Hearing Research 315 (2014) 88e98
Contents lists available at ScienceDirect
Hearing Research journal homepage: www.elsevier.com/locate/heares
Research paper
Morphological brain network assessed using graph theory and network filtration in deaf adults Eunkyung Kim a, b, c, 1, Hyejin Kang a, d, 1, Hyekyoung Lee a, b, Hyo-Jeong Lee e, Myung-Whan Suh g, Jae-Jin Song h, Seung-Ha Oh f, g, **, Dong Soo Lee a, b, c, i, * a
Department of Nuclear Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea Institute of Radiation Medicine, Medical Research Center, Seoul National University, Seoul, Republic of Korea Interdisciplinary Program in Cognitive Science, Seoul National University, Seoul, Republic of Korea d Data Science for Knowledge Creation Research Center, Seoul National University, Seoul, Republic of Korea e Department of Otorhinolaryngology-Head and Neck Surgery, Hallym University College of Medicine, Chuncheon, Republic of Korea f Department of Otorhinolaryngology-Head and Neck Surgery, Seoul National University College of Medicine, Seoul, Republic of Korea g Sensory Organ Research Institute, Seoul National University Medical Research Center, Seoul, Republic of Korea h Department of Otorhinolaryngology-Head and Neck Surgery, Seoul National University Bundang Hospital, Seongnam, Republic of Korea i Department of Molecular Medicine and Biopharmaceutical Sciences, Graduate School of Convergence Science and Technology, and College of Medicine or College of Pharmacy, Seoul National University, Seoul, Republic of Korea b c
a r t i c l e i n f o
a b s t r a c t
Article history: Received 10 March 2014 Received in revised form 13 June 2014 Accepted 24 June 2014 Available online 10 July 2014
Prolonged deprivation of auditory input can change brain networks in pre- and postlingual deaf adults by brain-wide reorganization. To investigate morphological changes in these brains voxel-based morphometry, voxel-wise correlation with the primary auditory cortex, and whole brain network analyses using morphological covariance were performed in eight prelingual deaf, eleven postlingual deaf, and eleven hearing adults. Network characteristics based on graph theory and network filtration based on persistent homology were examined. Gray matter density in the primary auditor cortex was preserved in prelingual deafness, while it tended to decrease in postlingual deafness. Unlike postlingual, prelingual deafness showed increased bilateral temporal connectivity of the primary auditory cortex compared to the hearing adults. Of the graph theory-based characteristics, clustering coefficient, betweenness centrality, and nodal efficiency all increased in prelingual deafness, while all the parameters of postlingual deafness were similar to the hearing adults. Patterns of connected components changing during network filtration were different between prelingual deafness and hearing adults according to the barcode, dendrogram, and single linkage matrix representations, while these were the same in postlingual deafness. Nodes in frontolimbic and left temporal components were closely coupled, and nodes in the temporo-parietal component were loosely coupled, in prelingual deafness. Patterns of connected components changing in postlingual deafness were the same as hearing adults. We propose that the preserved density of auditory cortex associated with increased connectivity in prelingual deafness, and closer coupling between certain brain areas, represent distinctive reorganization of auditory and related cortices compared with hearing or postlingual deaf adults. The differential network reorganization in the prelingual deaf adults could be related to the absence of auditory speech experience. © 2014 Elsevier B.V. All rights reserved.
Abbreviations: pre-LD, prelingual deaf; post-LD, postlingual deaf; A1, primary auditory cortex; VOIs, volumes of interest * Corresponding author. Department of Nuclear Medicine, Seoul National University College of Medicine, 28 Yeongeon-dong, Jongno-gu, Seoul, 110-744, Republic of Korea. Tel.: þ82 2 2072 2501; fax: þ82 2 745 7690. ** Corresponding author. Department of Otorhinolaryngology, Seoul National University College of Medicine, 28 Yeongeon-dong, Jongno-gu, Seoul, 110-744, Republic of Korea. Tel.: þ82 2 2072 2442; fax: þ82 2 831 2826. E-mail addresses: [email protected] (S.-H. Oh), [email protected] (D.S. Lee). 1 Contributed equally to this work. http://dx.doi.org/10.1016/j.heares.2014.06.007 0378-5955/© 2014 Elsevier B.V. All rights reserved.
