Research Articles, Systems/Circuits

A Novel Directed Seed-Based Connectivity Analysis Toolbox Applied to Human and Marmoset Resting-State FMRI

Takuto Okuno, Junichi Hata, Chino Kawai, Hideyuki Okano and Alexander Woodward
Journal of Neuroscience 6 November 2024, 44 (45) e0389242024; https://doi.org/10.1523/JNEUROSCI.0389-24.2024

Abstract

Estimating the direction of functional connectivity (FC) can help further elucidate complex brain function. However, the estimation of directed FC at the voxel level in fMRI data, and evaluating its performance, has yet to be done. We therefore developed a novel directed seed-based connectivity analysis (SCA) method based on normalized pairwise Granger causality that provides greater detail and accuracy over ROI-based methods. We evaluated its performance against 145 cortical retrograde tracer injections in male and female marmosets that were used as ground truth cellular connectivity on a voxel-by-voxel basis. The receiver operating characteristic (ROC) curve was calculated for each injection, and we achieved area under the ROC curve of 0.95 for undirected and 0.942 for directed SCA in the case of high cell count threshold. This indicates that SCA can reliably estimate the strong cellular connections between voxels in fMRI data. We then used our directed SCA method to analyze the human default mode network (DMN) and found that dlPFC (dorsolateral prefrontal cortex) and temporal lobe were separated from other DMN regions, forming part of the language-network that works together with the core DMN regions. We also found that the cerebellum (Crus I-II) was strongly targeted by the posterior parietal cortices and dlPFC, but reciprocal connections were not observed. Thus, the cerebellum may not be a part of, but instead a target of, the DMN and language-network. Summarily, our novel directed SCA method, visualized with a new functional flat mapping technique, opens a new paradigm for whole-brain functional analysis.

Significance Statement

We developed a novel directed seed-based connectivity analysis (SCA) method. To evaluate its performance, we used 145 retrograde tracer injections into the common marmoset left cortex as ground truth cellular connectivity on a voxel-by-voxel basis. We achieved an area under the curve (AUC) of 0.95 for undirected and 0.942 for directed SCA; thus SCA can estimate strong cellular connections with high accuracy. Then, we analyzed the human default mode network (DMN) with directed SCA and found that the dorsolateral prefrontal cortex and temporal lobe were not part of the DMN but part of the language-network. Moreover, our analysis showed that the cerebellum may not be part of, but instead a target of, the DMN and language-network.

Introduction

Understanding the connectivity between brain regions is important for elucidating the network structure of the brain (the connectome). Connectome analysis can provide the basis for understanding pathological conditions, and the uncovered structure can help formulate new hypothesis-driven studies. The brain can be analyzed using both invasive and noninvasive approaches. An example of an invasive approach is the common marmoset retrograde tracing study of Majka et al. (2020), used to investigate cellular connectivity. The common marmoset (Callithrix jacchus) is a primate with a brain that is near-lissencephalic, with many regions anatomically and functionally homologous to the human brain (Okano, 2021). Noninvasive approaches generally have a lower spatiotemporal resolution compared with invasive ones, but they are important because invasive approaches cannot be applied to humans. Therefore, in this research, we directly compared noninvasive and invasive approaches at the voxel level to investigate the relationship between structural (cellular) and functional connectivity (FC). To do this, we mapped the injection sites of Majka et al.'s study and the number of connected cells to the left cortex at the submillimeter-voxel level. This was used as ground truth connectivity that was compared with 48 sessions of awake marmoset resting-state (rs-)fMRI data, analyzed by our seed-based connectivity analysis (SCA) approach at the voxel level.

SCA is usually performed by calculating the correlation between a seed and all other voxels. A one-sample t test in each voxel is performed for second-level (group) analysis (Holmes and Friston, 1998). However, correlations do not detect directional influence (Okuno and Woodward, 2021), and in this work, we propose a novel directed SCA method. Here, directional influence (causality estimation) at the voxel was calculated for both input and output directions by normalized pairwise Granger causality (npwGC; Granger, 1969), and a one-sample Wilcoxon signed-rank test in each voxel was performed for second-level analysis. Retrograde tracing can be used as a ground truth to assess the accuracy of directed and undirected SCA by ROC (receiver operating characteristic) curve analysis. In this scheme, we investigated the optimal spatial smoothing size and nuisance factor removal method for fMRI preprocessing. We found a FWHM (full-width at half-maximum) of 3–4 voxels was optimal for spatial smoothing, and ICA-AROMA (Pruim et al., 2015) for nuisance factor removal greatly improved the AUC (area under the curve). We achieved an AUC of 0.95 for a strong cellular connectivity condition (cell count, >100, at a voxel) for undirected SCA and an AUC of 0.942 for directed SCA. These results show that undirected and directed FC are very reliable for detecting strong cellular connections, lending credence to future noninvasive studies.

Finally, we applied our novel directed SCA to investigate the directional influence among human default mode network (DMN) voxels. The full voxel FC matrix is huge, making it hard to interpret, so we developed a novel visualization technique called “functional flat mapping.” Here we applied dimension reduction using Uniform Manifold Approximation and Projection (UMAP; McInnes et al., 2018), and the FC matrix was reduced to a unique 2D planar map. We found that the cerebellum (Crus I-II) is strongly targeted from the posterior parietal cortices (PPC) and dorsolateral prefrontal cortex (dlPFC) regions, but reciprocal connections were not observed. Thus, the cerebellum appears to not be a part of but a target of the DMN. The temporal lobe and dlPFC regions were previously defined as the brain's language network (Mineroff et al., 2018), separated from the DMN regions. Graph analytics of directional SCA showed a clear separation between these two networks, even in rs-fMRI data. Thus, directional analysis can clarify the functional separation of dlPFC-temporal versus posterior cingulate cortex (PCC)–PPC–mPFC as DMN regions, and furthermore, zero-lag FC (including undirected SCA) indicates these two networks operate together. Our novel directed SCA can reliably detect voxel-based directional influence and complements zero-lag FC analysis. We believe that our approach described in this article will be useful for further neuroimaging studies.

Materials and Methods

Preprocessing of marmoset rs-fMRI data

Awake rs-fMRI data of the common marmoset (Callithrix jacchus) were acquired as part of the Brain/MINDS project (Okano et al., 2015; Muta et al., 2022). A Bruker BioSpec 9.4 T MRI machine (Biospin) was used. The experimental settings of the gradient recalled echo planar imaging (EPI) sequence were as follows: flip angle, 65; repetition time (TR), 2,000 ms; echo time (TE), 16 ms; pixel size, 0.7 × 0.7 mm; slice thickness, 0.7 mm; matrix size, 60 × 42 × 52; and frame length, 150.

For our experiments, T1WI, T2WI, and rs-fMRI NIfTI files of healthy awake marmosets aged 3–6 years (three males, one female; 12 sessions × 4 = 48 sessions of fMRI data) were used. Preprocessing and image registration were performed using Statistical Parametric Mapping (SPM12; Penny et al., 2007). SPM12 registered NIfTI images to the Marmoset MRI Standard Brain (Iriki, 2017). We removed the first 10 frames of the rs-fMRI data. For finding the optimal spatial smoothing size for group SCAs, a FWHM from 1 voxel (0.7 mm) to 5 voxels (3.5 mm) at 0.2 voxel steps was investigated. Temporal high-pass filters were investigated from 0.1 to 0.001 Hz. Combinations of several nuisance factor removal methods were investigated. The methods were global mean (gm; the average signal across all voxels), global signal (the average signal across all brain voxels), mean CSF (cerebrospinal fluid; the average signal across CSF voxels), mean white matter (the average signal across white matter voxels), 6HMP (head motion parameters), 24HMP, aCompCor, tCompCor (Behzadi et al., 2007), ICA-AROMA (Pruim et al., 2015), and polynomial detrending (Friman et al., 2004).

Undirected SCA

Undirected SCA was done by calculating the correlation coefficients between a seed voxel and all other voxels. MATLAB scripts for this analysis were developed in-house and worked together with the VARDNN toolbox (Okuno and Woodward, 2021). After calculating the correlation coefficients in each voxel from individual sessions, a mixed-effects model was applied to acquire final group results. A one-sample t test in each voxel was performed for second-level (group) analysis (Holmes and Friston, 1998). Bonferroni’s correction was then applied to correct for the familywise error (FWE) rate and a T value threshold was applied to acquire significantly correlated voxels.

We used the BrainSpace toolbox (Vos de Wael et al., 2020) to generate the principal gradient of the marmoset left cortex. The mixed-effects model T value matrix (9,862 × 9,862 voxels) of the marmoset left cortex was processed with the “kernel: normalized angle” and “approach: diffusion mapping” settings, and 10 components of the gradient map were obtained. Then, the first component was used as the principal gradient (Gradient 1).

Directed SCA

Directed SCA was calculated by npwGC between a seed voxel and all other voxels. MATLAB scripts for this analysis were developed in-house and worked together with the VARDNN toolbox. We applied pairwise Granger causality (pwGC), first proposed by Granger (1969), to describe the temporal relationship between two nodes as follows:xi(t)=a0+bpxj(t−p)+εi(t),(1) xi(t)=a0+apxi(t−p)+bpxj(t−p)+εij(t),(2) where xi={xi(t)}t=1…L and xj={xj(t)}t=1…L are time-series data of node i and j; L is the number of frames of data; p is the spot time lag [because of the peak time of BOLD (blood oxygenation level dependent) signal issue (Extended Data Fig. 1-3), we did not use other lag terms]; a0 , ap , and bp are coefficients; and εi={εi(t)}t=1…L and εij={εij(t)}t=1…L are time-series of residuals. The coefficients a0 , ap , and bp and residuals εi and εij of the autoregression formulae are usually calculated by ordinary least squares. Then, the pwGC from j (source) to i (target) is calculated by the following:pwGCj→i=−log(var(εi)/var(εij)).(3) However, we found that pwGC resulted in very large noise in the vertical direction of the mixed-effect negative log p value matrix (Extended Data Fig. 1-1a). The accuracy was very low in this case, so we developed a new normalization method as follows:npwGCj→i=−log(dij/d¯i),(4) d¯i=1n∑j=1ndij,(5) dij=var(εi)/var(εij).(6) This technique reduced the vertical noise in the mixed-effects matrix (Extended Data Fig. 1-1b) and greatly improved the accuracy of causality estimation. After calculating npwGC in each voxel from individual sessions, a mixed-effects model was applied to acquire final group results. A one-sample Wilcoxon signed-rank test in each voxel was performed for second-level (group) analysis. We investigated the one-sample t test and Wilcoxon signed-rank test methods for group analysis (Extended Data Fig. 1-2). The Wilcoxon signed-rank test showed higher AUCs for directed SCA. Bonferroni’s correction was then applied to correct for the FWE rate, and a Z-score threshold was applied to acquire significantly influenced voxels.

Transforming retrograde tracing data into the whole-brain voxel space

Majka et al. (2020) published a comprehensive retrograde tracer dataset (145 injections) for the left hemisphere cortex of the common marmoset brain. The data can be downloaded from the Marmoset Brain Connectivity Atlas (https://www.marmosetbrain.org/), and they were integrated into the STPT (serial two-photon tomography) template image space (Skibbe et al., 2022). We transformed the retrograde tracer results from the STPT template to the Marmoset MRI Standard Brain (Iriki, 2017) by manual affine transformation, and this was used for our comparison between functional and cellular connectivity.

Preprocessing of human rs-fMRI data for investigating spot time lag

Human rs-fMRI data from Tokyo Metropolitan University was used for investigating the spot time lag parameter of directed SCA. A 3.0 Tesla scanner (GE HealthCare, SIGNA Premier), 48-channel coil for the head (GE HealthCare) was used. The data have multiple TRs (750, 1,200, 2,000 ms) for the same 10 subjects aged 23.8 ± 4.87 years (three males, seven females). Experimental settings of the gradient-echo echo–planar imaging (EPI) sequence were as follows: flip angle, 52; TR, 750, 1,200, and 2,000 ms; TE, 36 ms; pixel size, 3 × 3 mm; slice thickness, 2.6 mm; spacing, 0.6 mm; matrix size, 80 × 80; multiband factor, 8; and frame length, 400, 250, and 150. The CONN toolbox (Whitfield-Gabrieli and Nieto-Castanon, 2012) was used for preprocessing. CONN performed the realignment and coregistration of NIfTI images to the standard Montreal Neurological Institute brain space. The first 10 frames of the rs-fMRI data were removed, and the remaining data were smoothed using a FWHM of 6.8 mm (3.4 voxels). ICA-AROMA was applied for nuisance factor removal.

Preprocessing of HCP rs-fMRI data for DMN analysis

Rs-fMRI data from the WU-Minn HCP consortium (the S500 release; Van Essen et al., 2013) was used for human DMN analysis. Scanning used a customized SC72 gradient insert and a body transmitter coil with 56 cm bore size, and data were saved in NIfTI format. Experimental settings of the gradient-echo EPI sequence were as follows: flip angle, 52; TR, 720 ms; TE, 33.1 ms; pixel size, 2 × 2 mm; slice thickness, 2 mm; matrix size, 104 × 104 × 90; multiband factor, 8; and frame length, 1,200. More information on the rs-parameters can be found at the HCP website (https://www.humanconnectome.org/storage/app/media/documentation/s500/HCP_S500_Release_Reference_Manual.pdf).

T1WI, T2WI, and rs-fMRI NIfTI files from the S500 release were downloaded, and a total of 200 sessions (50 male subjects × 2 sessions, 50 female subjects × 2 sessions) were used in our experiments. Preprocessing was the same as for the data used to investigate the spot time lag. Spatial smoothing using a FWHM of 6.8 mm (3.4 voxels) and ICA-AROMA were applied for directed and undirected SCA of human DMN.

Independent component analysis of HCP rs-fMRI data

After preprocessing, group ICA was applied to acquire 15 components from the human rs-fMRI data. The systematic checking of the optimal component number was done in a previous study (Okuno et al., 2024). Here, MELODIC (Beckmann and Smith, 2004) was used to obtain group ICA from 200 sessions. Multisession temporal concatenation was performed, and a spatial map was obtained. Finally, the DMN component used in our study was manually selected from the 15 rs-fMRI data components. For surface mappings of human data, the command “wb_command -volume-to-surface-mapping” of the Connectome Workbench visualization software (Marcus et al., 2011) was used to map NIfTI image data onto the human cortical surface. Finally, the cortical surface (in gray), the mapped functional data, and the Brodmann label mapping (included in the HCP data) were overlaid to produce our visualizations.

Graph analytics of the human DMN

Directed and undirected SCA of human rs-fMRI data generated a huge FC matrix (24,860 × 24,860 voxels, 1 voxel size, 4 × 4 × 4 mm3 downsampled from the CONN atlas and used as mask). The threshold to obtain the adjacency matrix is taken from p < 0.05 and Bonferroni’s correction or top 5% of values in the T value/Z-score matrix (chosen based on our analysis of 145 retrograde tracer injections in the common marmoset). For directed SCA, the threshold was Z > 4.61, and for undirected SCA, the threshold was T > 18.82. Based on these adjacency matrices, graphs were visualized by the Scalable Force-Directed Placement (SFDP) algorithm (Hu, 2005). Hierarchical circle graphs of DMN voxels based on directed and undirected SCA results were visualized by the Nested Stochastic Block Model (Peixoto, 2014).

Preprocessing of mouse rs-fMRI data

Awake rs-fMRI data of the mouse were acquired by a Bruker BioSpec 9.4 T MRI machine (Biospin). The experimental settings of the gradient recalled EPI sequence were as follows: flip angle, 90; TR, 1,500 ms; TE, 12 ms; pixel size, 0.24 × 0.24 mm; slice thickness, 0.6 mm; matrix size, 75 × 60 × 25; and frame length, 400.

For our experiments, rs-fMRI NIfTI files of healthy awake mice aged 9 weeks (six males; 22 sessions of fMRI data) were used. Preprocessing and image registration were performed using SPM12 (Penny et al., 2007) and ANTs (Avants et al., 2011). SPM12 performed slice timing correction and motion correction, and ANTs registered NIfTI images to the mouse average template (47 × 42 × 25), downsampled from Allen Mouse Common Coordinate Framework (CCF; Wang et al., 2020). We removed the first 10 frames of the rs-fMRI data, and spatial smoothing of FWHM of 4 voxels (0.96 mm) was applied. Combinations of several nuisance factor removal methods were investigated. The methods were gm, global signal, mean CSF, mean white matter, 6HMP, 24HMP, aCompCor, tCompCor (Behzadi et al., 2007), ICA-AROMA (Pruim et al., 2015), and polynomial detrending (Friman et al., 2004).

Statistical analysis

A mixed-effects model was used for group analysis. For undirected SCA, a one-sample t test in each voxel was done for second-level (group) analysis. For directed SCA, a one-sample Wilcoxon signed-rank test in each voxel was done for second-level (group) analysis. Statistical significance was set at p < 0.05. Bonferroni’s correction was then applied to correct for the FWE rate and a T value or a Z-score threshold was applied to acquire significantly correlated voxels.

Software accessibility

The code used in this study (directed and undirected SCA) is provided as open source and available from https://github.com/takuto-okuno-riken/dirsca.

Results

Comparison between tracer injections and SCAs in the marmoset left cortex

The matrix of 145 retrograde tracer injections and the structural (cellular) connectivity in left cortical voxels (0.7 mm × 0.7 mm × 0.7 mm) of the common marmoset is shown in Figure 1a, top. In a previous study, 97% of all possible corticocortical connections were observed in mouse anatomical results (Gămănuţ et al., 2018). However, we found that only ∼5% (mean value in each injection) of voxels were connected in the common marmoset data. Thus, ROI (region of interest)-level and voxel-level parcellations can give very different results. Directed and undirected SCA from injection voxels to left cortical voxels are shown in Figure 1a, middle and bottom. The cellular connectivity matrix, with specific cell count connection threshold (Fig. 1a), can be used as a ground truth, and the ROC curves of FC–SC (structural connectivity) detection were calculated for directed and undirected SCA matrices. From this we achieved an AUC of 0.95 (Fig. 1b) in the strong cellular connectivity case (a cell count, >100, at a voxel) for undirected SCA and an AUC of 0.942 (Fig. 1c) for directed SCA with an optimal spatial smoothing size (FWHM, 3.4 voxel) and nuisance factor removal (by ICA-AROMA). These results suggest that undirected and directed SCA is very reliable for detecting strong cellular connectivity. However, the correlation coefficient between the structure (cell count, >0, at a voxel) and function (absolute T value of undirected SCA) was not so high (an FC–SC correlation of r = 0.282). This shows that although a strong FC indicates the presence of cellular connectivity, FC is not proportional to SC.

Figure 1.
Figure 1.

Comparison between tracer injections and SCA in the marmoset left cortex. a, Schematic image of the comparison between cellular connectivity and FC. Top, Cellular connectivity matrix for the marmoset left cortex voxels; the vertical axis is the 145 injection voxels; the horizontal axis is the voxels of the left cortex. Middle, Undirected SCA matrix (N = 48; seed voxels are the same as injection voxels) in the marmoset left cortex voxels. Bottom, Directed SCA matrix in the marmoset left cortex voxels. b, FC–SC detection result. ROC curve analysis of undirected SCA matrix with ground truth matrix (threshold set to be >100 cellular connections per voxel). Gray line shows ROC curves at each injection. Sky blue line shows total result (AUC, 0.95). c, FC–SC detection result. ROC curve analysis of directed SCA matrix with ground truth matrix (threshold set to be >100 cellular connections per voxel). Sky blue line shows total result (AUC, 0.942). d, FC–SC correlation result between undirected SCA matrix and SC matrix (threshold set to be >0 cellular connections per voxel, 76,301 voxels). e, Principal gradient of the marmoset left cortex from a mixed-effects model T value matrix of 9,862 × 9,862 voxels. f, The DMN-like ICA component from Okuno et al. (2024). g, 3D projection of injection points into the marmoset left cortex, left lateral view. The circle size represents the correlation coefficient |R| between the Gradient 1 value and FC–SC correlation. The circle color represents the Gradient 1 value. h, Scatterplot of g; X axis is FC–SC correlation; Y axis is Gradient 1. See Extended Data Figure 1-1 for example analysis results of pwGC versus npwGC, Extended Data Figure 1-2 for investigating the group analysis test methods, Extended Data Figure 1-3 for investigating optimal spot time lag for directed SCA, Extended Data Figure 1-4 for interindividual variability of SC, Extended Data Figure 1-5 for relationship between ROI-based and voxel-based time-series, and Extended Data Figure 1-6 for relationship between Euclidean distance and SCA.

Fig. 1-1

Example of whole brain voxel wise analysis by pairwise Granger Causality. a, Matrix of the mixed-effects (one sample t-test) pwGC(1) result. Noise remains in the vertical direction. b, Matrix of the mixed-effects (one sample t-test) npwGC(1) result. Noise in the vertical direction was removed by normalization. Download Fig. 1-1, TIF file.

Fig. 1-2

Investigating the group analysis test method for directed and undirected SCA. a, ROC curve analysis between tracer injections and undirected SCA (N = 48, smoothing with FWHM = 3.4 voxel and high-pass filtering 1/128  Hz) with a one sample t-test or Wilcoxon signed rank test. The vertical axis is the AUC (higher is better). The horizontal axis is the cell count threshold for the ground truth. The blue line (t-test) shows the mean AUC between 145 injections and undirected SCA, the error bar shows the standard error of 145 AUCs. The t-test showed a higher AUC for undirected SCA. b, ROC curve analysis between tracer injections and directed SCA with one sample t-test or Wilcoxon signed rank test. The Wilcoxon signed rank test showed a higher AUC for directed SCA. Download Fig. 1-2, TIF file.

Fig. 1-3

Investigating optimal spot time-lag for directed SCA. a, ROC curve analysis between tracer injections and directed SCA in the marmoset (N = 48, smoothing with FWHM = 3.4 voxel and ICA-AROMA) for each spot time-lag. The vertical axis is the AUC (higher is better). The horizontal axis is the cell count threshold for the ground truth. The blue line (spot time-lag = 1) shows the mean AUC between 145 injections and directed SCA, the error bar shows the standard error of 145 AUCs. When the spot time-lag was 2 or more, the AUC clearly decreases. b, Canonical hemodynamic response function (HRF) for the marmoset and human. c, Histogram of the Z-score matrix (24860 × 24860 voxels, 1 voxel size: 4 × 4 × 4  mm) of directed SCA (N = 29 [TR = 0.75  s], N = 30 [TR = 1.2  s], N = 29 [TR = 2  s]) for the human brain. 3 to 4 seconds gave an optimal spot time-lag, thus lag = 3 for TR = 1.2  s, and lag = 2 for TR = 2  s. d, Directed SCA matrix of the default mode network for each spot time-lag. Lag = 3 of TR = 2  s clearly loses sensitivity. Download Fig. 1-3, TIF file.

Fig. 1-4

Inter-individual variability of structural connectivity in the common marmoset. a, Cosine similarities of SC (DWI Tractography) matrices among 126 subjects (126 × 126 matrix). SC matrix data was downloaded from the Brain/MINDS data-portal: https://dataportal.brainminds.jp/marmoset-mri-na216). Top upper right of matrix shows inter-individual variability of SC matrices (104 × 104 ROIs) in the common marmoset. The color bar shows the cosine similarity value. b, Box plot of A (left) and thresholded SC (>10 and >100) versions. Higher SC thresholds improve the similarity. c, The two most different SC matrices (subject IDs 9 and 125). The cosine similarity was 0.886. Download Fig. 1-4, TIF file.

Fig. 1-5

Relationship between ROI-based and voxel-based time-series. a, Time-series of Middle Temporal Gyrus (MTG), left ROI (206 voxels), from one subject of the Human Connectome Project. X axis is frame number (TR = 0.72). The light-colored lines are the time-series of each voxel. The black line is the average time-series of the MTG left ROI voxels. Compared to the time-series of each voxel, the ROI time-series had a narrower amplitude range. b, Correlation result between the ROI time-series and each voxel time-series. The correlation with the ROI time-series varied widely, from 0.2 to 0.8, and it may be difficult to say that the ROI time-series was representative for each voxel. c, Correlation result for MTG left (206 voxels) vs. Middle Frontal Gyrus (MFG) left (426 voxels). d, Box plot of C. Compared with the correlation between the MTG left and MFG left ROI time-series (R = 0.435), the correlation coefficients of each voxel varied widely from -0.2 to 0.5. The time series correlation of the ROIs may be too high to be representative of the two ROIs. e, Pairwise Granger Causality result from MTG left to MFG left voxels. Many significant connections (P < .05, Bonferroni corrected) were observed (1965 connections), but not in the ROI-based calculation (P = 0.36). Download Fig. 1-5, TIF file.

Fig. 1-6

Relationship between Euclidean distance and undirected and directed SCA. a, Scatter plot between log10 of undirected SCA |T-value| and Euclidean distance. Correlation R = -0.638. b, ROC curve analysis between tracer injections and undirected SCA (N = 48) for each distance category (≤5  mm, 5 to 10  mm, > 10  mm). The vertical axis is the AUC (higher is better). The horizontal axis is the cell count threshold for the ground truth. The blue line (≤5  mm) is the mean AUC between 145 injections and undirected SCA. The error bar shows the standard error of 145 AUCs at each cell count threshold and distance category. c, Scatter plot between directed SCA Z-value and Euclidean distance. Correlation R = -0.679. d, ROC curve analysis between tracer injections and directed SCA (N = 48) for each distance category (≤5  mm, 5 to 10  mm, > 10  mm). Download Fig. 1-6, TIF file.

Fig. 1-7

FC-SC detection using SC (DWI) and undirected FC and directed FC in the human and marmoset brain. a. Mean SC matrix, generated by 90 ROIs (AALv4 atlas) from 88 human subjects. b. Group analysis (mixed effects model) of undirected FC (Pearson correlation) matrix, generated from 200 sessions of HCP rs-fMRI data. c. Group analysis (mixed effects model) of directed FC (npwGC) matrix, generated from 200 sessions of HCP rs-fMRI data. d. FC-SC detection in the human. The X axis is log10 of the fiber count threshold used for ground truth. As the threshold increased, the ground truth network density decreased (green line, right Y axis). As the threshold increased, the AUC increased (red & blue lines, left Y axis). e. FC-SC detection in the marmoset. The mean SC matrix was generated by 104 ROIs from 126 marmosets. Group analyses of undirected and directed FC were generated from 48 sessions of awake marmoset rs-fMRI data. X and Y axis are the same as d. Download Fig. 1-7, TIF file.

The Euclidean distance versus undirected and directed SCA was also investigated (Extended Data Fig. 1-6). Shorter-distance voxels showed higher FC, as seen in the previous study by Hori et al. (2020). Thus, there was a strong anticorrelation between distance and undirected SCA (R = −0.638) and directed SCA (R = −0.679; Extended Data Fig. 1-6a,c). FC–SC detection was also checked at different distance categories (≤5 mm, 5–10 mm, and >10 mm), and it was observed that the longer the distance, the lower the FC and the lower the detection accuracy (Extended Data Fig. 1-6b,d).

Additionally, we generated the principal gradient of the marmoset left cortex from a mixed-effects model [T value matrix of 9,862 × 9,862 voxels (Fig. 1e). This result is similar to the DMN-like ICA component of the marmoset cortex (Fig. 1f from Okuno et al. (2024)]. This suggests that the DMN region in the marmoset is transmodal, and this result is consistent with human results (Margulies et al., 2016). However, we found some injections in unimodal areas (V1, V2, etc.) showed a low FC–SC correlation, and some injections in transmodal areas (PE, etc.) showed a high FC–SC correlation (Fig. 1g). The correlation between the principal gradient and FC–SC correlation values for 145 injections did not show any significant tendency (R = 0.1; p = 0.23; Fig. 1h). Thus, voxel-based analysis was not consistent with previous studies (Preti and Van De Ville, 2019; Vázquez-Rodríguez et al., 2019). This observation will be explored in the discussion section.

Investigating the optimal spatial smoothing size for group SCAs

Spatial smoothing is typically applied at the preprocessing stage of fMRI dataset analysis (Mikl et al., 2008). Especially, it is useful for suppressing the influence of functional variability within and across individual subjects. We investigated the optimal spatial smoothing size for group SCAs with FWHM by a Gaussian smoothing kernel from 1 voxel (0.7 mm) to 5 voxels (3.5 mm) at 0.2 voxel steps. Figure 2a shows an example of spatial smoothing of rs-fMRI data of the common marmoset. We found that a FWHM of ∼4 voxels gave the highest AUC for group analysis of undirected SCA (s40 in Fig. 2b). We also investigated the FC–SC correlation in each tracer injection; a FWHM from 3 to 4 voxels was optimal for group analysis (Fig. 2c). For directed SCA, a FWHM of ∼4 voxels showed the highest AUC (Fig. 2d) and was optimal for FC–SC correlation analysis (Fig. 2e). Therefore, we applied an FWHM of 3.4 voxels as the optimal spatial smoothing size for subsequent analyses.

Figure 2.
Figure 2.

Investigating the optimal spatial smoothing size for group analysis of directed and undirected SCAs. a, Example of spatial smoothing result of marmoset whole-head fMRI image. The left shows a raw transverse slice image. The right shows smoothed slice images (FWHMs are from 1 voxel to 4 voxels). b, ROC curve analysis between tracer injections and undirected SCA (N = 48) for each spatial smoothing setting. The vertical axis is the AUC (higher is better). The horizontal axis is the cell count threshold for the ground truth. The blue line (raw) shows mean AUC between 145 injections and undirected SCA; the error bar shows the standard error of 145 AUCs at each cell count threshold. c, Box plot of correlation between 145 tracer injections and undirected SCA for each spatial smoothing setting. The vertical axis is the correlation coefficient. The horizontal axis is spatial smoothing size (FWHM by 0.2 voxel step). d, ROC curve analysis between tracer injections and directed SCA for each spatial smoothing setting. e, Box plot of the correlation between 145 tracer injections and directed SCA for each spatial smoothing setting.

Investigating the optimal nuisance factor removal method for group SCAs

Nuisance factor removal is important for removing motion artifacts and physiological noise such as pulsation and respiration. Several studies (Ciric et al., 2017; Parkes et al., 2018; Chuang et al., 2019) provide a comprehensive comparison among nuisance factor removal methods. We also investigated the optimal nuisance factor removal method for group analysis of undirected SCA (Fig. 3a,b) and directed SCA (Fig. 3c,d). We found that most methods improved the AUCs of undirected SCA; the ICA-AROMA (Pruim et al., 2015) and gm combination showed the highest AUC (Fig. 3b, orange dashed line), and ICA-AROMA alone showed the next highest AUC. For directed SCA, ICA-AROMA showed the highest AUC (Fig. 3d, orange dashed line), and the ICA-AROMA and gm combination showed the next highest AUC. On face value, the use of ICA-AROMA alone for both directed and undirected SCA seems ideal, because it showed the highest AUC for directed SCA. The gm and aCompCor (Behzadi et al., 2007) combination appears the next best option because it showed the next highest AUC for directed SCA. Chuang et al. (2019) presented a comprehensive investigation of nuisance factor removal methods for the anesthetized rat brain and concluded that some methods for the human case did not work for the rat brain. For example, motion regression in motionless anesthetized rat may increase artifacts and alone does not contribute to signal detection. Then, we conducted additional experiments in this regard using the awake mouse with skull fixed and anterograde axonal tract tracing (Harris et al., 2012; Oh et al., 2014) data as the ground truth. Anterograde axonal tract tracing data were downloaded from Allen Mouse CCF (Wang et al., 2020). Only the left cortex of the mouse was used, and the injection point located in the left cortex was extracted and used as the ground truth (N = 578). Extended Data Figure 3-1 shows our investigation into the optimal nuisance factor removal method between ground truth tracing data and undirected and directed SCA. FWHM of 4 voxel was used for smoothing. Since anterograde axonal tract tracing highlights axons, it is unclear whether functional connections exist between voxels on the highlighted pathway. Therefore, increasing the intensity threshold increased the accuracy of detecting structural connectivity by FC (Extended Data Fig. 3-1a). At a threshold above 95% of intensity, only very strong tracing signal became the ground truth, and such voxels could be detected with high accuracy in both undirected and directed SCA (AUC, >0.91; Extended Data Fig. 3-1a,b). This result is consistent with previous ROI-based analyses (Stafford et al., 2014). Compared with the anesthetized rat (Chuang et al., 2019), the skull was fixed in the awake mouse and awake marmoset experiments, but this does not mean that the position of the brain tissue was completely fixed; at 6hmp and 24hmp, the AUC was improved. The brain tissue itself is floating in cerebral spinal fluid, suggesting that the regression of positional variability compensates for the shaking of the brain tissue. Additionally, different from the rat brain, the white matter and CSF voxels can be easily separated in the marmoset. Thus, we were able to apply ICA-AROMA and aCompCor and were able to achieve a higher AUC. This supports the use of nuisance factor removal methods used in humans for the analysis of the marmoset brain.

Figure 3.
Figure 3.

Investigating optimal nuisance factor removal method for directed and undirected SCAs. a, ROC curve analysis between tracer injections and undirected SCA (N = 48) for each nuisance factor removal method. The vertical axis is AUC (higher is better). The horizontal axis is cell count threshold for the ground truth. The blue line (raw, smoothing with FWHM = 3.4 voxel) shows the mean AUC between 145 injections and undirected SCA; the error bar shows the standard error of 145 AUCs for each nuisance factor removal method. b, Box plot of a. the ICA-AROMA and gm combination showed the highest AUC (orange dashed line). Sky blue-dashed line shows the gm and aCompCor combination. c, ROC curve analysis between tracer injections and directed SCA for each nuisance factor removal method. d, Box plot of c. ICA-AROMA shows highest AUC (orange dashed line). Sky blue-dashed line shows the gm and aCompCor combination (second highest).

Fig. 3-1

Investigating optimal nuisance factor removal method for directed and undirected SCAs in the mouse left cortex. a. ROC curve analysis between tracer injections and undirected SCA (N = 22) for each nuisance factor removal method. The vertical axis is the AUC (higher is better). The horizontal axis is the percentile threshold for the ground truth. The blue line (raw, smoothing with FWHM = 4 voxel) shows the mean AUC between 578 injections and undirected SCA, the error bar shows the standard error of 578 AUCs for each nuisance factor removal method. b, Box plot of a. The 24hmp, gm & aCompCor combination showed the highest AUC (orange dashed line) for directed and undirected SCA. c, ROC curve analysis between tracer injections and directed SCA for each nuisance factor removal method. d, Box plot of c. Download Fig. 3-1, TIF file.

Investigating the effect of high-pass filtering for group SCAs

A high-pass filter is typically applied to remove motion artifacts and physiological noise from BOLD signals. Thus, this process is essentially redundant if nuisance factor removal is used. Figure 4a shows FC–SC detection by undirected SCA (smoothing with an FWHM of 3.4 voxels and ICA-AROMA) for each high-pass filter setting. The high-pass filter did not show a clear improvement in the AUCs. On the contrary, the high-pass filter destroyed the time dependence of the BOLD signal, resulting in a decrease in the accuracy of directed SCA (Fig. 4b). Therefore, we do not recommend a high-pass or bandpass filter with a nuisance factor removal method in the preprocessing stage for both directed and undirected SCAs.

Figure 4.
Figure 4.

Investigating high-pass filtering for directed and undirected SCAs. a, ROC curve analysis between tracer injections and undirected SCA (N = 48; smoothing with FWHM, 3.4 voxel, and ICA-AROMA) for each high-pass filter setting. There was no large difference between no filter and high-pass filters. b, ROC curve analysis between tracer injections and directed SCA (smoothing with FWHM, 3.4 voxel, and ICA-AROMA) for each high-pass filter. High-pass filtering reduced AUCs.

Investigating optimal spot time lag for directed SCA

Directed SCA requires an appropriate spot time lag to estimate directional influence. Extended Data Figure 1-3a shows FC–SC detection by directed SCA in the marmoset (smoothing with an FWHM of 3.4 voxels and ICA-AROMA) with spot time lags from 1 to 5. Clearly, spot time lags from 2 to 5 lost sensitivity in causality estimation. Since the peak time of the hemodynamic response function (HRF) of the common marmoset is ∼3.1 s (Extended Data Fig. 1-3b; Yen et al., 2018), a spot time lag of 2 or more, which is ∼4 s for TR = 2 s, would exceed the peak HRF time. From these observations it appears that a long time lag reduces the accuracy of causality estimation. For the human case, the peak HRF time is ∼5–6 s (Bonakdarpour et al., 2007), so the lag time should be smaller than this. To further explore this topic, we investigated several spot time lags (1–3) and TRs (0.75, 1.2, 2 s) obtained for the same 10 subjects. Extended Data Figure 1-3c shows a histogram of the Z-score matrix (24,860 × 24,860 voxels; 1 voxel size, 4 × 4 × 4 mm3) of directed SCA [N = 29 (TR = 0.75 s); N = 30 (TR = 1.2 s); N = 29 (TR = 2 s)]. At a shorter time lag (0.75 s), the histogram was shifted in the negative direction, and some voxels were found to be saturated. At a longer time lag (6.0 s), the histogram was close to a normal distribution, but it lost sensitivity in causality estimation (Extended Data Fig. 1-3d, right). Therefore, we recommend using 3–4 s as an optimal spot time lag (lag of 3 for TR of 1.2 s; lag of 2 for TR of 2 s) for the human case.

We also performed FC–SC detection comparing Pearson's correlation (undirected FC) and npwGC (lag of 5 for TR of 0.72 s; directed FC) with SC on diffusion-weighted imaging (DWI). Extended Data Figure 1-7 shows this experimental result. A ground truth SC matrix (Extended Data Fig. 1-7a) was generated by 90 ROIs of the AALv4 atlas from 88 healthy human subjects (Škoch et al., 2022). We calculated the T value matrix (Extended Data Fig. 1-7b) of Pearson's correlation and the Z value matrix (Extended Data Fig. 1-7b) of npwGC from 200 sessions (50 male subjects × 2 sessions and 50 female subjects × 2 sessions) of HCP rs-fMRI data. The npwGC showed similar accuracy to Pearson's correlation for the human brain. We also investigated whether this analysis framework showed a similar trend in marmosets and confirmed that it does (Extended Data Fig. 1-7e). Although the AUC of Pearson's correlation tended to be higher in the marmoset result by DWI SC-based analysis, tracer-based analysis showed a closer result between correlation and npwGC (Fig. 1b,c). This may be because fiber counting from ROI-to-ROI end points generated a more symmetrical SC matrix and the ground truth became advantageous to undirected FC. Summarily, we confirmed the accuracy of npwGC by FC–SC detection using tracer data and DWI data in the marmoset and human.

Functional flat mapping of human brain voxels

Due to the large number of relationships, it is very difficult to visually interpret anything from the full voxel connectivity matrix (Fig. 5a; one voxel was downsampled to 4 × 4 × 4 mm3) in the human brain. Therefore, we developed a novel visualization method called “functional flat mapping” to simplify and display these many relationships. A dimension reduction technique called UMAP (McInnes et al., 2018) was applied to reduce and map the huge matrix (24,860 × 24,860) of relationships to a 2D scatterplot (Fig. 5a, left). One point corresponds to one voxel and close points (voxels) are functionally close. Since close voxels in the human brain also have close features, topological features are well preserved. Figure 5b shows the abbreviated brain region names on top of the functional flat mapping visualization, and from left to right, the cortical, subcortical, and cerebellar region points are shown in color. Interestingly, the topographic nature of the structural and functional relationships is well plotted. For example, the anterior and PCC and precuneus compose the central line through the cortical points, the hippocampal points spread out from the thalamic ones, and the amygdala points are attached to the tip of the hippocampal points. Functional flat mapping has many advantages, including the ability to display subcortical and cerebellar voxel information that is not displayed by surface mapping alone, and the ability to display information from many voxels that are missed when using maximum projection in either the sagittal, coronal, or transverse planes. We used this visualization technique to display our subsequent analysis results in this article.

Figure 5.
Figure 5.

Schematic image of functional flat mapping in human brain voxels. a, Schematic image of functional flat mapping. The FC matrix (N = 200; 24,860 × 24,860 voxels; 1 voxel size, 4 × 4 × 4 mm) for the human brain was reduced to a two-dimensional flat map by UMAP. The 24,860 points on the functional flat map correspond to the voxels in the 3D brain data. Close points are functionally close because of the UMAP algorithm. b, Cortical (left), subcortical (center), and cerebellum (right) points on the functional flat mapping are displayed separately. Abbreviations are based on the Conn Atlas (Whitfield-Gabrieli and Nieto-Castanon, 2012).

Analysis of the human DMN using directed and undirected SCA

Figure 6 shows visualizations of the human DMN going from surface mapping to functional flat mapping. Group ICA (MELODIC; Beckmann and Smith, 2004) was used to acquire 15 components from 200 sessions of HCP rs-fMRI data, and the DMN component used in our study was manually selected from them (Okuno et al., 2024). Figure 6a shows a surface mapping of the DMN component. The characteristic positive value regions of the DMN, such as the mPFC, dlPFC, PPC, PCC, and temporal lobe are shown. We manually annotated these regions by ITK-SNAP (Yushkevich et al., 2006; Fig. 6b), and the DMN voxels were plotted on the functional flat mapping (Fig. 6c). Also, the group ICA result can be directly plotted onto the flat map (Fig. 6d), and we could visually confirm the similarity between the original group ICA voxels and the manually annotated voxels. Finally, analysis of the human DMN using directed and undirected SCA is shown in Figure 7. The top row shows the internal DMN voxel connectivity by undirected SCA (Fig. 7a) and directed SCA (outputs-from and inputs-to each voxel; Fig. 7b,c). For visibility, only the most significant connections for the DMN voxels were displayed. To see more detailed connectivity, Extended Data Figure 7-2 shows the max projection of directed and undirected SCA results for voxels from each DMN region. Each region of the DMN is connected, and synchronization between regions, observed by zero-lag FC (undirected SCA) is thought to occur due to the mutual input and output (directed SCA) between regions. The bottom row shows the significant connections from DMN voxels to voxels outside the DMN (Fig. 7d,e) or from voxels outside the DMN to DMN voxels (Fig. 7f). Only the top five most significant connections and a distance limit >3 were displayed for visibility. If there was no distance limit in place then basically the closer points outside the DMN showed the strongest connections (Extended Data Fig. 7-1a). This is probably because the time-series of voxels around the DMN became similar due to spatial smoothing. For this reason, we set a distance limit to examine the connectivity with voxels that were slightly distant (voxels that were not the close neighbors). In each figure of the bottom row, the temporal lobe (red dashed ellipses) showed a different tendency to the other DMN regions. The temporal lobe showed widespread output to regions outside the DMN, and this widespread input/output was also observed in the max projection of directed SCA (Extended Data Fig. 7-2). The PPC, dlPFC, and temporal lobe regions showed strong output to the Crus I-II regions (Fig. 7e, gray dashed ellipses), but there was no strong input from the cerebellar voxels to any DMN voxels (Fig. 7f, gray dashed ellipses). These results indicate that the cerebellum (Crus I-II) may not be a part of but a target of the DMN.

Figure 6.
Figure 6.

Functional flat mapping of the DMN of the human brain. a, The right cortical surface of the human brain. Human rs-DMN based on group ICA result (N = 200) was mapped onto the surface. Z-score range is 2–10 for positive and −2 to −10 for negative. b, Manually annotated eight DMN ROIs by the ITK-SNAP software. A superior view of the brain; the left is posterior; the right is anterior. c, Functional flat mapping of eight DMN ROIs in each color and all other voxels. d, Functional flat mapping of the DMN component voxels based on the group ICA result.

Figure 7.
Figure 7.

Directed and undirected SCA of the DMN of the human brain. a, Functional flat mapping of eight DMN ROIs in each color and all other voxels. Lines indicate correlations inside DMN voxels. For better visibility, only the top-most significant connections (T > 18.8) of undirected SCA (N = 200) were visualized. The line color is based on ROI voxel color. Colors with lower transparency indicate higher T values. b, Functional flat mapping of eight DMN ROIs. Lines indicate output from DMN voxels to other DMN voxels. Top one significant connections (Z-score > 6) of directed SCA were visualized. c, Functional flat mapping of eight DMN ROIs. Lines indicate input to DMN voxels from other DMN voxels. Top one significant connections (Z-score > 6) of directed SCA were visualized. d, Functional flat mapping of eight DMN ROIs. Lines indicate correlations from DMN voxels to other (non-DMN) voxels. Top five most significant connections (T > 18.8) and distance > 3 (on UMAP) of undirected SCA were visualized. e, Functional flat mapping of eight DMN ROIs. Lines indicate output from DMN voxels to other (non-DMN) voxels. