulators between EB and SC also failed to reveal any significant brain regions.

S6 Fig illustrates the effect of the two linguistic variables (N = 48; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). Word duration was localized to bilateral auditory cortices and bilateral STG. Word frequency was mainly localized to the right-lateralized ventral attention network and the salience network, characterized by their sensitivity to salient stimuli (e.g., [69,70]).

ILOTC represented shape similarity, not conceptual association in both EB and SC

We then used RSA to investigate whether the ILOTC identified implemented shape representation (Fig 4, left and right panels corresponding to left and right ILOTC). A three-way mixed ANOVA was first performed between groups (EB versus SC), tasks (shape versus conceptual tasks), and representations (shape similarity versus conceptual association). The groups factor was between-subject, whereas tasks and representations were within-subject factors. In bilateral ILOTC, we only found a significant effect in representations and a significant interaction between tasks and representations (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Neural representations of bilateral ILOTC.

The left column showed the neural representation in the left ILOTC. The right column showed the neural representation of the right ILOTC. (A) The RSA results across all participants (N = 48). (B) The RSA results in the EB (N = 16). (C) Inter-subject correlation between brain RDMs within and between the EB and the SC. (D) The MDS visualization of the mean brain RDM of the ILOTC across all participants (N = 48). ns: not significant, *: p < 0.05, **: p < 0.01, ***: p < 0.001. The underlying data for this figure can be found in S1 Data. EB, early blind; ILOTC, inferolateral occipitotemporal cortex; RDM, representational dissimilarity matrix; RSA, representational similarity analysis; SC, sighted control.

https://doi.org/10.1371/journal.pbio.3001930.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Neural representation in bilateral ILOTC.

https://doi.org/10.1371/journal.pbio.3001930.t002

Fig 4A illustrates the RSA results in bilateral ILOTC across all participants (N = 48). Bilateral ILOTC represented shape similarity in both the shape task (left ILOTC: t(47) = 10.367, p < 0.001; right ILOTC: t(47) = 7.705, p < 0.001) and the conceptual task (left ILOTC: t(47 = 4.066), p < 0.001; right ILOTC: t(47) = 3.209, p = 0.002). The shape representation was stronger in the shape task than in the conceptual task (left ILOTC: paired t(47) = 5.183, p < 0.001; right ILOTC: paired t(47) = 3.776, p < 0.001). We found no clear evidence that bilateral ILOTC represented the conceptual association in either the shape or the conceptual tasks—only the conceptual effect in the left ILOTC in the conceptual task was marginally significant (t(47) = 2.123, p = 0.039). No significant difference was found in conceptual representation between shape and conceptual tasks (left ILOTC: paired t(47) = 0.558, p = 0.580; right ILOTC: paired t(47) = 0.395, p = 0.695).

Fig 4B highlighted that the population without visual experience (i.e., the EB) showed a largely similar pattern. Bilateral ILOTC represented shape similarity in the shape task (left ILOTC: paired t(15) = 4.568, p < 0.001; right ILOTC: paired t(15) = 3.610, p = 0.003), whereas their shape representation in the conceptual task was less evident (left ILOTC: paired t(15) = 1.220, p = 0.241; right ILOTC: paired t(15) = 1.852, p = 0.084). The paired t test revealed a significant difference between the two tasks in the left ILOTC (paired t(15) = 3.361, p = 0.004) but not in the right ILOTC (paired t(15) = 1.466, p = 0.163). No evidence supported bilateral ILOTC represented conceptual association in either shape or conceptual tasks (t(15) < 1.282, ps > 0.219).

We also investigated whether bilateral ILOTC in EB and SC share a matched shape representation (Fig 4C). By doing so, we measured the within-group coherence—the correlation between each participant’s neural RDM and the mean neural RDM of the other participants within the same group (i.e., EB-EB and SC-SC) and the between-group coherence—the correlation between each participant’s neural RDM and the mean neural RDM of all the other participants in the other group (i.e., EB-SC). A two-way ANOVA was performed between tasks (shape versus conceptual tasks) and group pairs (EB-EB versus SC-SC versus EB-SC). No significant interaction was found between tasks and group pairs (left ILOTC: F(2,90) = 1.366, p = 0.260; right ILOTC: F(2, 90) = 1.446, p = 0.241). There is a significant difference between tasks (left ILOTC: F(1,90) = 90.743, p < 0.001; right ILOTC: F(1, 90) = 75.809, p < 0.001), suggesting the shape task induced more coherent representations in bilateral ILOTC across participants. A weak effect in group pairs was also spotted in the left ILOTC (F(2,90) = 4.746, p = 0.011) but not in the right one (F(2, 90) = 1.065, p = 0.349). The post hoc comparison found that the mean value across levels of tasks was significantly different between SC-SC and EB-EB in the left ILOTC (Tukey’s test: p = 0.011), suggesting that the neural representation in the left ILOTC was more homogeneous in the SC group than in the EB group. However, there was no significant difference between EB-EB and EB-SC (Tukey’s test: p = 0.742) or between SC-SC and EB-SC (Tukey’s test: p = 0.073), suggesting no significant evidence showing a boundary effect between the neural representations across groups.

We averaged the neural RDMs of bilateral ILOTC across all participants (N = 48) and provided a planar visualization of the representational pattern using multidimensional scaling (Fig 4D). The color of words denoted the three clusters in the model RDM of shape similarity, mainly corresponding to elongated, round, and square objects. Representations of the three shape categories were separated in bilateral ILOTC.

We also investigated the multivariate object representation in other regions showing an enhanced univariate response to the shape task than the conceptual task. S4 to S6 Tables show the three-way mixed ANOVA results between groups (EB versus SC), tasks (shape versus conceptual task), and representations (shape similarity versus conceptual association) in bilateral aIPS, bilateral pIPS, and bilateral vPMC, respectively. They all had the same pattern, with a significant effect in representations and a significant interaction between tasks and representations. S7 Fig shows that all these regions represented shape similarity in the shape tasks (t(47): 5.531–10.074, ps < 0.001). Bilateral aIPS and pIPS also represented shape similarity in the conceptual tasks (t(47): 2.216–2.902, ps: 0.032–0.006), whereas shape representation in bilateral vPMC was not evident in the conceptual task (left: t(47) = 1.875, p = 0.067; right: (47) = 1.677, p = 0.100). Shape representation was more apparent in the shape task than in the conceptual task in all these regions (right vPMC: paired t(47) = 2.602, p = 0.012; other regions: paired t(47): 0.408–5.055, ps < = 0.001).

S8A Fig illustrated the whole-brain searchlight results of shape similarity in the shape tasks across all participants (N = 48) (vertex-wise FWE corrected p < 0.005, cluster size > 400 mm²). The ILOTC was one of the epicenters showing the strongest shape effect. Direct contrast between EB and SC revealed a region in the right lateral occipital cortex showing a stronger shape representation in the EB than SC (S8B Fig; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05).

Conceptual representation in the brain

We also used the RSA to investigate whether the brain areas sensitive to the conceptual task in the univariate analyses represented multivariate conceptual association (Fig 5). Interestingly, although all these regions showed significantly stronger univariate activation in the conceptual task than in the shape task, only the left AG represented the conceptual association in the conceptual task across all the participants (orbital IFG: t(47) = 2.395, p = 0.021; aLTC: t(47) = −1.268, p = 0.211; pSTG: t(47) = −0.621, p = 0.537; AG: t(47) = 3.337, p = 0.002, SMG: t(47) = 1.174, p = 0.246; only the AG survived from multiple comparison correction, as Bonferroni corrected p < 0.05 for five ROIs is p < 0.01). The conceptual representation in the left AG was more evident in the conceptual task than in the shape task (paired t(47) = 2.163, p = 0.036), and no group differences were found between EB and SC (F(1, 30) = 0.192, p = 0.664).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. RSA results of conceptual association in the brain areas with greater activation in the conceptual task than in the shape task across all participants (N = 48).

(A) Brain areas with significantly greater activation in the conceptual task than in the shape task defined in Fig 2A. (B) RSA results of these conceptual-relevant areas in shape and conceptual tasks. *: p < 0.05, **: p < 0.01. The underlying data for this figure can be found in S1 Data. RSA, representational similarity analysis.

https://doi.org/10.1371/journal.pbio.3001930.g005

S9 Fig illustrates the whole-brain searchlight results of conceptual association in the conceptual tasks across all participants (N = 48) (vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). The effects were mainly on bilateral dorsal AG, the left pIPS, the left precuneus, and the left dorsal medial prefrontal cortex. Given that some of the regions could also be spotted in the shape effect in the shape task (S8A Fig), they were likely to be driven by the task context [71].

Shape and conceptual brain network in both EB and SC

We last used the seed-based RSFC to trace the regions having the neural activity synchronized with bilateral ILOTC (left ILOTC: Fig 6A; right ILOTC: Fig 6B; vertex-wise FWE corrected p < 0.005, cluster size > 400 mm²). The ILOTC had strong RSFC to the other bilateral regions sensitive to the shape task—the aIPS, the pIPS, and the vPMC in both EB and SC. The left ILOTC in EB had stronger connectivity to the “visual” cortex than in SC (S10 Fig; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Shape and conceptual brain network.

The left panel shows the RSFC results in the EB, and the right panel shows the RSFC results in the SC. (A, B) The significant seed-based RSFC results in the left ILOTC (A) and the right ILOTC (B) (vertex-wise FWE corrected p < 0.005, cluster size > 400 mm²). Dots in colors denote the location of LOtv in three representative studies. (C) The mean RSFC matrix across participants in EB and SC among the shape- and conceptual-relevant brain areas. (D) Comparison among the mean RSFC among the shape-relevant regions (“Within Shape”), among the conceptual-relevant regions (“Within Conceptual”), and between the shape- and the conceptual-relevant regions (“Between Modules”). ***: p < 0.001. The underlying data for this figure can be found in S1 Data. EB, early blind; FWE, family-wise error; ILOTC, inferolateral occipitotemporal cortex; LOtv, lateral occipital tactile-visual; RSFC, resting-state functional connectivity; SC, sighted control.

https://doi.org/10.1371/journal.pbio.3001930.g006

Fig 6C illustrates the mean RSFC matrix across participants in EB and SC among the regions showing stronger activation in the shape tasks or in the conceptual task. It shows that the brain areas sensitive to the shape task and those sensitive to the conceptual task belonged to separate network modules in both EB and SC. Fig 6D further compares the mean RSFC across all the pairs among the shape-sensitive regions, among the conceptual-sensitive regions, and between the shape- and the conceptual-sensitive regions. In both EB and SC, the mean RSFC within the shape module (EB: paired t(15) = 10.650, p < 0.001; SC: paired t(15) = 9.563, p < 0.001) and within the conceptual module (EB: paired t(15) = 10.024, p < 0.001; SC: paired t(15) = 8.014, p < 0.001) were significantly stronger than the mean RSFC between the two network modules.

Ethics statement

This study was conducted according to the principles expressed in the Declaration of Helsinki. The ethical committee of the University of Trento approved the experimental protocol in this study (protocol 2014–007). All participants provided written informed consent and were paid for their time.

Participants

Forty-eight native Italian speakers with no history of neurobiological or psychiatric disorders participated in the fMRI experiment. Thirty-two participants were sighted and 16 participants were early blind. Further recruitment of blind participants was stalled due to COVID-19 restrictions. The early blind (EB) group reported, at most, faint light perception and had no visual memories (10 females; age: M = 32.8, SD = 4.5; all right-handed). To match the demographic information of the early blind group, we divided the sighted participants into two groups. Sixteen formed the sighted control (SC) group, matching the early blind in pairs on gender and age (10 females; age: M = 32.5, SD = 5.9; all right-handed). There was no significant difference between the early blind and the sighted control in head motion measured by the mean framewise displacement index [106] (EB: M = 0.20 mm, SD = 0.05 mm; SC: M = 0.17 mm, SD = 0.06 mm; t(30) = 1.79, p = 0.083). The other 16 formed the independent sighted (IS) group (7 females; age: M = 28.3, SD = 8.1; 2 left-handed). We investigated the group-general effect by pooling EB, SC, and IS together to increase the statistical power and provide the most stable results. We investigated the between-group difference by contrasting EB and its matched SC.

S1 Table shows the demographic information of the early blind and their matched sighted control. In each matched pair, the gender was the same, and the age difference was no more than 3 years. All blind participants were blind since birth except for three participants, who also had visual trouble since birth but fully lost their vision at 8 months, 2 years, and 4 years. These participants’ data did not differ from those of the other blind participants.

Stimuli

To disentangle shape and conceptual representation, we aimed to select a set of words referring to objects, among which the pairwise shape similarity was orthogonal to the pairwise conceptual association.

Here, we differentiated two types of conceptual relations. One assumes concepts are componential, consisting of a set of shared semantic features (e.g., shape, action, motion, and emotion); similarity across semantic features leads to taxonomic relations or categories (e.g., forks and plates are manmade objects, not animals). The other assumes concepts are holistic; complementary roles within the same scenario lead to thematic relations (e.g., folks and plates relate to eating, not sleeping). This study focused on thematic relations by confining its stimuli to one taxonomic category—manmade objects, based on the following considerations: (1) Growing evidence suggests taxonomic and thematic relations rely on dissociable neural systems (e.g., [95,96,107]). Confusing two quality-different conceptual relations into one unified conceptual RDM might be problematic. (2) It is challenging for a neuroimaging study to distinguish whether a brain area represents taxonomic knowledge per se or shows preferences for specific taxonomic categories. Such distinction matters. The former assumes a dedicated brain area representing taxonomic relations among concepts, whereas the latter could mean a brain area representing category-specific features at the pre-conceptual stage. The category-specific features could be the distinguishing features across brain systems. For example, as mainly manmade objects have manipulation-related features and humans have social features, brain systems processing action or social information will exhibit taxonomic preference. The category-specific features could also be within one brain system. For example, different patches along the visual pathway show category-specific effects (e.g., [108,109]) serving the perceptual purpose (e.g., [110]). However, in both cases, we can hardly say these brain systems represent taxonomic relations among concepts. Taxonomic representation at the conceptual level is assumed to emerge from the converge zones when multiple features have already been bound onto a concept (e.g., [95,111]). (3) Some hypothesis argues that category-specific representations at the conceptual level do exist; however, they cannot be represented in local brain areas but emerge from the connectivity among distributed categorical-specific regions across different brain systems [112]. This hypothesis explains why category-specific semantic deficits are well documented in neuropsychological literature (e.g., [113,114]) but are difficult to localize in the brain. However, category-specific representation at the conceptual level does not equal taxonomic knowledge representation, and the proposed connectivity-based neural representations are beyond the scope of this study. (4) Compared to the sighted, the early blind lack perceptual experience with many concepts in the natural world and have different neural representations of these “imperceptible” concepts [115]. Using only manmade objects ensures a relatively fair comparison between the sighted and the early blind (see ratings on touch experience in Fig 1B).

As a starting point, we preselected a set of Italian words referring to 60 everyday manmade objects based on our subjective impressions so that, among these objects, the shape similarity did not always correlate to the conceptual association. For example, a plate (“piatto”) is perceptually similar to a coin (“moneta”) but conceptually relates to a fork (“forchetta”).

Next, we recruited 19 sighted native Italian speakers (age: M = 25.4, SD = 3.6) who did not participate in the fMRI experiments to rate the shape similarity and the conceptual association among the 60 objects. As pairwise rating among numerous items is time-consuming (60 objects require 1,770 pairs of comparison), we adopted the multi-arrangement method [116]. By doing so, participants arranged Italian words on a computer screen by mouse drag-and-drop operations in two 45-min task sessions. The closeness among the words was required to reflect shape similarity in the shape task session and conceptual association in the conceptual task session. Participants were instructed to disregard other object properties like color and size. The pairwise dissimilarity matrix of shape and conceptual information was estimated as the weighted mean of the scale-adjusted on-screen distances from individual arrangements. We averaged the ratings across participants and obtained a mean pairwise dissimilarity matrix for shape and conceptual information, respectively.

Then, these participants rated the potential confounding factors, i.e., object size (big versus small), toolness (tools versus non-tool manmade objects), and contextual association (strong versus weak contextual association objects). Participants were instructed to rate these three unidimensional variables by sliding a horizontal slider from left to right on a computer screen. To assess the variance in familiarity across objects, participants also rated each object on a 7-point Likert scale about the degree to which they knew its typical shape and primary function (1: do not know it at all; 7: know it very well). We also conducted a telephone interview with 16 early blind participants (8 females; age: M = 33.0, SD = 6.6; 6 of the participants took part in the fMRI experiment). We let them perform the same shape and conceptual familiarity rating tasks and asked them whether they had ever touched the objects. We averaged the rating score across participants to obtain a mean rating score for each object and each rating task.

