Individualized Cortical Parcellation Based on Diffusion ...

11
© The Author(s) 2019. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: [email protected]. Cerebral Cortex, 2019;00: 1–11 doi: 10.1093/cercor/bhz303 Advance Access Publication Date: Original Article ORIGINAL ARTICLE Individualized Cortical Parcellation Based on Diffusion MRI Tractography Meizhen Han 1,2,3 , Guoyuan Yang 1,2,3 , Hai Li 2,3,4 , Sizhong Zhou 1,2,3 , Boyan Xu 1,2,3 , Jun Jiang 1,2,3 , Weiwei Men 1,2,3 , Jianqiao Ge 1,2,3 , Gaolang Gong 5 , Hesheng Liu 6,7 and Jia-Hong Gao 1,2,3 1 Beijing City Key Lab for Medical Physics and Engineering, Institute of Heavy Ion Physics, School of Physics, Peking University, Beijing 100871, China, 2 Center for MRI Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing 100871, China, 3 McGovern Institute for Brain Research, Peking University, Beijing 100871, China, 4 Beijing Intelligent Brain Cloud Inc., Beijing 100036, China, 5 National Key Laboratory of Cognitive Neuroscience and Learning, School of Brain and Cognitive Sciences, Beijing Normal University, Beijing 100875, China, 6 Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Charlestown, MA 02129, USA, and 7 Beijing Institute for Brain Disorders, Capital Medical University, Beijing 100069, China Address correspondence to Jia-Hong Gao, Center for MRI Research, Peking University, Beijing 100871, China. Email: [email protected]; Hesheng Liu, Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Charlestown, MA 02129, USA. Email: [email protected] Abstract The spatial topological properties of cortical regions vary across individuals. Connectivity-based functional and anatomical cortical mapping in individuals will facilitate research on structure–function relationships. However, individual-specific cortical topographic properties derived from anatomical connectivity are less explored than those based on functional connectivity. We aimed to develop a novel individualized anatomical connectivity-based parcellation framework and investigate individual differences in spatial topographic features of cortical regions using diffusion magnetic resonance imaging (dMRI) tractography. Using a high-quality, repeated-session dMRI dataset (42 subjects, 2 sessions per subject), cortical parcels were derived through in vivo anatomical connectivity-based parcellation. These individual-specific parcels demonstrated good within-individual reproducibility and reflected interindividual differences in anatomical brain organization. Connectivity in these individual-specific parcels was significantly more homogeneous than that based on the group atlas. We found that the position, size, and topography of these anatomical parcels were highly variable across individuals and demonstrated nonredundant information about individual differences. Finally, we found that intersubject variability in anatomical connectivity was correlated with the diversity of anatomical connectivity patterns. Overall, we identified cortical parcels that show homogeneous anatomical connectivity patterns. These parcels displayed marked intersubject spatial variability,which may be used in future functional studies to reveal structure–function relationships in the human brain. Key words: brain atlas, connectivity-based parcellation, diffusion tractography, individual variability Downloaded from https://academic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing Normal University Library user on 17 December 2019

Transcript of Individualized Cortical Parcellation Based on Diffusion ...

Page 1: Individualized Cortical Parcellation Based on Diffusion ...

© The Author(s) 2019. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: [email protected].

Cerebral Cortex, 2019;00: 1–11

doi: 10.1093/cercor/bhz303Advance Access Publication Date:Original Article

O R I G I N A L A R T I C L E

Individualized Cortical Parcellation Based on DiffusionMRI TractographyMeizhen Han1,2,3, Guoyuan Yang1,2,3, Hai Li2,3,4, Sizhong Zhou1,2,3,Boyan Xu1,2,3, Jun Jiang1,2,3, Weiwei Men1,2,3, Jianqiao Ge1,2,3,Gaolang Gong5, Hesheng Liu6,7 and Jia-Hong Gao1,2,3

1Beijing City Key Lab for Medical Physics and Engineering, Institute of Heavy Ion Physics, School of Physics,Peking University, Beijing 100871, China, 2Center for MRI Research, Academy for Advanced InterdisciplinaryStudies, Peking University, Beijing 100871, China, 3McGovern Institute for Brain Research, Peking University,Beijing 100871, China, 4Beijing Intelligent Brain Cloud Inc., Beijing 100036, China, 5National Key Laboratory ofCognitive Neuroscience and Learning, School of Brain and Cognitive Sciences, Beijing Normal University,Beijing 100875, China, 6Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology,Massachusetts General Hospital, Harvard Medical School, Charlestown, MA 02129, USA, and 7Beijing Institutefor Brain Disorders, Capital Medical University, Beijing 100069, China

Address correspondence to Jia-Hong Gao, Center for MRI Research, Peking University, Beijing 100871, China. Email: [email protected]; Hesheng Liu,Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Charlestown,MA 02129, USA. Email: [email protected]

AbstractThe spatial topological properties of cortical regions vary across individuals. Connectivity-based functional and anatomicalcortical mapping in individuals will facilitate research on structure–function relationships. However, individual-specificcortical topographic properties derived from anatomical connectivity are less explored than those based on functionalconnectivity. We aimed to develop a novel individualized anatomical connectivity-based parcellation framework andinvestigate individual differences in spatial topographic features of cortical regions using diffusion magnetic resonanceimaging (dMRI) tractography. Using a high-quality, repeated-session dMRI dataset (42 subjects, 2 sessions per subject),cortical parcels were derived through in vivo anatomical connectivity-based parcellation. These individual-specific parcelsdemonstrated good within-individual reproducibility and reflected interindividual differences in anatomical brainorganization. Connectivity in these individual-specific parcels was significantly more homogeneous than that based on thegroup atlas. We found that the position, size, and topography of these anatomical parcels were highly variable acrossindividuals and demonstrated nonredundant information about individual differences. Finally, we found that intersubjectvariability in anatomical connectivity was correlated with the diversity of anatomical connectivity patterns. Overall, weidentified cortical parcels that show homogeneous anatomical connectivity patterns. These parcels displayed markedintersubject spatial variability, which may be used in future functional studies to reveal structure–function relationships inthe human brain.

Key words: brain atlas, connectivity-based parcellation, diffusion tractography, individual variability

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 2: Individualized Cortical Parcellation Based on Diffusion ...

2 Cerebral Cortex, 2019, Vol. 00, No. 00

IntroductionThe human brain is characterized by marked interindividualvariability in many aspects, including macroscopic and micro-scopic brain anatomy (Amunts et al., 1999, 2005; Bürgel et al.1999; Frost and Goebel 2012), functional organization (Muelleret al. 2013; Finn et al. 2015; Wang et al. 2015; Kong et al. 2018; Liet al. 2019), and anatomical connectivity (Anwander et al. 2007).Recent individual-level functional imaging studies have indi-cated that the size, location, and spatial arrangement of corticalfunctional areas vary substantially across individuals, and it hasalso been demonstrated that individual-specific functional net-work topography is predictive of human behavioral phenotypes(Kong et al. 2018; Li et al. 2019). To investigate the spatial topolog-ical properties of cortical regions, one can parcellate the cortexinto subareas based on specific macro- and microanatomicalcriteria (i.e., cytoarchitecture and distribution of neurorecep-tors). Although cytoarchitectonic areas derived from histologi-cal examination have provided invaluable information (Amuntset al. 1999, 2007, 2010; Amunts and Zilles 2015), the correspond-ing relationships between cytoarchitectonic areas and functioncannot be measured in individuals in vivo. One alternative cri-terion for parcellating cortical areas is the unique connectivitycharacteristics of each cortical area, for example, functionalconnectivity provided by resting-state functional MRI (rs-fMRI)(Johansen-Berg et al. 2004; Anwander et al. 2007; Yeo et al.2011; Shen et al. 2013; Wang et al. 2015; Fan et al. 2016; Glasseret al. 2016; Gordon et al. 2017; Schaefer et al. 2017) or anatomicalconnectivity provided by diffusion magnetic resonance imaging(dMRI). Although functional regions defined by rs-fMRI showgood correspondence with task-induced activation and invasivecortical stimulation mapping (Wang et al. 2015; Fox et al. 2016),the exact structure–function relationship remains poorly under-stood (Langs et al. 2016). Recent studies have indicated thatcortical areas specified using dMRI tractography show a highdegree of overlap with cytoarchitectonic areas (Johansen-Berget al. 2004; Mars et al. 2011; Henssen et al. 2016) and thatanatomical connectivity fingerprints can be linked to functionalactivity in many cognitive tasks (Saygin et al., 2012, 2016; Osheret al. 2016). Thus, anatomical connectivity-based cortical map-ping at the individual level may provide critical information forunderstanding the relation between anatomical and functionalbrain organization. To date, an individual-specific, anatomicalconnectivity-based whole-cortex parcellation is not yet avail-able, and the individual differences in anatomical connectivitycharacteristics remain unexplored.

In this study, we aimed to develop a novel framework to par-cellate the cerebral cortex at the individual level using anatomi-cal connectivity profiles derived from dMRI tractography, andwe used these anatomical regions to investigate individualdifferences in brain organization. We adopted a population-level anatomical connectivity-based atlas, known as the HumanBrainnetome Atlas (Fan et al. 2016), as the starting point ofan iterative, individualized parcellation procedure, which issimilar to the functional parcellation procedure proposed byWang et al. (2015). We leveraged a high-quality, repeated-sessiondMRI dataset (n = 42), which allowed us to differentiate theintrasubject variability caused by technical instability fromthe intersubject variability. Individualized connectivity-basedparcellation was performed independently using the first andsecond scans of each subject. Connectivity homogeneity wasquantified in each parcel. We then examined the intersubjectvariability in the position, size, and overall spatial topography of

connectivity regions. In contrast to previous studies that focusedon individual differences in regional fractional anisotropy, meandiffusivity, and the geometry of macrostructural white matter(Oishi et al. 2008; Lee et al. 2009; Yendiki et al. 2011; Veenithet al. 2013; Giordano et al. 2017), the present study aims toinvestigate individual differences in spatial features of theanatomical connectivity-based cortical regions. Finally, thesemaps of intersubject variability were compared to the diversityof connectivity patterns quantified at each cortical region.

Materials and MethodsParticipants

Forty-two right-handed healthy subjects (19 males, age (mean ±standard deviation), 21.6 ± 2.5; range 19–30 years) were recruitedto participate in the experiment. All participants underwent2 dMRI scanning sessions at an interval of 10 months. Allparticipants provided written informed consent prior to theexperiment. The study was approved by the Peking UniversityInstitutional Review Board.

Data Acquisition

All imaging data were collected at Peking University using a 3 TSiemens Prisma MRI scanner (Siemens Healthineers, Erlangen,Germany) equipped with a 64-channel head coil.

The dMRI images were acquired using a 2D spin-echosingle-shot multiband echo-planar imaging (EPI) sequencewith a multiband factor of 3 and a monopolar gradientpulse. Parameters for the dMRI scan were as follows: slicethickness = 1.5 mm without gap; 100 axial slice coverage ofthe whole brain; repetition time (TR) = 3500 ms; echo time(TE) = 86 ms; flip angle (FA) = 90◦; acquisition matrix = 140 × 140;and field of view (FOV) = 210 × 210 mm2. A full dMRI sessionincluded 2 runs with 2 different gradient tables, collectedin opposite (anterior-to-posterior and posterior-to-anterior)phase-encoding polarities (Andersson et al. 2003). We usedthe gradient tables published in the Connectome CoordinationFacility protocol by the Human Connectome Project (HCP) group.In each gradient table, there were either 92 or 93 diffusion-weighted directions and 7 b = 0 s/mm2 acquisitions (b0 images)spread throughout each run uniformly. Diffusion weighting wascomposed of 2 shells of b = 1000 and 2000 s/mm2, with nearlyequivalent numbers of acquisitions distributed on each shell ineach run.

The structural T1-weighted images were collected with a3D magnetization-prepared rapid acquisition with gradientecho sequence (sagittal plane; TR = 2400 ms; TE = 2.22 ms;inversion time (TI) = 1000 ms; iPAT = 2; 0.8 mm isotropic voxels,224 interleaved slices; FOV = 256 × 240 mm2; FA = 8◦; total scantime = 6 min 36 s).

Data Preprocessing

The structural images were first processed with the HCP Pre-FreeSurfer pipeline, which included gradient distortion correc-tion, alignment, brain extraction (Smith 2002), readout distortioncorrection, and bias field correction, followed by registration tostandard Montreal Neurological Institute space (Collins et al.1994). Then, the FreeSurfer toolbox was used to produce tissuemaps and surface files for both the pia mater and the whitematter for each individual.

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 3: Individualized Cortical Parcellation Based on Diffusion ...

Individualized Cortical Parcellation Based Han et al. 3

The diffusion images were preprocessed by the HCP Diffusionpreprocessing pipeline using the FMRIB diffusion toolbox (FSL5.0; http://www.fmrib.ox.ac.uk/fsl). The diffusion preprocessingpipeline begins by intensity normalization of the average b0images in all diffusion successions. After that, EPI susceptibility-induced field distortions were calculated using the b0 images ofboth phase-encoding directions (i.e., anterior-to-posterior andposterior-to-anterior) with the “topup” tool (Andersson et al.2003) in FSL 5.0. Next, all distortions, including motion andeddy-current distortions, were corrected utilizing the “eddy” tool(Andersson et al. 2016). The data were then converted fromthe native dMRI space into the native structural space. Finally,the local probability distribution of fiber orientation at eachvoxel was calculated with FSL’s multishell spherical deconvolu-tion toolbox (bedpostx) (Behrens et al. 2007; Jbabdi et al. 2012;Hernández et al. 2013). Here, the computation model allowedautomatic assessment of 2 fiber directions in each voxel, bywhich considerable improvement of the tracking sensitivity ofnondominant fiber populations can be achieved in the humanbrain (Sotiropoulos et al. 2013).