1. Introduction Both auditory and visual deprivations have major effects on the brain (Rauschecker, 1999, 2001; Sadato et al., 2002). In deaf children, auditory and adjacent cortices showed hypo-metabolism (Lee et al., 2001), but in prelingual deaf (pre-LD) adults these areas showed hyper-metabolism in long-term follow-up (CatalanAhumada et al., 1993). In pre-LD adults, gray matter volume was
E. Kim et al. / Hearing Research 315 (2014) 88e98
also found to be preserved in the auditory cortex on morphometric analyses (Emmorey et al., 2003; Penhune et al., 2003; Shibata, 2007). The increased metabolism and preserved morphological volume represent “brain reorganization by neural plasticity” in preLD adults. In postlingual deaf (post-LD) adults, metabolism in the auditory cortices decreased shortly after the onset of deafness, but had recovered upon long-term follow-up (Lee et al., 2003). However, very little has been reported about the morphological changes of auditory cortices in post-LD adults, especially with long-term follow-up. These functional or morphological changes, however, could not entirely explain the pattern of brain reorganization in pre-LD and post-LD adults, especially considering the network characteristics of the brain. The characteristics of brain network can be measured using graph theoretical analysis (Sporns, 2011; Sporns et al., 2004) or network filtration of persistent homological framework (Lee et al., 2011, 2012). In this study, we investigated the morphological reorganization of deaf adults using both graph theoretical analysis and network filtration methods based on the morphological covariance of the unmodulated concentrations of gray matter tissue (i.e. gray matter density) of the brain regions. Several papers have substantiated that the individual's variability in morphometric or morphological features such as gray matter density, volume, or cortical thickness can imply structural association/interaction between brain regions like functional connectivity or association (Alexander-Bloch et al., 2013). For instance, the significant covariance of bilateral homologue regions was observed by using gray matter volume (Mechelli et al., 2005), which is similar to metabolic covariance (Lee et al., 2008). The morphological covariance of the default mode network was associated with general cognitive status, and as expected declined in Alzheimer's disease (Spreng and Turner, 2013). Using the morphological feature of cortical thickness as obtained by structural T1 images, significant covariance between Broca's and Wernicke's areas was observed overlapping with the arcuate fasciculus (Lerch et al., 2006). More recently, an intrinsic similarity of information traffic patterns was observed between the covariance of cortical thickness and the diffusion connections of the brain (Gong et al., 2012). Morphological covariance is related to the fibre tracts or functional interactions in the brain. Thus, if crosssubject covariance in tissue density maps is correctly analyzed and understood its correlates, it could be a companion to understanding the active reorganization of brain networks (Evans, 2013; Evans et al., 2008), including those in deaf adults. Using a graph theoretical approach and network filtration methods on data acquired from a correlation analysis of MRI density of specified volumes of interest (VOIs), we described characteristics of morphological covariance network quantitatively in deaf adults. The complex network patterns of human brains have been analyzed by graph theoretical measures that quantify the topological properties of the network (Sporns, 2011; Sporns et al., 2004). To understand the network characteristics of deaf adults' brains, we measured segregation, centrality, and efficiency. Network segregation can be quantified by the clustering coefficient (Bassett et al., 2008; He et al., 2008), which measures the connections between sets of nodes that belong to the neighbors of a specific node and normalized by all the possible connections of these neighbor nodes. Betweenness centrality quantifies the number of shortest paths between two arbitrary brain nodes passing through a specific node in relation to all the numbers of shortest paths connecting these two brain nodes (Gong et al., 2009; Iturria-Medina et al., 2008). Network efficiency is measured using nodal efficiency, which shows how efficient the fault-tolerance of a connection is in a specific node by calculating the sum of the inverse length of the shortest paths from that node to other nodes (Latora and Marchiori, 2001). Intuitively, the clustering coefficient of a node indicates the
89
strength of connections between its neighbor nodes. Betweenness centrality indicates how many network shortcuts pass through the node. Nodal efficiency indicates how efficiently a node is connected to all other nodes in the brain. Network filtration using a persistent homological framework (Lee et al., 2011, 2012) was applied to characterize the topological features of networks using data acquired from interregional correlations of density on MRIs. Thresholding of the brain network was found to affect network topology (Rubinov and Sporns, 2010). In terms of network edge, the sparsest network with a giant all-nodeconnected component could be chosen in each group, but the number of edges could not be the same between groups. Optimal sparsity, selecting the giant connected components with the same edge numbers between groups, was also arbitrary in that it used specific thresholds for each group. Threshold choice led to disregarding weaker correlations, which could be a sensitive indicator to differentiate and characterize the brain networks of different groups (Bassett et al., 2012). Instead of choosing one arbitrary threshold, we applied network filtration using varying distance thresholds, which allowed us to look into network changes with varying thresholds. Counting the connected components in the thresholded network was based upon the topological interpretation of networks on a persistent homological framework, which was once successfully applied to differentiation of psychiatric patients using correlations acquired from PET images (Lee et al., 2012). In this study, we investigated whether the gray matter density of the primary auditory cortex (A1) was preserved in pre-LD or post-LD adults. Reduction of gray matter density was considered to represent the loss of neurons and/or glial cells (Rusch et al., 2003), and thus the decrease of the local amount of the gray matter (Chen et al., 2006). We assumed that the preserved density of the gray matter reflected recovery in the affected brain regions following long-term adaptation to the sensory deficit in deaf adults. If pre-LD adults demonstrated morphological recovery while post-LD adults did not, it would confirm that pre-LD adults have network properties different from those of post-LD adults or the hearing adults. We assumed that this morphological recovery would lead to preserved tissue MRI density, but with a different organization expressed as different interregional correlations. Regarding the persistent homological characteristics of the brain networks, we expected that network filtration would disclose differential characteristics between pre-LD or post-LD adults and normal hearing adults. We explored how the networks in deaf adults were reorganized compared with hearing adults after prolonged auditory deprivation had influenced and changed the topological characteristics of the networks. 2. Materials and methods 2.1. Participants Three groups of subjects participated in this study: eight pre-LD (5M/3F; mean age, 50.4 ± 6.1 years; duration of deafness, 45.8 ± 6.5 years; onset age of deafness, 4.6 ± 1.4 years), eleven post-LD (4M/ 7F; mean age, 50.9 ± 12.2 years; duration of deafness, 15.3 ± 14.0 years; onset age of deafness, 35.6 ± 18.2 years), and eleven agematched normal hearing adults (6M/5F; mean age, 49.5 ± 8.9 years). All participants were right-handed and were given at least 9 years of formal education. All deaf participants had a mean threetone unaided threshold of >70 dB hearing level (HL), the International Organization for Standardization (ISO) criterion for severe to profound hearing loss. The average hearing threshold of the hearing adults was 13 dB HL. In this study, the pre-LD adults became deaf before or around the usual language acquisition period, which did not exceed 7 years of age. The pre-LD adults used sign language as their primary
90
E. Kim et al. / Hearing Research 315 (2014) 88e98
communication tool, and were not aided by hearing aids or cochlear implants. In contrast, the post-LD became deaf after language acquisition, and used lip-reading skills to communicate with others and spoke fluently with hearing aids. Participants had normal/corrected-normal visual acuity and no previous history of neurological or psychiatric disorders. This study was approved by the institutional review board (IRB) of the Seoul National University College of Medicine. All participants gave informed consent as provided by the Seoul National University Hospital. 2.2. MR image acquisition Structural MR images were acquired using GE Signa 3.0 T EXCITE systems (General Electric Healthcare, Milwaukee, WI). T1 images of deaf groups were acquired using a three-dimensional (3D) spoiled gradient-recall (SPGR) inversion recovery (IR) acquisition protocol with the following parameters: axial acquisition with a 320 mm 192 mm; TR ¼ 5.9 ms; TE ¼ 1.5 ms; FOV 200 mm, except for one pre-LD adult whose TR ¼ 6 ms. T1 images of the hearing adults were acquired using the same parameters (eight subjects) as the deaf groups, except for 3 subjects who had different parameters: TR ¼ 6 ms, 5.8 ms, 4.6 ms; TE ¼ 1.5 ms, 1.4 ms, 1.2 ms, respectively. All scans had 104 or 106 continuous slices, and a 512 x 512 image matrix with a 0.39 0.39 1.5 mm voxel size, except one of the hearing adults whose voxel size was 0.47 0.47 1.5 mm. The thickness was 1.5 mm and the flip angle was 20 . 2.3. Image pre-processing The unmodulated concentration of gray matter tissue - a density map of the gray matter - was generated using voxel-based morphometry (VBM) with the DARTEL (Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra) tool, implemented in Statistical Parametric Mapping (SPM 8, www.fil.ion.ucl. ac.uk/spm) to acquire better registration results (Ashburner, 2007). First, each 3D T1 image was segmented after being aligned to the anterior and posterior commissures, and an average shaped brain template and flow field files of rigidly aligned gray matter images were created. The flow fields were used to normalize the gray matter images to Montreal Neurological Institute (MNI) space by non-affine warping. For VBM analysis, an 8 mm FWHM was used for spatial smoothing. To construct the morphological covariance network, correlation coefficients between VOIs were calculated from gray matter density maps without spatial smoothing, to avoid the effects of spatial overlapping between VOIs (Van den Heuvel et al., 2008). Network nodes were defined as 90 regions chosen from the Automated Anatomical Labeling (AAL) template (http://www.cyceron.fr/web/ aal__anatomical_automatic_labeling.html). 2.4. Voxel-based morphometric analysis After removing non-brain voxels by masking the brain, ANCOVA with three conditions (groups) was performed with the age and global covariates as nuisance variables. In this analysis, the significance level was set at P < 0.005 within a cluster of an extent threshold of k > 50 voxels. We also investigated the effect of duration and onset age of deafness on gray matter density of bilateral A1 using Spearman's correlation coefficients in the pre-LD and post-LD groups. 2.5. Mapping correlation with A1 region seed density After removing non-brain voxels, the effects of age and total gray matter density were controlled using ANCOVA to discount the
global variation in the gray matter caused by age and head size. Correlation analyses within groups were performed with the density values of A1 regions as seed regions in each hemisphere. The density value of A1 was extracted from the AAL. Correlation r maps were transformed to z maps by Fisher's z transformation and compared between the deaf groups and hearing adults to yield group differences (P < 0.005 within a cluster of an extent threshold of k > 50). 2.6. Mapping the whole brain covariance network over the VOIs as nodes 2.6.1. Network parametric characterization based on graph theory To construct the whole brain covariance network, gray matter density of 90 VOIs was extracted from individual maps. In each group, a positive correlation matrix was constructed across the 90 VOIs with age as a nuisance variable. If we choose a threshold arbitrarily for the hearing adults, pre-LD, and post-LD, the networks can be biased because the groups have a different number of edges (He et al., 2008). To avoid these biases, we chose the threshold while considering the sparsity of the matrix (Achard and Bullmore, 2007; Bassett et al., 2008; He et al., 2007, 2008; Wang et al., 2010), where sparsity was estimated and chosen for each group to create the same number of edges for each group. In addition, the sizes of connected components were made the largest (and the same) for all groups. Thus, the number of connected components is equal to one. Sparsity follows the equation S ¼ K/(N (N1)/2), where K represents the total number of edges, and N represents the number of nodes. We chose the threshold as a 12.3%, at which point the cost of the network was minimized, while the fully connected components in each group were maintained. From these sparsityoptimized, weighted network matrices, network properties such as the clustering coefficient, betweenness centrality, and nodal efficiency were calculated using the Brain Connectivity Toolbox (http://www.brain-connectivity-toolbox.net/). Each measure characterized the nodal properties. The clustering coefficient reflects local segregation of a given specific node in a network (Watts and Strogatz, 1998). It represents local connectivity by calculating connections between neighbors of a given node among all possible connections of these neighbor nodes. It ranges from zero to one, where a value close to zero indicates weaker connections between neighbors of a given node, while the value close to one indicates that more neighbors of a given node are connected. Betweenness centrality represents the centrality of information flow through a given specific node in a network (Rubinov and Sporns, 2010). It is measured per node by counting the fraction of all the shortest paths in the network which pass through a specific node in the correlation network (Rubinov and Sporns, 2010). The shortest path stands for a path between two nodes in a network which pass from node i to node j as a shortcut. The node having low value indicates that the specific node rarely has the shortest path passing through itself, while that having high value indicates the specific node has many paths passing through itself. Nodal efficiency represents the nodal faulttolerance of information flow when the specific node is removed (Latora and Marchiori, 2001). It is measured per node by the sum of the inverse of the shortest path lengths of a specific node yielding paths to the other nodes. The node having low value indicates that the specific node is inefficiently connected to the other nodes in a network, while that having high value indicates that the specific node is efficiently connected to the other nodes in a network. After obtaining each measure per nodes in each group, we drew the frequency distribution of the clustering coefficient, betweenness centrality, and nodal efficiency over the nodes, and compared the distributions of the groups (hearing adults, pre-LD, and post-
E. Kim et al. / Hearing Research 315 (2014) 88e98
LD) using a two-sample KolmogoroveSmirnov (KS) test based on the general form of a bootstrap procedure. We created a null distribution of the parameter of the KS test, statistic D, as repeated 10,000 times, and set the statistical significance from the null distribution. We also investigated which node(s) had significantly different network properties compared with the other groups by generating a null distribution of network properties. More specifically, the gray matter density of 90 VOIs extracted from individual maps was concatenated into one single matrix. The size of the matrix was 30 (number of subjects)-by-90 (number of VOIs). We randomly selected data comprised of pseudo-groups. Partial correlation matrices were generated in each pseudo-group using the corresponding age. Using this partial correlation matrix, we estimated the sparsity by selecting the smallest number of edges across groups, while maintaining the fully connected components, and then obtained each network property per node in each pseudogroup. To compare network properties between groups, the differences of the network properties were computed. This was done 5000 times to generate a null distribution of the difference value of network properties per node. Significance was set at P < 0.005 (two-tailed) for group comparison. 2.6.2. Network filtration to yield all-threshold correlation of density based on a persistent homological framework To construct brain networks of MRI density across all thresholds, where r was the correlation coefficient between the gray matter density in VOIs with age as a nuisance variable, the distance matrix (1-r) was used to generate a barcode, single linkage dendrogram, and single linkage matrix for each group (hearing adults, pre-LD, and post-LD) based on persistent homological perspectives (Lee et al., 2012). When the network was filtered, two nodes were considered to be connected only if they had a shorter distance than the threshold. By increasing the threshold distances, more nodes were allowed to be connected to each other (Lee et al., 2011). The barcode represented changes in the number of connected components of the filtered network when the threshold was varied during network filtration. In the barcode, each connected component is represented by a bar which starts and ends at the filtration values when the corresponding connected component appears and disappears. The number of connected components is monotonically decreasing function for all filtration values. Its maximum value is equal to the number of nodes (number of VOIs) at the filtration value zero, and its minimum value is one when all nodes are connected. If bars in the barcode are rearranged according to the location of the VOIs, it becomes a dendrogram which represents the hierarchical clustering of brain regions. A single linkage matrix is the matrix representation of a single linkage dendrogram, and was used to disclose the statistical differences between the hearing adults, pre-LD, and post-LD groups. The single linkage matrix shows the local change of connected structure of brain network during filtration. Since its element represents the filtration value when two brain regions belong to the same connected component for the first time, we can find which areas are connected earlier than the other regions. A nonparametric, unpaired permutation test was used to obtain the probability that the observed difference between two groups occurred by chance (the null hypothesis) by generating the pseudo-matrices of the single linkage matrix. To yield two pseudo-group data (pseudohearing adults and pseudo-pre-LD), individuals were randomly chosen from data of mixed population, and single linkage distances between nodes i and j (dij) were calculated for each pseudo-group, with partialling out corresponding age. For each pseudo-group's data, single linkage distances were converted to z-transformed metrics following the equation, zij ¼ 1/2 log [(2 dij)/(dij)]. These z-
91
values were then compared between groups by z statistics using Zdiffij ¼ (zij1 zij2)/√(1/(n13)þ1/(n23)), where n1 and n2 represent the number of subjects in each group. This procedure was repeated 5000 times to generate the null distribution of Zdiffij, which represents a group difference between single linkage matrices. Significance was set at P < 0.001 (two-tailed) for group comparison. 3. Results 3.1. Gray matter density in pre-LD and post-LD compared with hearing adults After correcting for multiple comparisons via Bonferroni correction and the False Discovery Rate (FDR) correction (Benjamini and Hochberg, 1995), there was no significant difference between each deaf group and hearing adults. However, when analyzed using the uncorrected P < 0.005 (k > 50), the gray matter density of bilateral A1 was preserved in the pre-LD group (Fig. 1A), but decreased in the post-LD group (Fig. 1B). There was no significant correlation between the density of bilateral A1 and duration or onset of deafness in both groups. 3.2. Correlation of bilateral superior temporal regions with A1 density in pre-LD and post-LD compared with hearing adults The left and right A1s were correlated positively with bilateral superior temporal regions, including A1, in all three groups (Supplementary Fig. 1). Group comparisons between the pre- or post-LD groups and the hearing adults showed a differential correlation using the uncorrected P < 0.005 (k > 50). Compared with the hearing adults, the pre-LD group showed increased correlation coefficients of left A1 with the bilateral superior temporal gyri, superior temporal pole, middle temporal gyri, insulae, and left inferior frontal operculum (Fig. 2A, upper row). Also, the pre-LD group showed similarly increased correlation coefficients of right A1 with the bilateral superior temporal gyri, right superior temporal pole, bilateral middle temporal gyri, and insulae as compared with the hearing adults (Fig. 2B, upper row). In contrast, there were no significant differences between the post-LD group and the hearing adults with regard to the correlation between the bilateral A1s and other brain regions (Fig. 2A and B, lower rows). 3.3. Network characteristics of pre-LD, post-LD, and hearing adults using graph theory The distribution of the nodal properties of the network characteristics based on graph theoretical analysis yielded unique patterns in the hearing adults (Fig. 3, left column). In terms of the clustering coefficient (Fig. 3A), betweenness centrality (Fig. 3B), and nodal efficiency (Fig. 3C), the pre-LD group showed a different distribution from the hearing adults (Fig. 3, middle column). In contrast, the post-LD group showed a similar distribution of nodal properties to the hearing adults (Fig. 3, right column). In the pre-LD group, the nodal distribution of clustering coefficients was shifted to a higher value than that of hearing adults (P < 0.001), indicating there were more nodes with greater clustering coefficients in the pre-LD group. In addition, more nodes with higher betweenness centrality were found in the pre-LD group than in the hearing adults (P < 0.05). The skewness of betweenness centrality of the pre-LD group was also higher than that of the post-LD or the hearing adults (pre-LD; 1.43, post-LD; 0.68, hearing adults; 0.62). The nodal efficiency of the pre-LD group spread into the upper and lower extremes of the
92
E. Kim et al. / Hearing Research 315 (2014) 88e98
Fig. 1. Results of voxel-based morphometry. Compared with the hearing adults, (A) the density of bilateral A1 was preserved in the pre-LD group, while (B) the density of bilateral A1 was decreased in the post-LD group. Blue represents the areas where density was lower in the deaf groups than the hearing adults, while red represents the areas where density was higher in the deaf groups than the hearing adults. The A1 region is overlaid in purple. The statistical threshold was set as P value
Contents lists available at ScienceDirect
Hearing Research journal homepage: www.elsevier.com/locate/heares
Research paper
Morphological brain network assessed using graph theory and network filtration in deaf adults Eunkyung Kim a, b, c, 1, Hyejin Kang a, d, 1, Hyekyoung Lee a, b, Hyo-Jeong Lee e, Myung-Whan Suh g, Jae-Jin Song h, Seung-Ha Oh f, g, **, Dong Soo Lee a, b, c, i, * a
Department of Nuclear Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea Institute of Radiation Medicine, Medical Research Center, Seoul National University, Seoul, Republic of Korea Interdisciplinary Program in Cognitive Science, Seoul National University, Seoul, Republic of Korea d Data Science for Knowledge Creation Research Center, Seoul National University, Seoul, Republic of Korea e Department of Otorhinolaryngology-Head and Neck Surgery, Hallym University College of Medicine, Chuncheon, Republic of Korea f Department of Otorhinolaryngology-Head and Neck Surgery, Seoul National University College of Medicine, Seoul, Republic of Korea g Sensory Organ Research Institute, Seoul National University Medical Research Center, Seoul, Republic of Korea h Department of Otorhinolaryngology-Head and Neck Surgery, Seoul National University Bundang Hospital, Seongnam, Republic of Korea i Department of Molecular Medicine and Biopharmaceutical Sciences, Graduate School of Convergence Science and Technology, and College of Medicine or College of Pharmacy, Seoul National University, Seoul, Republic of Korea b c
a r t i c l e i n f o
a b s t r a c t
Article history: Received 10 March 2014 Received in revised form 13 June 2014 Accepted 24 June 2014 Available online 10 July 2014
Prolonged deprivation of auditory input can change brain networks in pre- and postlingual deaf adults by brain-wide reorganization. To investigate morphological changes in these brains voxel-based morphometry, voxel-wise correlation with the primary auditory cortex, and whole brain network analyses using morphological covariance were performed in eight prelingual deaf, eleven postlingual deaf, and eleven hearing adults. Network characteristics based on graph theory and network filtration based on persistent homology were examined. Gray matter density in the primary auditor cortex was preserved in prelingual deafness, while it tended to decrease in postlingual deafness. Unlike postlingual, prelingual deafness showed increased bilateral temporal connectivity of the primary auditory cortex compared to the hearing adults. Of the graph theory-based characteristics, clustering coefficient, betweenness centrality, and nodal efficiency all increased in prelingual deafness, while all the parameters of postlingual deafness were similar to the hearing adults. Patterns of connected components changing during network filtration were different between prelingual deafness and hearing adults according to the barcode, dendrogram, and single linkage matrix representations, while these were the same in postlingual deafness. Nodes in frontolimbic and left temporal components were closely coupled, and nodes in the temporo-parietal component were loosely coupled, in prelingual deafness. Patterns of connected components changing in postlingual deafness were the same as hearing adults. We propose that the preserved density of auditory cortex associated with increased connectivity in prelingual deafness, and closer coupling between certain brain areas, represent distinctive reorganization of auditory and related cortices compared with hearing or postlingual deaf adults. The differential network reorganization in the prelingual deaf adults could be related to the absence of auditory speech experience. © 2014 Elsevier B.V. All rights reserved.
Abbreviations: pre-LD, prelingual deaf; post-LD, postlingual deaf; A1, primary auditory cortex; VOIs, volumes of interest * Corresponding author. Department of Nuclear Medicine, Seoul National University College of Medicine, 28 Yeongeon-dong, Jongno-gu, Seoul, 110-744, Republic of Korea. Tel.: þ82 2 2072 2501; fax: þ82 2 745 7690. ** Corresponding author. Department of Otorhinolaryngology, Seoul National University College of Medicine, 28 Yeongeon-dong, Jongno-gu, Seoul, 110-744, Republic of Korea. Tel.: þ82 2 2072 2442; fax: þ82 2 831 2826. E-mail addresses: [email protected] (S.-H. Oh), [email protected] (D.S. Lee). 1 Contributed equally to this work. http://dx.doi.org/10.1016/j.heares.2014.06.007 0378-5955/© 2014 Elsevier B.V. All rights reserved.
1. Introduction Both auditory and visual deprivations have major effects on the brain (Rauschecker, 1999, 2001; Sadato et al., 2002). In deaf children, auditory and adjacent cortices showed hypo-metabolism (Lee et al., 2001), but in prelingual deaf (pre-LD) adults these areas showed hyper-metabolism in long-term follow-up (CatalanAhumada et al., 1993). In pre-LD adults, gray matter volume was
E. Kim et al. / Hearing Research 315 (2014) 88e98
also found to be preserved in the auditory cortex on morphometric analyses (Emmorey et al., 2003; Penhune et al., 2003; Shibata, 2007). The increased metabolism and preserved morphological volume represent “brain reorganization by neural plasticity” in preLD adults. In postlingual deaf (post-LD) adults, metabolism in the auditory cortices decreased shortly after the onset of deafness, but had recovered upon long-term follow-up (Lee et al., 2003). However, very little has been reported about the morphological changes of auditory cortices in post-LD adults, especially with long-term follow-up. These functional or morphological changes, however, could not entirely explain the pattern of brain reorganization in pre-LD and post-LD adults, especially considering the network characteristics of the brain. The characteristics of brain network can be measured using graph theoretical analysis (Sporns, 2011; Sporns et al., 2004) or network filtration of persistent homological framework (Lee et al., 2011, 2012). In this study, we investigated the morphological reorganization of deaf adults using both graph theoretical analysis and network filtration methods based on the morphological covariance of the unmodulated concentrations of gray matter tissue (i.e. gray matter density) of the brain regions. Several papers have substantiated that the individual's variability in morphometric or morphological features such as gray matter density, volume, or cortical thickness can imply structural association/interaction between brain regions like functional connectivity or association (Alexander-Bloch et al., 2013). For instance, the significant covariance of bilateral homologue regions was observed by using gray matter volume (Mechelli et al., 2005), which is similar to metabolic covariance (Lee et al., 2008). The morphological covariance of the default mode network was associated with general cognitive status, and as expected declined in Alzheimer's disease (Spreng and Turner, 2013). Using the morphological feature of cortical thickness as obtained by structural T1 images, significant covariance between Broca's and Wernicke's areas was observed overlapping with the arcuate fasciculus (Lerch et al., 2006). More recently, an intrinsic similarity of information traffic patterns was observed between the covariance of cortical thickness and the diffusion connections of the brain (Gong et al., 2012). Morphological covariance is related to the fibre tracts or functional interactions in the brain. Thus, if crosssubject covariance in tissue density maps is correctly analyzed and understood its correlates, it could be a companion to understanding the active reorganization of brain networks (Evans, 2013; Evans et al., 2008), including those in deaf adults. Using a graph theoretical approach and network filtration methods on data acquired from a correlation analysis of MRI density of specified volumes of interest (VOIs), we described characteristics of morphological covariance network quantitatively in deaf adults. The complex network patterns of human brains have been analyzed by graph theoretical measures that quantify the topological properties of the network (Sporns, 2011; Sporns et al., 2004). To understand the network characteristics of deaf adults' brains, we measured segregation, centrality, and efficiency. Network segregation can be quantified by the clustering coefficient (Bassett et al., 2008; He et al., 2008), which measures the connections between sets of nodes that belong to the neighbors of a specific node and normalized by all the possible connections of these neighbor nodes. Betweenness centrality quantifies the number of shortest paths between two arbitrary brain nodes passing through a specific node in relation to all the numbers of shortest paths connecting these two brain nodes (Gong et al., 2009; Iturria-Medina et al., 2008). Network efficiency is measured using nodal efficiency, which shows how efficient the fault-tolerance of a connection is in a specific node by calculating the sum of the inverse length of the shortest paths from that node to other nodes (Latora and Marchiori, 2001). Intuitively, the clustering coefficient of a node indicates the
89
strength of connections between its neighbor nodes. Betweenness centrality indicates how many network shortcuts pass through the node. Nodal efficiency indicates how efficiently a node is connected to all other nodes in the brain. Network filtration using a persistent homological framework (Lee et al., 2011, 2012) was applied to characterize the topological features of networks using data acquired from interregional correlations of density on MRIs. Thresholding of the brain network was found to affect network topology (Rubinov and Sporns, 2010). In terms of network edge, the sparsest network with a giant all-nodeconnected component could be chosen in each group, but the number of edges could not be the same between groups. Optimal sparsity, selecting the giant connected components with the same edge numbers between groups, was also arbitrary in that it used specific thresholds for each group. Threshold choice led to disregarding weaker correlations, which could be a sensitive indicator to differentiate and characterize the brain networks of different groups (Bassett et al., 2012). Instead of choosing one arbitrary threshold, we applied network filtration using varying distance thresholds, which allowed us to look into network changes with varying thresholds. Counting the connected components in the thresholded network was based upon the topological interpretation of networks on a persistent homological framework, which was once successfully applied to differentiation of psychiatric patients using correlations acquired from PET images (Lee et al., 2012). In this study, we investigated whether the gray matter density of the primary auditory cortex (A1) was preserved in pre-LD or post-LD adults. Reduction of gray matter density was considered to represent the loss of neurons and/or glial cells (Rusch et al., 2003), and thus the decrease of the local amount of the gray matter (Chen et al., 2006). We assumed that the preserved density of the gray matter reflected recovery in the affected brain regions following long-term adaptation to the sensory deficit in deaf adults. If pre-LD adults demonstrated morphological recovery while post-LD adults did not, it would confirm that pre-LD adults have network properties different from those of post-LD adults or the hearing adults. We assumed that this morphological recovery would lead to preserved tissue MRI density, but with a different organization expressed as different interregional correlations. Regarding the persistent homological characteristics of the brain networks, we expected that network filtration would disclose differential characteristics between pre-LD or post-LD adults and normal hearing adults. We explored how the networks in deaf adults were reorganized compared with hearing adults after prolonged auditory deprivation had influenced and changed the topological characteristics of the networks. 2. Materials and methods 2.1. Participants Three groups of subjects participated in this study: eight pre-LD (5M/3F; mean age, 50.4 ± 6.1 years; duration of deafness, 45.8 ± 6.5 years; onset age of deafness, 4.6 ± 1.4 years), eleven post-LD (4M/ 7F; mean age, 50.9 ± 12.2 years; duration of deafness, 15.3 ± 14.0 years; onset age of deafness, 35.6 ± 18.2 years), and eleven agematched normal hearing adults (6M/5F; mean age, 49.5 ± 8.9 years). All participants were right-handed and were given at least 9 years of formal education. All deaf participants had a mean threetone unaided threshold of >70 dB hearing level (HL), the International Organization for Standardization (ISO) criterion for severe to profound hearing loss. The average hearing threshold of the hearing adults was 13 dB HL. In this study, the pre-LD adults became deaf before or around the usual language acquisition period, which did not exceed 7 years of age. The pre-LD adults used sign language as their primary
90
E. Kim et al. / Hearing Research 315 (2014) 88e98
communication tool, and were not aided by hearing aids or cochlear implants. In contrast, the post-LD became deaf after language acquisition, and used lip-reading skills to communicate with others and spoke fluently with hearing aids. Participants had normal/corrected-normal visual acuity and no previous history of neurological or psychiatric disorders. This study was approved by the institutional review board (IRB) of the Seoul National University College of Medicine. All participants gave informed consent as provided by the Seoul National University Hospital. 2.2. MR image acquisition Structural MR images were acquired using GE Signa 3.0 T EXCITE systems (General Electric Healthcare, Milwaukee, WI). T1 images of deaf groups were acquired using a three-dimensional (3D) spoiled gradient-recall (SPGR) inversion recovery (IR) acquisition protocol with the following parameters: axial acquisition with a 320 mm 192 mm; TR ¼ 5.9 ms; TE ¼ 1.5 ms; FOV 200 mm, except for one pre-LD adult whose TR ¼ 6 ms. T1 images of the hearing adults were acquired using the same parameters (eight subjects) as the deaf groups, except for 3 subjects who had different parameters: TR ¼ 6 ms, 5.8 ms, 4.6 ms; TE ¼ 1.5 ms, 1.4 ms, 1.2 ms, respectively. All scans had 104 or 106 continuous slices, and a 512 x 512 image matrix with a 0.39 0.39 1.5 mm voxel size, except one of the hearing adults whose voxel size was 0.47 0.47 1.5 mm. The thickness was 1.5 mm and the flip angle was 20 . 2.3. Image pre-processing The unmodulated concentration of gray matter tissue - a density map of the gray matter - was generated using voxel-based morphometry (VBM) with the DARTEL (Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra) tool, implemented in Statistical Parametric Mapping (SPM 8, www.fil.ion.ucl. ac.uk/spm) to acquire better registration results (Ashburner, 2007). First, each 3D T1 image was segmented after being aligned to the anterior and posterior commissures, and an average shaped brain template and flow field files of rigidly aligned gray matter images were created. The flow fields were used to normalize the gray matter images to Montreal Neurological Institute (MNI) space by non-affine warping. For VBM analysis, an 8 mm FWHM was used for spatial smoothing. To construct the morphological covariance network, correlation coefficients between VOIs were calculated from gray matter density maps without spatial smoothing, to avoid the effects of spatial overlapping between VOIs (Van den Heuvel et al., 2008). Network nodes were defined as 90 regions chosen from the Automated Anatomical Labeling (AAL) template (http://www.cyceron.fr/web/ aal__anatomical_automatic_labeling.html). 2.4. Voxel-based morphometric analysis After removing non-brain voxels by masking the brain, ANCOVA with three conditions (groups) was performed with the age and global covariates as nuisance variables. In this analysis, the significance level was set at P < 0.005 within a cluster of an extent threshold of k > 50 voxels. We also investigated the effect of duration and onset age of deafness on gray matter density of bilateral A1 using Spearman's correlation coefficients in the pre-LD and post-LD groups. 2.5. Mapping correlation with A1 region seed density After removing non-brain voxels, the effects of age and total gray matter density were controlled using ANCOVA to discount the
global variation in the gray matter caused by age and head size. Correlation analyses within groups were performed with the density values of A1 regions as seed regions in each hemisphere. The density value of A1 was extracted from the AAL. Correlation r maps were transformed to z maps by Fisher's z transformation and compared between the deaf groups and hearing adults to yield group differences (P < 0.005 within a cluster of an extent threshold of k > 50). 2.6. Mapping the whole brain covariance network over the VOIs as nodes 2.6.1. Network parametric characterization based on graph theory To construct the whole brain covariance network, gray matter density of 90 VOIs was extracted from individual maps. In each group, a positive correlation matrix was constructed across the 90 VOIs with age as a nuisance variable. If we choose a threshold arbitrarily for the hearing adults, pre-LD, and post-LD, the networks can be biased because the groups have a different number of edges (He et al., 2008). To avoid these biases, we chose the threshold while considering the sparsity of the matrix (Achard and Bullmore, 2007; Bassett et al., 2008; He et al., 2007, 2008; Wang et al., 2010), where sparsity was estimated and chosen for each group to create the same number of edges for each group. In addition, the sizes of connected components were made the largest (and the same) for all groups. Thus, the number of connected components is equal to one. Sparsity follows the equation S ¼ K/(N (N1)/2), where K represents the total number of edges, and N represents the number of nodes. We chose the threshold as a 12.3%, at which point the cost of the network was minimized, while the fully connected components in each group were maintained. From these sparsityoptimized, weighted network matrices, network properties such as the clustering coefficient, betweenness centrality, and nodal efficiency were calculated using the Brain Connectivity Toolbox (http://www.brain-connectivity-toolbox.net/). Each measure characterized the nodal properties. The clustering coefficient reflects local segregation of a given specific node in a network (Watts and Strogatz, 1998). It represents local connectivity by calculating connections between neighbors of a given node among all possible connections of these neighbor nodes. It ranges from zero to one, where a value close to zero indicates weaker connections between neighbors of a given node, while the value close to one indicates that more neighbors of a given node are connected. Betweenness centrality represents the centrality of information flow through a given specific node in a network (Rubinov and Sporns, 2010). It is measured per node by counting the fraction of all the shortest paths in the network which pass through a specific node in the correlation network (Rubinov and Sporns, 2010). The shortest path stands for a path between two nodes in a network which pass from node i to node j as a shortcut. The node having low value indicates that the specific node rarely has the shortest path passing through itself, while that having high value indicates the specific node has many paths passing through itself. Nodal efficiency represents the nodal faulttolerance of information flow when the specific node is removed (Latora and Marchiori, 2001). It is measured per node by the sum of the inverse of the shortest path lengths of a specific node yielding paths to the other nodes. The node having low value indicates that the specific node is inefficiently connected to the other nodes in a network, while that having high value indicates that the specific node is efficiently connected to the other nodes in a network. After obtaining each measure per nodes in each group, we drew the frequency distribution of the clustering coefficient, betweenness centrality, and nodal efficiency over the nodes, and compared the distributions of the groups (hearing adults, pre-LD, and post-
E. Kim et al. / Hearing Research 315 (2014) 88e98
LD) using a two-sample KolmogoroveSmirnov (KS) test based on the general form of a bootstrap procedure. We created a null distribution of the parameter of the KS test, statistic D, as repeated 10,000 times, and set the statistical significance from the null distribution. We also investigated which node(s) had significantly different network properties compared with the other groups by generating a null distribution of network properties. More specifically, the gray matter density of 90 VOIs extracted from individual maps was concatenated into one single matrix. The size of the matrix was 30 (number of subjects)-by-90 (number of VOIs). We randomly selected data comprised of pseudo-groups. Partial correlation matrices were generated in each pseudo-group using the corresponding age. Using this partial correlation matrix, we estimated the sparsity by selecting the smallest number of edges across groups, while maintaining the fully connected components, and then obtained each network property per node in each pseudogroup. To compare network properties between groups, the differences of the network properties were computed. This was done 5000 times to generate a null distribution of the difference value of network properties per node. Significance was set at P < 0.005 (two-tailed) for group comparison. 2.6.2. Network filtration to yield all-threshold correlation of density based on a persistent homological framework To construct brain networks of MRI density across all thresholds, where r was the correlation coefficient between the gray matter density in VOIs with age as a nuisance variable, the distance matrix (1-r) was used to generate a barcode, single linkage dendrogram, and single linkage matrix for each group (hearing adults, pre-LD, and post-LD) based on persistent homological perspectives (Lee et al., 2012). When the network was filtered, two nodes were considered to be connected only if they had a shorter distance than the threshold. By increasing the threshold distances, more nodes were allowed to be connected to each other (Lee et al., 2011). The barcode represented changes in the number of connected components of the filtered network when the threshold was varied during network filtration. In the barcode, each connected component is represented by a bar which starts and ends at the filtration values when the corresponding connected component appears and disappears. The number of connected components is monotonically decreasing function for all filtration values. Its maximum value is equal to the number of nodes (number of VOIs) at the filtration value zero, and its minimum value is one when all nodes are connected. If bars in the barcode are rearranged according to the location of the VOIs, it becomes a dendrogram which represents the hierarchical clustering of brain regions. A single linkage matrix is the matrix representation of a single linkage dendrogram, and was used to disclose the statistical differences between the hearing adults, pre-LD, and post-LD groups. The single linkage matrix shows the local change of connected structure of brain network during filtration. Since its element represents the filtration value when two brain regions belong to the same connected component for the first time, we can find which areas are connected earlier than the other regions. A nonparametric, unpaired permutation test was used to obtain the probability that the observed difference between two groups occurred by chance (the null hypothesis) by generating the pseudo-matrices of the single linkage matrix. To yield two pseudo-group data (pseudohearing adults and pseudo-pre-LD), individuals were randomly chosen from data of mixed population, and single linkage distances between nodes i and j (dij) were calculated for each pseudo-group, with partialling out corresponding age. For each pseudo-group's data, single linkage distances were converted to z-transformed metrics following the equation, zij ¼ 1/2 log [(2 dij)/(dij)]. These z-
91
values were then compared between groups by z statistics using Zdiffij ¼ (zij1 zij2)/√(1/(n13)þ1/(n23)), where n1 and n2 represent the number of subjects in each group. This procedure was repeated 5000 times to generate the null distribution of Zdiffij, which represents a group difference between single linkage matrices. Significance was set at P < 0.001 (two-tailed) for group comparison. 3. Results 3.1. Gray matter density in pre-LD and post-LD compared with hearing adults After correcting for multiple comparisons via Bonferroni correction and the False Discovery Rate (FDR) correction (Benjamini and Hochberg, 1995), there was no significant difference between each deaf group and hearing adults. However, when analyzed using the uncorrected P < 0.005 (k > 50), the gray matter density of bilateral A1 was preserved in the pre-LD group (Fig. 1A), but decreased in the post-LD group (Fig. 1B). There was no significant correlation between the density of bilateral A1 and duration or onset of deafness in both groups. 3.2. Correlation of bilateral superior temporal regions with A1 density in pre-LD and post-LD compared with hearing adults The left and right A1s were correlated positively with bilateral superior temporal regions, including A1, in all three groups (Supplementary Fig. 1). Group comparisons between the pre- or post-LD groups and the hearing adults showed a differential correlation using the uncorrected P < 0.005 (k > 50). Compared with the hearing adults, the pre-LD group showed increased correlation coefficients of left A1 with the bilateral superior temporal gyri, superior temporal pole, middle temporal gyri, insulae, and left inferior frontal operculum (Fig. 2A, upper row). Also, the pre-LD group showed similarly increased correlation coefficients of right A1 with the bilateral superior temporal gyri, right superior temporal pole, bilateral middle temporal gyri, and insulae as compared with the hearing adults (Fig. 2B, upper row). In contrast, there were no significant differences between the post-LD group and the hearing adults with regard to the correlation between the bilateral A1s and other brain regions (Fig. 2A and B, lower rows). 3.3. Network characteristics of pre-LD, post-LD, and hearing adults using graph theory The distribution of the nodal properties of the network characteristics based on graph theoretical analysis yielded unique patterns in the hearing adults (Fig. 3, left column). In terms of the clustering coefficient (Fig. 3A), betweenness centrality (Fig. 3B), and nodal efficiency (Fig. 3C), the pre-LD group showed a different distribution from the hearing adults (Fig. 3, middle column). In contrast, the post-LD group showed a similar distribution of nodal properties to the hearing adults (Fig. 3, right column). In the pre-LD group, the nodal distribution of clustering coefficients was shifted to a higher value than that of hearing adults (P < 0.001), indicating there were more nodes with greater clustering coefficients in the pre-LD group. In addition, more nodes with higher betweenness centrality were found in the pre-LD group than in the hearing adults (P < 0.05). The skewness of betweenness centrality of the pre-LD group was also higher than that of the post-LD or the hearing adults (pre-LD; 1.43, post-LD; 0.68, hearing adults; 0.62). The nodal efficiency of the pre-LD group spread into the upper and lower extremes of the
92
E. Kim et al. / Hearing Research 315 (2014) 88e98
Fig. 1. Results of voxel-based morphometry. Compared with the hearing adults, (A) the density of bilateral A1 was preserved in the pre-LD group, while (B) the density of bilateral A1 was decreased in the post-LD group. Blue represents the areas where density was lower in the deaf groups than the hearing adults, while red represents the areas where density was higher in the deaf groups than the hearing adults. The A1 region is overlaid in purple. The statistical threshold was set as P value