After that, we selected 21 from the 60 Italian words based on the above ratings. This set of words met the following criteria: (1) Both sighted and early blind participants knew each object’s typical shape and primary function. The shape and the conceptual familiarity rating scores were higher than 5.8 (7-point Likert scale) in both groups. (2) Most early blind participants we interviewed (i.e., at least 14 among 16 participants) had touched the objects. (3) Shape similarity and conceptual association were orthogonal across pairs of objects. The absolute value of Spearman’s correlation coefficient was 0.039. (4) Both shape similarity and conceptual association were orthogonal to the potential confounding factors, including both shape and conceptual familiarity from both early blind and sighted participants, word length (i.e., number of letters), word frequency (i.e., the Zipf value of the word occurrence in film and television subtitles; http://crr.ugent.be/subtlex-it/), object size, toolness, and contextual association. Since all these confounding factors were unidimensional, we measured the pairwise dissimilarity of these variables as the absolute difference between each pair of objects and correlated it to the shape and the conceptual information, respectively. The absolute values of Spearman’s correlation coefficients were all below 0.15. (5) The variances across pairwise shape similarity (variance = 0.54) and pairwise conceptual association (variance = 0.53) were maximized while kept comparable. (6) Each object had at least one shape-matched item and one conceptual-associated item. S2 Table shows the 21 Italian words and their English translation.

Finally, a professional narrator recorded his pronunciation of these 21 words. We cut out the silence period at the beginning and the end of each auditory word with the same threshold and equalized the average intensity of all the auditory words as 70 dB using Praat 6.1.01 (https://www.fon.hum.uva.nl/praat/).

Procedures

Before the fMRI scanning, all participants rated each object on a 7-point Likert scale about the degree to which they knew its typical shape and primary function (1: do not know it at all; 7: know it very well). They also rated how frequently they touched each object (1: have never touched it before; 7: touch it every day). We then explained the items of which either shape or conceptual familiarity rating score was below 6 points to ensure that all participants knew each object’s typical shape and primary function. S1 Text shows the survey questions of these ratings.

During the fMRI scanning, we presented audio stimuli using Psychotoolbox-3 (http://psychtoolbox.org/). The sound was delivered through in-ear headphones. Before the formal scanning, we adjusted the volume for each participant so that they could hear the pronunciation clearly under the scanning noise but did not feel too loud. To ensure both sighted and blind participants received the same input during the scanning, we blindfolded all participants and turned off the lights in the scanning room.

The scanning session included one resting-state run at the beginning (8 min), 10 task-state runs (5 min 30 s each), and one run collecting T1 weighted images after the first five task-state runs (S1A Fig). During the resting-state run, participants were instructed to keep their heads still, not fall asleep, and not think about particular things. During the task-state runs, participants performed verification tasks on the words they heard.

Each task-state run was divided into two even blocks (S1A Fig). One corresponded to the shape verification task, and the other corresponded to the conceptual verification task. The order of the two task blocks was interleaved across runs within each subject, and whether the first run started with a shape or a conceptual block was counterbalanced across subjects within the EB and the SC group. Each block started with a 10 s rest, followed by a 20 s task probe. In the shape verification block, we instructed participants to think carefully about objects’ shape (“Pensa attentamente alla forma”) and judge whether they were elongated (“allungato”), angular (“angolare”), hollow (“cavo”), circular (“circolare”), and discal (“discoidale”). In the conceptual verification block, we instructed participants to think carefully about objects’ function (“Pensa attentamente alla funzione”) and judge whether they were used for eating (“per mangiare”), writing (“per scrivere”), sleeping (“per dormire”), lighting (“per illuminazione”), and purchasing (“per fare acquisti”). These five shape and conceptual verification tasks were randomly assigned to each participant’s first five task-state runs, and the second five task-state runs repeated these tasks in the same order. In this way, gaps between the same tasks were evenly distributed, and the same tasks could not be repeated in close time proximity. Participants made a yes/no judgment by pressing buttons using their right index/middle fingers. The button configuration (correspondence between yes/no judgments and index/middle fingers) in the first five runs was counterbalanced across subjects within the EB and the SC group. To counterbalance the motor effects of different fingers within subjects, we told each participant that the button configuration was switched in the second five runs.

Each block included 21 trials after the task probe, with 21 words presented once (S1B Fig). Each trial started with a 100 ms beep to capture participants’ attention, followed by a 300 ms silence and an auditory word (word duration: M = 662 ms, SD = 165 ms). The stimulus onset asynchrony was jittered as either 5 s or 8 s—11 trials lasted 5 s, and 10 trials lasted 8 s. The order of the words and the jitter intervals were randomized for each block. Participants were instructed to press buttons within 5 s after the stimulus onset. The RT was measured as the interval between the stimulus onset and the button press.

After the fMRI scanning, participants also rated object properties. For pairwise shape similarity and pairwise conceptual association, we adapted the paradigm for both sighted and blind populations by presenting the stimuli in the auditory modality. In each trial, participants heard two words in sequence and rated on a 7-point Likert scale (for shape rating, 1: not similar at all, 7: identical in shape; for conceptual rating, 1: not associated at all, 7: strongly associated). Both rating tasks consisted of 210 trials covering all the possible object pairs. For the other three object properties as potential confounding factors, participants rated item-wise on a 7-point Likert scale. They were object size (1: as small as a needle, 7: as big as a television), toolness (1: non-tools like a lamp, 7: tools like a hammer), and conceptual association (1: weak contextually associated like a cellphone, 7: strong contextually associated like a bowling ball). S1 Text shows the survey questions of these ratings.

Behavior analysis

For pairwise shape similarity and conceptual association ratings, we averaged the rating scores across all participants who took part in the fMRI experiment and calculated the model RDMs for the following RSA (i.e., 7 minus the mean rating score). To investigate the organizational structure of the two model RDMs, we performed the clustering analysis using the k-means clustering algorithm [38,39]. The maximum number of iterations was 10,000, the number of times to repeat clustering using new initial cluster centroid positions was 100, and the silhouette criterion was adopted to decide the optimal number of clusters in the range from 2 to 10 [40]. We conducted this analysis using the kmeans and evalclusters function in Matlab 2021.

For the ratings on other object properties (i.e., object size, contextual association, and toolness) and touch experience, we averaged the rating scores across all participants to obtain a mean vector for each rating item. These mean rating vectors, together with word duration and word frequency, constituted the potential confounding factors. To investigate the effect of these factors in the subsequent parametric modulation analysis, we orthogonalized these unidimensional variables using principal component analysis. Varimax rotation was applied to increase the interpretability of components, and five RCs of which the eigenvalues were greater than 1 were selected. The principal component analysis was performed using the principal function in the R package psych 2.1.9.

Analyses of the performance during fMRI scanning were conducted using JASP (Version 0.16).

MRI acquisition

MRI data were acquired using a MAGNETOM Prisma 3T MR scanner (Siemens) with a 64-channel head-neck coil at the Center for Mind/Brain Sciences, University of Trento. Functional images were acquired using the simultaneous multislices echoplanar imaging sequence: the scanning plane was parallel to the bicommissural plane, the phase encoding direction was from anterior to posterior, repetition time (TR) = 1000 ms, echo time (TE) = 28 ms, flip angle (FA) = 59°, multiband factor = 5. All participants in the early blind and sighted control groups and seven participants in the independent sighted group used a 3 mm spatial resolution: field of view (FOV) = 198 mm × 198 mm, matrix size = 66 × 66, 65 axial slices, slices thickness (ST) = 3 mm, gap = 0.3 mm, voxel size = 3 × 3 × (3 + 0.3) mm. The rest nine participants in the independent sighted group used a 2 mm spatial resolution: FOV = 200 mm × 200 mm, matrix size = 100 × 100, 65 axial slices, ST = 2 mm, gap = 0.2 mm, voxel size = 2 × 2 × (2 + 0.2) mm. Three-dimensional T1-weighted images were acquired using the magnetization-prepared rapid gradient-echo sequence, sagittal plane, TR = 2,140 ms, TE = 2.9 ms, inversion time = 950 ms, FA = 12°, FOV = 288 mm × 288 mm, matrix size = 288 × 288, 208 continuous sagittal slices, ST = 1 mm, voxel size = 1 × 1 × 1 mm.

MRI preprocessing

We performed MRI preprocessing using fMRIPrep 20.0.5 ([117]; RRID: SCR_016216), based on Nipype 1.4.2 ([118]; RRID: SCR_002502). Please see S2 Text, a boilerplate text directly generated by the fMRIPrep. It describes the detailed preprocessing steps used in the current study, aiming for a clear and consistent description to improve experimental reproducibility.

As surface-based analysis can significantly improve the spatial localization compared to the traditional volume-based analysis [119], we analyzed the images in the surface space generated by fMRIPrep (i.e., the fsaverage5 or the fsnative space). We conducted the surface smoothing of the functional images with a full width at half maximum (FWHM) of 6 mm using the mri_surf2surf command in FreeSurfer (http://surfer.nmr.mgh.harvard.edu/).

First-level neuroimaging analysis

We performed the first-level analysis using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). Individual-level GLMs were built separately for univariate contrast, parametric modulation, and RSA. In all three GLMs, six rigid-body transformation parameters and constant variables indicating each of the 10 runs were involved as nuisance regressors. A high-pass filter with a cutoff of 512 s was used to remove low-frequency noise and slow drifts. The RSA used unsmoothed images, while the other two analyses used smoothed images.

The GLM for the univariate contrast analysis involved three events—the shape task, the conceptual task, and the task probe. The duration of shape and conceptual tasks was set as each trial’s RT, and the duration of task probes was set as the auditory period before each block introducing the task ahead. The resulting boxcar function was convolved with a canonical hemodynamic response function (HRF). In this way (i.e., the variable epoch approach), the trial-by-trial RT variability was modeled [46]. To further control the domain-general effect of RT across the two tasks, we also used stick functions to model the trial-by-trial RT variability. We pooled the trials in the two tasks together, modulated the amplitude of sticks by the mean-centered RT, and convoluted the RT-modulated stick function with the canonical HRF (i.e., the variable impulse approach). The resulting RT variable was involved in the GLM as one regressor. We contrasted the shape task, the conceptual task, and the RT regressor to the resting state and contrasted between shape and conceptual tasks. The obtained combined beta images were used in the second-level analysis.

The GLM for the parametric modulation analysis only involved two conditions—the trials (i.e., shape and conceptual tasks pooled together as one condition) and the task probes. The duration of trials was set as its RT, and the duration of task probes was set as the auditory period before each block introducing the task ahead. We modulated the condition of the trials with a set of parametric variates, including the task type (i.e., the shape task coded as one and the conceptual task coded as −1), the z-scores of the RT across all the trials in each run, the RCs corresponding to word duration, word frequency, object size, toolness, and touch experience. The option for orthogonalizing modulations in the SPM was turned off [120]. We contrasted each parametric modulator to zero. The obtained combined beta images were used in the second-level analysis.

The GLM for the RSA involved each word in each task as a separate condition and the task probes as one condition. We concatenated 10 runs to improve the reliability of the model estimation [35]. The duration of trials was set as its RT, and the duration of task probes was set as the auditory period before each block introducing the task ahead. The trial-by-trial RT variability across the two tasks was also modeled using the variable impulse approach. We contrasted each word in each task to the resting state. The obtained T images instead of the beta images were used in the following RSA [121].

Representational similarity analysis

The RSA was conducted among the 21 object conditions within shape and conceptual tasks separately. It included two steps of correlation [35]. In the first-order correlation, we calculated the Spearman distance of the activity patterns across vertices between each pair of conditions and obtained a 21 × 21 neural RDM for a particular region. In the second-order correlation, we correlated the neural RDM and each model RDM (i.e., shape similarity and conceptual association) across the 210 pairs using Spearman correlation. The resulting correlation coefficients were Fisher z-transformed using the inverse hyperbolic function.

The ROI-based RSA focused on two sets of ROIs derived from significant brain areas in the second level of the parametric modulation analysis (see below). The shape ROIs were bilateral and had significantly greater activation in the shape task than in the conceptual task—the ILOTC, the aIPS, the pIPS, and the vPMC. The conceptual ROIs were left-lateralized and were significant in the opposite contrast—the orbital IFG, the aLTC, the pSTG, the AG, and the SMG. In cases when clusters were stuck together under the conventional threshold (vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05), we raised the vertex-wise threshold until they were isolated.

The searchlight-based RSA was performed to provide a global view of the results [122]. The searchlight spot went through all the vertices on the fsaverage5 surface. For each vertex, the spot included the six vertices directly connecting to the central vertex and the more peripheral vertices connecting to the six vertices (i.e., 19 vertices in total) [123]. The Fisher z-transformed second-order correlation coefficient was assigned back to the central vertex, and a surface smoothing with a 6 mm FWHM was applied to the resulting maps.

To investigate whether the ILOTC in EB and SC represented the same content, we compared the inter-subject neural RDM correlation within the same group (i.e., EB-EB and SC-SC) and between different groups (i.e., EB-SC). The within-group inter-subject correlation was calculated in a leave-one-subject-out manner. The neural RDM of the ILOTC of one participant was correlated to the mean neural RDMs of all the other participants within the same group across the 210 object pairs. This procedure ended up with 16 correlation coefficients for each group. The between-group inter-subject correlation was calculated in two steps. First, the neural RDM of the ILOTC of each participant in one group was correlated to the mean neural RDMs of all the participants in the other group across the 210 object pairs, which generated 16 correlation coefficients for each group. Second, we averaged the correlation coefficients from the EB and SC participants in the same pair to obtain 16 between-group correlation coefficients. These correlation coefficients were calculated using Spearman’s correlation and were Fisher z-transformed.

To provide a planar visualization of the representational pattern in bilateral ILOTC in the shape task, we performed the multidimensional scaling analysis using the mdscale function in Matlab 2021. The input dissimilarity matrix was the mean Euclidean distance between each pair of conditions averaged across all participants (N = 48). We used the squared stress, normalized with the sum of fourth powers of the dissimilarities, as the goodness-of-fit criterion to minimize.

Resting-state functional connectivity

We started with the unsmoothed resting-state images. To remove nonneuronal nuisance variables, we built a GLM to predict the timecourse of each vertex using the 24 head motion regressors [124], the mean timecourses in a conservative mask of the white matter and the cerebrospinal fluid extracted by the fMRIPrep, and the linear trend with the time points. We estimated the beta coefficients using the fitglm function in Matlab 2021 and subtracted all the terms (i.e., the dot product of all the nuisance variables and their estimated beta coefficients) from the original timecourses. A band-pass filter (0.01 to 0.1 Hz) was then performed on the resulting timecourses using the infinite impulse response filter method, and surface smoothing was carried out with a 6 mm FWHM. The functional connectivity between the two regions was defined as Pearson’s correlation between their timecourses. The correlation coefficients were Fisher z-transformed before the second-level analysis. The ROIs used in the seed-based RSFC and the interregional RSFC also came from the parametric modulation analysis.

Second-level neuroimaging analysis

We performed the group-level one-sample test or two-sample test (i.e., EB versus SC) on the first-level beta images from the univariate contrast and parametric modulation analyses, the searchlight-based RSA images, and the RSFC images. The statistic inference was made using the permutation method with PALM (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PALM). Five thousand sign flips were performed [125]. It is worth noting that, in the two-sample tests, we also chose the sign-flipping method assuming independent and symmetric errors instead of the traditional permutation method assuming exchangeable errors. This is because the variance of the early-blind group, on many occasions, is greater than the variance of the sighted-control group (Fig 4C; e.g., [126]), which violates the equal variance assumption of exchangeability. For the p-value below 0.01, we fit a generalized Pareto distribution to model the tail of the permutation distribution, aiming to improve the precision of the p-values [127,128].

In most cases, we controlled the FWE rate using a conventional cluster-forming threshold (i.e., vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). In the cases when the cluster-forming threshold was not suitable (i.e., distributed clusters spliced together), we controlled the FWE rate using a more conservative vertex-wise threshold (i.e., vertex-wise FWE corrected p < 0.005). We corrected the multiple comparisons of the two hemispheres using Bonferroni correction—the threshold set on each hemisphere was vertex-wise p < 0.001, cluster-level FWE corrected p < 0.025, or vertex-wise FWE corrected p < 0.0025.

Brain visualization

The brain results were illustrated using the Connectome Workbench 1.5.0 (https://www.humanconnectome.org/software/connectome-workbench). We mapped the significant brain areas from the fsaverage5 surface to the fsLR surface using the ADAP_BARY_AREA method for visualization purposes. They were displayed on an inflated surface against the group-averaged all sulcus image of 1,096 young adults from the dataset of the Human Connectome Project (https://balsa.wustl.edu/reference/pkXDZ).