Individualized Connectivity-Based ParcellationFramework

The individualized connectivity-based parcellation approach(Fig. 1) included the following steps:

Step 1: The population-based atlas was projected onto anindividual’s cortical ribbon. For each subject, the gray matterribbon and the voxels located at the gray–white matter interfacewere combined to obtain a cortical ribbon, which served as a per-sonalized cortical atlas mask. Then, the probability maps of 210cortical regions in the group-level Human Brainnetome Atlas,including 105 cortical areas per hemisphere, were separatelyprojected onto the individual subject’s gray matter tissue mapusing forward and inverse nonlinear transformations. Theseprobability-based cortical seeds were used as the initial conjec-ture of the anatomical parcellation of an individual’s cortex.

Step 2: Probabilistic tractography was performed to gener-ate anatomical connectivity profiles of each probability corti-cal seed derived in the first step. The voxel-wise probabilis-tic tractography was performed at the voxels that fell withinthe personalized cortical atlas mask in diffusion data space(native structural space) by sampling 5000 streamline fibersto approximate its whole-brain connectivity pattern (Behrenset al. 2007). Then, a small threshold value was set to remove theconnectivity information of the voxels that were only reached byno more than 2/5000 samples, which were considered as “noise”(Heiervang et al. 2006; Fan et al. 2016). In addition, the whole-brain connectivity pattern of each cortical seed voxel was down-sampled (e.g., 5 mm isotropic voxels) to improve computationalefficiency (Johansen-Berg et al. 2004). Based on the probabilityacross the voxels that fell within the probability map of a corticalseed, the diffusion connectivity pattern of each cortical voxelwas then averaged in a weighted manner; then the atlas-basedconnectivity profile of the corresponding region was derived.These atlas-based connectivity patterns of 210 probability cor-tical seeds were utilized as the initial “reference connectivityprofiles” in the following optimization procedures.

Step 3: Each voxel in an individual’s cortical ribbon wasassigned to one of the regions based on maximal Pearson’scorrelation with the reference connectivity profiles. In orderto improve computational efficiency, the assignment was per-formed separately in the cortices of the 2 hemispheres (i.e., a

Figure 1. Framework of the individualized connectivity-based parcellation

method. The method involved the following steps: (1) The population-basedHuman Brainnetome Atlas was separately projected onto individual corticalribbons. (2) The individual connectivity profile of each cortical seed was derivedusing probabilistic tractography and was utilized as the initial “reference con-

nectivity profile” in the following optimization procedures. (3) Each voxel in anindividual’s cortical ribbon was assigned to 1 of the 105 subregions based onmaximal correlation with the reference connectivity profiles. (4) After assign-ment, a new reference connectivity profile was computed in a weighted manner

for the next iteration. Then, the cortical voxels were further reassigned to 1 ofthe 105 subregions. (5) Step (4) was iterated until the iteration procedure reachedconvergence.

voxel would only be assigned to 1 of the 105 cortical regionsof the hemisphere where it was located). The individualizedconnectivity-based parcellation-optimization procedure in thisstudy is similar to the strategy proposed by Wang and colleagues(Wang et al. 2015). The connectivity profile resulting from prob-abilistic tractography of each voxel in the individual subjectwas correlated to the 105 reference connectivity profiles. Eachvoxel was assigned to 1 of the 105 regions based on maximal

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 4: Individualized Cortical Parcellation Based on Diffusion ...

4 Cerebral Cortex, 2019, Vol. 00, No. 00

correlation with the reference connectivity profiles. In addition,we calculated a confidence value as the ratio between the max-imum and the second largest correlation values.

Step 4: New reference connectivity profiles were generated,and then each voxel in the cortical ribbon was reassigned.After all voxels were assigned to 1 of the 105 subregions in theprevious step, a mean connectivity profile was calculated in eachcortical region, during which the diffusion connectivity profilesof the voxels with a confidence value greater than a preselectedthreshold (here it was set to 1.1) were averaged with greaterweight. This weighting strategy guaranteed that the connectivityprofiles coming from core voxels with higher confidence levelswere weighted more than connectivity profiles of the voxelslocated at the edge of cortical regions. Then, for each region,the mean connectivity profile and the reference connectivityprofile were averaged in a weighted manner again, during whichthe reference connectivity profiles were weighted more than themean connectivity profiles to slow down the convergence speed.The resulting connectivity profile estimate was utilized as thenew reference connectivity profile for the next iteration. Then,the cortical voxels were further reassigned to 1 of the 105 regionsusing these new reference connectivity profiles.

Step 5: The iteration procedure reached convergence. Step 4was iterated until the algorithm reached a predefined stoppingcriterion. Here, the iteration procedure was stopped when theparcellation pattern remained the same for 99% of the voxels in2 consecutive iterations.

This individualized parcellation–optimization iteration pro-cess, using these new reference connectivity profiles in eachiteration, which contained both the population atlas informa-tion and the information of an individual subject, ensured thatthe individual-specific cortical topological features of the con-nectivity units’ organization would gradually emerge from theoriginal population atlas as the iteration proceeded. Ultimately,the anatomical connectivity-based cortical parcellation wouldlocalize 210 cortical regions in each individual.

Intrasubject Reproducibility Estimation

The reproducibility of this individualized connectivity-basedparcellation was evaluated using our repeated-session dataset.Individualized anatomical connectivity-based parcellation wasindependently performed using the first scan and the secondscan, and the intrasubject test–retest reproducibility was eval-uated using Dice’s coefficient (Dice 1945). Reproducibility wasthen averaged across 42 individuals to obtain an overall estima-tion.

For each cortical region, Dice’s coefficient between 2 parcel-lations was calculated as follows:

Dice′s coefficient = 2∣∣A

⋂B∣∣

|A| + |B| (1)

where∣∣∣A

⋂B∣∣∣ denotes the number of voxels in the intersection

of the cortical region between 2 parcellations,∣∣∣A

∣∣∣ represents the

number of voxels of the cortical region in one parcellation, and∣∣∣B

∣∣∣ represents the number of voxels of the same cortical region

in the other parcellation.

Diffusion Connectional Homogeneity Estimation

Determining the validity of individualized connectivity-basedparcellations is difficult due to a lack of ground truth. Here, we

evaluated connectivity homogeneity, which has been commonlyutilized in individual-level functional network parcellation stud-ies (Chong et al. 2017; Gordon et al. 2017; Schaefer et al. 2017;Kong et al. 2018).

The connectivity homogeneity of a pair of voxels withinthe cortical region was calculated as the Pearson’s correlationcoefficient between the voxels’ diffusion connectivity profiles.The overall connectivity homogeneity of a cortical region wascalculated as the mean Pearson’s correlation coefficient acrossall voxel pairs within the cortical region. The evaluation ofhomogeneity across the whole cortex was conducted by averag-ing the homogeneity coefficient across all cortical regions whileaccounting for region size:

Homogeneity coefficient =∑L

l=1 ρl∣∣l∣∣

∑Ll

∣∣l∣∣

(2)

where ρl is the homogeneity coefficient of cortical region l and∣∣∣l∣∣∣

is the number of voxels within it (Schaefer et al. 2017). For eachsubject in the dataset (N = 42), we computed the homogeneityfor both the group-level atlas and individual-level maps. Then, apaired sample t-test was performed between the homogeneitycoefficients of the atlas and individualized maps.

Estimating Intersubject Variability in Position, Size, andOverall Topography of Connectivity Regions

The position of a cortical region (i.e., parcel) in the connectivity-based parcellation was characterized by the 3D coordinates ofits center of mass. For each parcel, the intersubject variabilityin parcel position was calculated as the mean geodesic distanceamong the centers of mass of the parcel across individuals (Liet al. 2019). Intrasubject variability in position was calculated asthe geodesic distance between the centers of mass of the parcelin the 2 sessions. To control for the effects of random noise andtechnical imperfections on intersubject variability measures, wecorrected intersubject variability in position by regressing outthe average intrasubject variability via the general linear model(Mueller et al. 2013).

The size of a parcel in the connectivity-based parcellationwas computed as the number of voxels belonging to that parcel.For each parcel, the intersubject variability in size was computedas the standard deviation of size across individuals. Intrasubjectvariability in size was calculated as the difference in parcelsize between 2 parcellations from 2 different sessions. We alsocorrected the intersubject variability in parcel size by regressingout the average intrasubject variability.

The spatial topography of a cortical region depends on thesize, location, and shape of the region. For each parcel, the inter-subject variability in parcel topography was calculated as themean Dice’s coefficient across all pairs of subjects. Intrasubjectvariability in topography was calculated as Dice’s coefficientbetween the parcels in the 2 sessions. Intersubject variabilityin topography was also corrected by regressing out the averageintrasubject variability.

In addition, as the individualized parcellation was performedin the subject’s native cortical ribbon space, volumetric corticalparcellation of each individual was registered and projected tothe fsaverage surface template prior to subsequent analyses.Relationship between each 2 of the 3 metrics (intersubject vari-ability in position, size and topography of parcels) was assessedusing Pearson’s correlation.

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 5: Individualized Cortical Parcellation Based on Diffusion ...

Individualized Cortical Parcellation Based Han et al. 5

Estimating Intersubject Variability in ConnectivityProfile and the Diversity of Connectivity

For each parcel, the intersubject variability in diffusion con-nectivity profile was calculated as the inverse of the meanPearson’s correlation coefficient of the diffusion connectivityprofiles across all pairs of subjects. Here, we used the classicalJHU white matter template (Mori et al. 2008) to encode thediffusion connectivity profile of each cortical parcel using themethod proposed by Mars and colleagues (Mars et al. 2018). A fullconnectivity profile matrix (210 × 48) describes how each parcelis connected to each white matter tract. Intrasubject variabilityin diffusion connectivity profile was calculated as the inverseof the Pearson’s correlation coefficient between the diffusionconnectivity profiles in the 2 sessions. Intersubject variability indiffusion connectivity profile was also corrected by regressingout the average intrasubject variability.

Furthermore, we hypothesized that cortical regions withhigher intersubject variability in connectivity profile would bemore complex and diversely connected. In order to describethe diversity of structural connectivity, an entropy value (i.e.,whether a cortical parcel is reached strongly by few white mattertracts or equally strongly by all the tracts) was calculated forthe structural connectivity profile of each cortical parcel foreach subject, and the values were then averaged across subjects(Mars et al. 2018). Pearson’s correlation analysis was performedbetween the connectivity entropy and the intersubject variabil-ity in connectivity profile to test the hypothesis that corticalregions with more diverse structural connectivity show higherintersubject variability in connectivity profile. Furthermore,Pearson’s correlation analysis was performed between theconnectivity entropy and the intersubject spatial variability ofcortical regions to test the hypothesis that cortical regions withmore diverse structural connectivity are more variable acrossindividuals.

Correlation to Functional Connectivity Variability andMorphological Variability

To further estimate the relationships between the intersubjectvariabilities in diffusion connectivity and in other modalities(i.e., functional connectivity, sulcal depth, and cortical thick-ness), variability in diffusion connectivity based on evenly dis-tributed cortical seeds was additionally calculated. First, theHuman Brainnetome Atlas was resampled into 1000 evenly dis-tributed cortical seeds. Next, an abovementioned calculationmethod (Mars et al. 2018) was used to calculate the intersubjectvariability in diffusion connectivity at each cortical seed. Afterthat, the intersubject diffusion connectivity variability map wasresampled to the fsaverage 3 template. Then, Pearson’s corre-lation analysis was performed to estimate its relation to func-tional connectivity variability, sulcal depth variability, and cor-tical thickness variability, respectively. Maps of functional con-nectivity variability, sulcal depth variability, and cortical thick-ness variability were provided by Sophia Mueller and colleagues(Mueller et al. 2013).

ResultsConnectivity-Based Parcellation in Individuals

Using a group atlas derived from the HCP dataset (N = 40) as theinitialization, we employed an iterative approach to parcellateeach individual subject’s cerebral cortex based on anatomical

connectivity (Fig. 2 for the atlas and parcellation maps from4 representative subjects). The parcellation was performedindependently using the first and second scans from the sameindividuals. Intrasubject test–retest reproducibility of parcel-lation was evaluated across all cortical regions (i.e., parcels)using Dice’s coefficient. Across all subjects, the mean test–retestreproducibility was 0.72 ± 0.07 (mean ± SD). SupplementaryFigure S1 illustrates the intrasubject reproducibility values inthe 210 parcels. Intrasubject reproducibility was moderate in theregions in the basal frontal and temporal cortex, which are espe-cially liable to MRI EPI susceptibility artifacts. In addition, regionsin the precentral gyrus and medial cortex exhibited relativelylow intrasubject reproducibility. Importantly, individualizedcortical parcellation revealed the unique network topology ofeach subject (Supplementary Figs S2 and S3 for different viewsof the parcellation maps) and reflected individual differences inbrain organization. The mean intersubject similarity of corticalparcellation was 0.39 ± 0.11 (Dice’s coefficient).

Individual-Level Parcellation Yielded Regions withHomogeneous Connectivity

Connectivity homogeneity was evaluated in each parcel andcompared between the individual maps and the group atlas(Fig. 3). The diffusion connectivity homogeneity of each parcelwas averaged across all subjects. Compared to the group atlas,homogeneity estimated in the individual parcellation exhib-ited overall improvement, especially in the bilateral temporallobes. Homogeneity was then compared between the individ-ualized parcellation and the group atlas at each of the 210parcels (Fig. 3B). For visualization purposes, we sorted the cor-tical regions on the x-axis according to increasing homogeneityin the individualized parcellation. For nearly all the parcels (208out of 210), connectivity homogeneity was significantly higher inindividualized parcellation than in the group atlas (two-samplepaired t-test, P < 0.001). Two regions showing higher homogene-ity in the group atlas (the dorsal granular insular and lateralposterior parahippocampal gyrus) have very few voxels in theatlas after the voxels outside the cortical ribbon (white matterand cerebrospinal fluid) are removed, which may lead to a biasedestimate of homogeneity (Gordon et al. 2017; Kong et al. 2018).

The overall homogeneity across the whole cortex was thencalculated by averaging the homogeneity coefficient across allparcels while accounting for parcel size. When averaged across42 subjects, the diffusion connectivity homogeneity values ofthe group atlas and the individual maps were 0.33 ± 0.02 and0.47 ± 0.02, respectively. Compared to the group atlas, the indi-vidual maps showed an improvement of 41.69% (two-samplepaired t-test, P < 0.001) in connectivity homogeneity.

Intersubject Variability in Position, Size, andTopography of Connectivity Regions

The intersubject variability in position, size, and topography wasquantified for each of the 210 anatomical parcels (Fig. 4). Wefound that intersubject variabilities in the 3 metrics all demon-strated a nonuniform distribution across parcels. Intersubjectvariability was high in the lateral prefrontal lobe and the tem-poral–parietal junction and minimal in orbital frontal, insula,motor, and medial cortical areas. Notably, some parcels showeddifferent intersubject variability features among these 3 met-rics. For example, while the primary auditory cortex and mid-dle temporal gyrus displayed strong intersubject variability in

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 6: Individualized Cortical Parcellation Based on Diffusion ...

6 Cerebral Cortex, 2019, Vol. 00, No. 00

Figure 2. The comparison chart of the group atlas derived using the HCP dataset (N = 40) and individual atlases from 4 representative individuals in the lateral side

view and medial side view (visualized using ITK-SNAP (www.itksnap.org)). Individual-specific connectivity-based parcellations were independently estimated usingthe first scan and the second scan with a 10-month interval. By visualization, these cortical parcellations exhibited unique spatial configuration for each individual,which was stably replicated across sessions.

topography (Fig. 4C), the primary auditory cortex showed lowervariability in size of the parcels (Fig. 4B), and the middle tem-poral gyrus showed moderately variable parcel positions acrossindividuals (Fig. 4A). Additionally, while the parietal–occipitaljunction region showed strong intersubject variability in parcelsize, it showed moderately variable position and topographyacross individuals.

The correlation between the variability map of parcel posi-tion and the variability in parcel topography (r = 0.80, P < 0.001)was stronger than the correlation between the variability inparcel size and the position variability (r = 0.53, P < 0.001) or thetopography variability (r = 0.56, P < 0.001) (see Supplementary Fig.S4 for the scatterplots), indicating that although the spatialtopography of a cortical region depends on the size, location,and shape of the region at the same time, the intersubject vari-ability of the spatial topography is affected more by the positiondifference than by the size or shape difference. Collectively,these findings suggest that the size, position, and topography ofanatomical parcels may reflect distinct aspects of intersubjectvariability in brain organization.

Intersubject Spatial Variability of Connectivity RegionsIs Related to Connectivity Diversity

We explored whether regions showing high intersubjectvariability also show diversified anatomical connectivity.Connectivity diversity was evaluated using connectivity entropy(see Methods). High structural connectivity entropy was foundin the parietal lobe, the parietal–occipital junction, the superiorprefrontal cortex, the posterior cingulate gurus, and the insula,which is consistent with the voxel-wise structural connectivityentropy measures reported by Mars et al. (2018). We found thatthe intersubject variability in diffusion connectivity profileshowed a moderate correlation with diffusion connectivityentropy (r = 0.18, P = 0.009) (Fig. 5, also see Supplementary Fig. S5for the scatterplots). More specifically, structural connectivityentropy demonstrated a moderate correlation with intersubjectvariability in parcel position (r = 0.14, P = 0.046), parcel size(r = 0.29, P < 0.001), and parcel topography (r = 0.17, P = 0.015)across the entire cerebral cortex. These observations suggestthat regions showing higher interindividual variability in

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 7: Individualized Cortical Parcellation Based on Diffusion ...

Individualized Cortical Parcellation Based Han et al. 7

Figure 3. Greater diffusion connectivity homogeneity was observed in the individual atlas than in the group atlas. (A) Diffusion connectivity homogeneity maps of the

group atlas and individual atlas. The diffusion connectivity homogeneity of each cortical area was averaged across all subjects. (B) The x-axis includes a list of corticalareas, sorted from lower to greater homogeneity in the individual atlas. The thin solid line represents the mean homogeneity of 42 subjects at each cortical area. Theshaded area indicates the standard deviation (S.D.) across the 42 subjects. At most of the cortical areas (208 areas), with the exception of 2 areas, the homogeneity was

significantly higher than in the group atlas if the cortical area was individually specified (two-sample paired t-test, P < 0.001). (C) At the whole cortex level, the diffusionconnectivity homogeneity of the individual atlas (0.47 ± 0.02) was significantly higher than that of the group atlas (0.33 ± 0.02, two-sample paired t-test, P < 0.001).∗ P < 0.001.

Figure 4. Intersubject variability in the position, size, and topography of connectivity-based regions. Intersubject variability in (A) parcel position, (B) parcel size, and (C)

parcel topography was measured for each of the 210 parcels. The variability map of parcel position showed a strong correlation (r = 0.53, P < 0.001) with the variabilityin parcel size and an even stronger correlation (r = 0.80, P < 0.001) with the variability in parcel topography. Additionally, the variability map of parcel size showed astrong correlation (r = 0.56, P < 0.001) with the variability in parcel topography (see Supplementary Figure S4 for the scatterplots). All intersubject variability maps werecorrected by the underlying intrasubject variability.

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 8: Individualized Cortical Parcellation Based on Diffusion ...

8 Cerebral Cortex, 2019, Vol. 00, No. 00

Figure 5. Diffusion connectivity variability is positively associated with connectivity diversity, and both are positively related to the intersubject spatial variability ofconnectivity regions. Diffusion connectivity variability was quantified for each of the 210 cortical parcels, and the variability map (A) showed a moderate but significantcorrelation with diffusion connectivity entropy (B) (r = 0.18, P = 0.009). The diffusion connectivity variability showed significant correlations with intersubject variabilityin parcel position (r = 0.30, P < 0.001), parcel size (r = 0.23, P = 0.001), and parcel topography (r = 0.26, P < 0.001) across the entire of cerebral cortex. Further, the diffusion

connectivity entropy demonstrated significant correlations with intersubject variability in parcel position (r = 0.14, P = 0.046), parcel size (r = 0.29, P < 0.001), and parceltopography (r = 0.17, P = 0.015) across the entire cerebral cortex (see Supplementary Figure S5 for the scatterplots).

diffusion connectivity tend to have more diversified connec-tivity patterns.

DiscussionIn this study, we parcellated the cerebral cortex based onanatomical connectivity and investigated individual differencesin the spatial topographic features of cortical parcels. We foundthat the individualized parcellation approach can yield corticalparcels with good reproducibility and that the parcellation canreflect individual differences in brain organization. Further-more, the position, size, and topography of these anatomicalparcels demonstrated substantial intersubject variability. Theintersubject variability of diffusion connectivity may also berelated to the diversity of connectivity patterns. Collectively,our study provided an approach to mapping individual subjects’cortical regions based on anatomical connectivity, which mayfacilitate future in-depth investigations of structure–functionrelationships.

Revealing Individual Variability in CorticalOrganization Using Anatomical Connectivity

Traditional research methods in the field of neuroscience mainlyrely on group-level analysis to infer general principles. How-ever, recent studies, especially those based on rs-fMRI, havedemonstrated the importance of accounting for individual dif-ferences in brain organization when relating imaging featuresto behavior or clinical symptoms. Mueller et al. found that theintersubject variability in functional connectivity is a funda-mental property of the human brain and may be a result ofevolutionary expansion of the cortex (Mueller et al. 2013). Finnet al. showed that an individual subject’s functional connectivitypattern is unique and can be viewed as the “fingerprint,” whichcan predict cognitive abilities such as fluid intelligence (Finnet al. 2015). Recently, a few studies further found that not onlythe strength of connectivity but also the topographic features of

brain regions, such as the size, shape, and location, are highlyvariable across individuals and are behaviorally relevant (Konget al. 2018; Li et al. 2019). In contrast to the large amount ofwork on mapping individual-specific functional cortical net-works (Wang et al. 2015; Gordon et al. 2017; Kong et al. 2018;Li et al. 2019), an individualized anatomical connectivity-basedwhole-cortical parcellation technique is lacking. In the presentstudy, we proposed an individualized parcellation strategy thatcan capture the individual difference in the spatial arrangementof anatomical regions. A reliable structural connectivity-basedparcellation with high sensitivity to individual differences willallow one to investigate functional and anatomical propertiesin these parcels within individuals, thus facilitating research onthe associations between anatomy and function. Future studiesmay also reveal how these associations change during braindevelopment and in diseases.

The Association Between the OrganizationCharacteristics of Microscale Cytoarchitectonics andMacroscale Connectomics

A better understanding of the relationship between microstruc-ture and connectivity is needed to advance the research ofhuman brain organization. Previous studies have indicated thatconnectivity-based cortical parcellation closely follows histo-logical subregions (Anwander et al. 2007; Klein et al. 2007). Ina recent study, researchers conducted cross-scale analysis andfound a significant association between the macroscale struc-tural connectivity obtained by dMRI data and human corticalcytoarchitectonic features, especially the size of cortical layer 3neurons (van den Heuvel et al. 2015). Furthermore, in situationswhere the results could be compared to existing cytoarchitec-ture maps, high overlap was observed between cytoarchitectonicareas and cortical areas specified using anatomical connec-tivity profiles derived with dMRI tractography (Johansen-Berget al. 2004; Mars et al. 2011; Henssen et al. 2016). This previousevidence indicates that the spatial characteristics of cortical

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 9: Individualized Cortical Parcellation Based on Diffusion ...

Individualized Cortical Parcellation Based Han et al. 9

areas are consistent between the 2 modalities. In this study, theintersubject variability in cortical area topography derived fromstructural connectivity was consistent with the intersubjectvariability identified by cytoarchitectonic parcellations (Amuntset al., 1999, 2004, 2005; Zilles and Amunts 2013; Bludau et al.2014; Henssen et al. 2016). For example, reduced intersubjectvariability in the orbitofrontal cortex compared to the inferiorfrontal gyrus was supported by the cytoarchitectonic probabilitymaps of Fo1, Fo2, Fo3, and BA44/45 (Amunts et al. 2004; Zillesand Amunts 2013; Henssen et al. 2016). In addition, consistentlyreduced intersubject topographic variability was found in thefrontopolar cortex, presubiculum, and posterior insula, whereasconsistently increased intersubject topographic variability wasnoted in the inferior parietal lobe and superior parietal lobeboth in histologically defined cortical areas and in diffusionconnectivity-based subdivisions. These findings provided fur-ther evidence for the association between the organization char-acteristics of microscale cytoarchitectonics and macroscale con-nectomics.

Anatomical Connectivity of High IntersubjectVariability Region Is Diverse

An interesting observation in our study is that the anatomi-cal connectivity of regions with higher intersubject variabilitytend to have more diverse connectivity. Previous research onintersubject variability in functional connectivity has demon-strated the positive association between distant connectivityand intersubject functional variability (Mueller et al. 2013). Here,we showed a moderate association between the diversity ofdiffusion connectivity and the intersubject variability of con-nectivity. Increased individual variability and increased struc-tural connectivity entropy were observed in the parietal lobe,the parietal–occipital junction, the superior prefrontal cortex,the posterior cingulate gyrus, and the insula, which may beconsistent with the voxel-wise structural connectivity entropymeasures (Mars et al. 2018). These observations indicate thatbrain regions with high intersubject variability in anatomicalconnectivity may have the hub-like property and have diverseconnectivity profiles.

Toward a Multimodal Cortical Atlas

To date, various brain maps based on different modalities (i.e.,structure, function, and connectivity) have been developed.While many features of these modalities are closely relatedto each other, the exact relationships are yet to be unveiled.This question requires further research on the convergenceor divergence of neurobiological measurements of brainstructure, function, and connectivity (Eickhoff et al. 2017). Asthe classical brain parcellation is focused on the average brain,interindividual differences in cortical area arrangement may beobscured; thus, it is necessary to compare cortical parcellationsof different modalities at the individual level. Recent studiesalso suggested that the functional parcellation of the humancortex may be state dependent and reconfigures substantiallyaccording to task load (Salehi et al. 2018), which calls for anindividualized anatomical map to further explore whether thefunctional reconfiguration process is constrained by anatomicalconnectivity.

We also compared intersubject variability in diffusion con-nectivity and variability in functional connectivity, variability incortical thickness, and sulcal depth but found no correlation

(see Supplemental Material and Fig. S6). This divergence maybe due to the fact that anatomical features are not stronglytied to functional features. For example, a lack of associationbetween functional connectivity variability and cortical thick-ness variability has been previously reported (Mueller et al.2013). We speculated that the association between functionalconnectivity and anatomical connectivity may also vary acrossthe cortical mantle. These preliminary observations may sug-gest future research on exploring the interindividual differencesin function–anatomy associations across the brain.

Limitations and Future Directions

A potential limitation of this study is an absence of an invivo method to cross-validate the locations and boundaries ofthe connectivity-based parcellations. In contrast to parcellat-ing cortical functional networks at the individual level, duringwhich within-individual validation can be performed in com-bination with task fMRI activation or intraoperative electricalstimulation, the accuracy of the locations and boundaries ofthese individual-level parcellations cannot be verified in vivo.Future studies should verify the accuracy of some cortical areasin individualized parcellations by ex vivo anatomical methodsusing postmortem brains.

In addition, as the intra- and intersubject variability mea-sures could depend on the level of granularity in cortical par-cellation, the measures derived from this study only apply tothe specific parcellation scheme (210 parcels from Human Brain-netome Atlas) and should not be simply generalized to otherscales, and comparisons across studies should take the level ofgranularity into account.

ConclusionsIn this study, we parcellated the cerebral cortex based onanatomical connectivity in individuals and investigated indi-vidual differences in the spatial topographic features ofcortical parcels. Our data provide evidence that the corticalparcellations based on anatomical connectivity informationdisplayed marked intersubject variability in parcel position, size,and topography. This study provided an approach to mappingindividual subjects’ cortical regions based on anatomicalconnectivity, which may facilitate future in-depth investigationson structure–function relationships.

Supplementary MaterialSupplementary material is available at Cerebral Cortex online.

FundingNational Key Research and Development Program of China(2017YFC0108901); the National Natural Science Foundationof China (81790651, 81790650, 81727808, 81430037, 81627901,31771253 and 31421003); the Beijing Municipal Science & Tech-nology Commission (Z171100000117012; Z181100001518005);and the Beijing Brain Initiative of Beijing Municipal Science& Technology Commission (Z181100001518003).

NotesWe thank National Center for Protein Sciences at PekingUniversity in Beijing, China, for the assistance with MRI data

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 10: Individualized Cortical Parcellation Based on Diffusion ...

10 Cerebral Cortex, 2019, Vol. 00, No. 00

acquisition. We thank Beijing Intelligent Brain Cloud, Inc.,for assistance with MRI data management and data analysis.Conflict of Interest: The authors declare that the research was con-ducted in the absence of any commercial or financial relation-ships that could be construed as a potential conflict of interest.

ReferencesAmunts K, Kedo O, Kindler M, Pieperhoff P, Mohlberg H, Shah

NJ, Habel U, Schneider F, Zilles K. 2005. Cytoarchitectonicmapping of the human amygdala, hippocampal region andentorhinal cortex: intersubject variability and probabilitymaps. Anat Embryol. 210:343–352.

Amunts K, Lenzen M, Friederici AD, Schleicher A, Morosan P,Palomero-Gallagher N, Zilles K. 2010. Broca’s region: novelorganizational principles and multiple receptor mapping.PLoS Biol. 8:1–16.

Amunts K, Schleicher A, Bürgel U, Mohlberg H, Uylings HBM,Zilles K. 1999. Broca’s region revisited: cytoarchitecture andintersubject variability. J Comp Neurol. 412:319–341.

Amunts K, Schleicher A, Zilles K. 2007. Cytoarchitecture ofthe cerebral cortex - more than localization. NeuroImage.37:1061–1065.

Amunts K, Weiss PH, Mohlberg H, Pieperhoff P, Eickhoff S, GurdJM, Marshall JC, Shah NJ, Fink GR, Zilles K. 2004. Analysis ofneural mechanisms underlying verbal fluency in cytoarchi-tectonically defined stereotaxic space-the roles of Brodmannareas 44 and 45. NeuroImage. 22:42–56.

Amunts K, Zilles K. 2015. Architectonic mapping of the humanbrain beyond Brodmann. Neuron. 88:1086–1107.

Andersson JLR, Skare S, Ashburner J. 2003. How to correct suscep-tibility distortions in spin-echo echo-planar images: applica-tion to diffusion tensor imaging. NeuroImage. 20:870–888.

Andersson JLR, Sotiropoulos SN. 2016. An integrated approach tocorrection for off-resonance effects and subject movement indiffusion MR imaging. NeuroImage. 125:1063–1078.

Anwander A, Tittgemeyer M, Von Cramon DY, Friederici AD,Knösche TR. 2007. Connectivity-based parcellation of Broca’sarea. Cereb Cortex. 17:816–825.

Behrens TEJ, Berg HJ, Jbabdi S, Rushworth MFS, Woolrich MW.2007. Probabilistic diffusion tractography with multiple fibreorientations: what can we gain? NeuroImage. 34:144–155.

Bludau S, Eickhoff SB, Mohlberg H, Caspers S, Laird AR, FoxPT, Schleicher A, Zilles K, Amunts K. 2014. Cytoarchitecture,probability maps and functions of the human frontal pole.NeuroImage. 93:260–275.

Bürgel U, Schormann T, Schleicher A, Zilles K. 1999. Mapping ofhistologically identified long fiber tracts in human cerebralhemispheres to the MRI volume of a reference brain: positionand spatial variability of the optic radiation. NeuroImage.10:489–499.

Chong M, Bhushan C, Joshi AA, Choi S, Haldar JP, Shattuck DW,Spreng RN, Leahy RM. 2017. Individual parcellation of restingfMRI with a group functional connectivity prior. NeuroImage.156:87–100.

Collins DL, Neelin P, Peters TM, Evans AC. 1994. Automatic 3Dintersubject registration of MR volumetric data in standard-ized Talairach space. J Comput Assist Tomogr. 18:192–205.

Dice LR. 1945. Measures of the amount of ecologic associationbetween species. Ecology. 26:297–302.

Eickhoff SB, Constable RT, Yeo BTT. 2017. Topographic organiza-tion of the cerebral cortex and brain cartography. NeuroImage.170:332–347.

Fan L, Li H, Zhuo J, Zhang Y, Wang J, Chen L, Yang Z, Chu C, XieS, Laird AR et al. 2016. The human brainnetome atlas: a newbrain atlas based on connectional architecture. Cereb Cortex.26:3508–3526.

Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM,Papademetris X, Constable RT. 2015. Functional connectomefingerprinting: identifying individuals using patterns of brainconnectivity. Nat Neurosci. 18:1664–1671.

Fox MD, Qian T, Madsen JR, Wang D, Li M, Ge M, Zuo H, GroppeDM, Mehta AD, Hong B et al. 2016. Combining task-evoked andspontaneous activity to improve pre-operative brain map-ping with fMRI. NeuroImage. 124:714–723.

Frost MA, Goebel R. 2012. Measuring structural-functionalcorrespondence: spatial variability of specialised brainregions after macro-anatomical alignment. NeuroImage.59:1369–1381.

Giordano C, Zappalà S, Kleiven S. 2017. Anisotropic finite ele-ment models for brain injury prediction: the sensitivity ofaxonal strain to white matter tract inter-subject variability.Biomech Model Mechanobiol. 16:1269–1293.

Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwel J,Yacoub E et al. 2016. A multi-modal parcellation of humancerebral cortex. Nature. 536:171–178.

Gordon EM, Laumann TO, Gilmore A, Petersen SE, Nelson SM,Dosenbach NUF. 2017. Precision functional mapping of indi-vidual human brains. Neuron. 95:791–807.

Heiervang E, Behrens TEJ, Mackay CE, Robson MD, Johansen-BergH. 2006. Between session reproducibility and between subjectvariability of diffusion MR and tractography measures. Neu-roImage. 33:867–877.

Henssen A, Zilles K, Palomero-Gallagher N, Schleicher A,Mohlberg H, Gerboga F, Eickhoff SB, Bludau S, Amunts K. 2016.Cytoarchitecture and probability maps of the human medialorbitofrontal cortex. Cortex. 75:87–112.

Hernández M, Guerrero GD, Cecilia JM, García JM, Inuggi A,Jbabdi S, Behrens TEJ, Sotiropoulos SN. 2013. Accelerat-ing fibre orientation estimation from diffusion weightedmagnetic resonance imaging using GPUs. PLoS One. 8:e61892.

Jbabdi S, Sotiropoulos SN, Savio AM, Graña M, Behrens TE. 2012.Model-based analysis of multi-shell diffusion MR data fortractography: how to get over fitting problems. Magn ResonMed. 68:1846–1855.

Johansen-Berg H, Behrens TEJ, Behrens TEJ, Robson MD, RobsonMD, Drobnjak I, Drobnjak I, Rushworth MFS, Rushworth MFS,Brady JM et al. 2004. Changes in connectivity profiles definefunctionally distinct regions in human medial frontal cortex.Proc Natl Acad Sci U S A. 101:13335–13340.

Klein JC, Behrens TEJ, Robson MD, Mackay CE, Higham DJ,Johansen-Berg H. 2007. Connectivity-based parcellation ofhuman cortex using diffusion MRI: establishing reproducibil-ity, validity and observer independence in BA 44/45 andSMA/pre-SMA. NeuroImage. 34:204–211.

Kong R, Li J, Orban C, Sabuncu MR, Liu H, Schaefer A, Sun N, ZuoXN, Holmes A, Eickhoff SB et al. 2018. Spatial topography ofindividual-specific cortical networks predicts human cogni-tion, personality and emotion. Cereb Cortex. 29:2533–2551.

Langs G, Wang D, Golland P, Mueller S, Pan R, Sabuncu MR, SunW, Li K, Liu H. 2016. Identifying shared brain networks inindividuals by decoupling functional and anatomical vari-ability. Cereb Cortex. 26:4004–4014.

Lee C, Danielian L, Thomasson D, Baker E. 2009. Normal regionalfractional anisotropy and apparent diffusion coefficient of

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019

Page 11: Individualized Cortical Parcellation Based on Diffusion ...

Individualized Cortical Parcellation Based Han et al. 11

the brain measured on a 3 T MR scanner. Neuroradiology.51:3–9.

Li M, Wang D, Ren J, Langs G, Stoecklein S, Brennan BP, Lu J,Chen H, Liu H. 2019. Performing group-level functional imageanalyses based on homologous functional regions mapped inindividuals. PLoS Biol. 17:e2007032.

Mars RB, Jbabdi S, Sallet J, O’Reilly JX, Croxson PL, Olivier E,Noonan MP, Bergmann C, Mitchell AS, Baxter MG et al. 2011.Diffusion-weighted imaging tractography-based parcellationof the human parietal cortex and comparison with humanand macaque resting-state functional connectivity. J Neurosci.31:4087–4100.

Mars RB, Passingham RE, Jbabdi S. 2018. Connectivity finger-prints: from areal descriptions to abstract spaces. Trends CognSci. 22:1026–1037.

Mori S, Oishi K, Jiang H, Jiang L, Li X, Akhter K, Hua K, Faria AV,Mahmood A, Woods R et al. 2008. Stereotaxic white matteratlas based on diffusion tensor imaging in an ICBM template.NeuroImage. 40:570–582.

Mueller S, Wang D, Fox MD, Yeo BTT, Sepulcre J, Sabuncu MR,Shafee R, Lu J, Liu H. 2013. Individual variability in func-tional connectivity architecture of the human brain. Neuron.77:586–595.

Oishi K, Zilles K, Amunts K, Faria A, Jiang H, Li X, Akhter K,Hua K, Woods R, Toga AW et al. 2008. Human brain whitematter atlas: identification and assignment of commonanatomical structures in superficial white matter. NeuroIm-age. 43:447–457.

Osher DE, Saxe RR, Koldewyn K, Gabrieli JDE, Kanwisher N,Saygin ZM. 2016. Structural connectivity fingerprints predictcortical selectivity for multiple visual categories across cor-tex. Cereb Cortex. 26:1668–1683.

Salehi M, Greene AS, Karbasi A, Shen X, Scheinost D, ConstableRT. 2018. There is no single functional atlas even for a singleindividual: parcellation of the human brain is state depen-dent. bioRxiv. 431833.

Saygin ZM, Osher DE, Koldewyn K, Reynolds G, Gabrieli JDE,Saxe RR. 2012. Anatomical connectivity patterns predict faceselectivity in the fusiform gyrus. Nat Neurosci. 15:321–327.

Saygin ZM, Osher DE, Norton ES, Youssoufian DA, Beach SD,Feather J, Gaab N, Gabrieli JDE, Kanwisher N. 2016. Connectiv-

ity precedes function in the development of the visual wordform area. Nat Neurosci. 19:1250–1255.

Schaefer A, Kong R, Gordon EM, Laumann TO, Zuo XN, HolmesAJ, Eickhoff SB, Yeo BTT. 2017. Local-global parcellation of thehuman cerebral cortex from intrinsic functional connectivityMRI. Cereb Cortex. 28:3095–3114.

Shen X, Tokoglu F, Papademetris X, Constable RT. 2013.Groupwise whole-brain parcellation from resting-state fMRIdata for network node identification. NeuroImage. 82:403–415.

Smith SM. 2002. Fast robust automated brain extraction. HumBrain Mapp. 17:143–155.

Sotiropoulos SN, Jbabdi S, Xu J, Andersson JL, Moeller S, AuerbachEJ, Glasser MF, Hernandez M, Sapiro G, Jenkinson M et al. 2013.Advances in diffusion MRI acquisition and processing in thehuman connectome project. NeuroImage. 80:125–143.

van den Heuvel MP, Scholtens LH, Feldman Barrett L, Hilge-tag CC, de Reus MA. 2015. Bridging cytoarchitectonicsand connectomics in human cerebral cortex. J Neurosci.35:13943–13948.

Veenith TV, Carter E, Grossac J, Newcombe VFJ, Outtrim JG, Lup-son V, Williams GB, Menon DK, Coles JP. 2013. Inter subjectvariability and reproducibility of diffusion tensor imagingwithin and between different imaging sessions. PLoS One.8:e65941.

Wang D, Buckner RL, Fox MD, Holt DJ, Holmes AJ, Stoeck-lein S, Langs G, Pan R, Qian T, Li K et al. 2015. Parcellat-ing cortical functional networks in individuals. Nat Neurosci.18:1853–1860.

Yendiki A, Panneck P, Srinivasan P, Stevens A, Zöllei L, Augusti-nack J, Wang R, Salat D, Ehrlich S, Behrens T et al. 2011. Auto-mated probabilistic reconstruction of white-matter path-ways in health and disease using an atlas of the underlyinganatomy. Front Neuroninform. 5:1–12.

Yeo BTT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D,Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JRet al. 2011. The organization of the human cerebral cortexestimated by intrinsic functional connectivity. J Neurophysiol.106:2322–2345.

Zilles K, Amunts K. 2013. Individual variability is not noise. TrendsCogn Sci. 17:153–155.

Dow

nloaded from https://academ

ic.oup.com/cercor/advance-article-abstract/doi/10.1093/cercor/bhz303/5669906 by Beijing N

ormal U

niversity Library user on 17 Decem

ber 2019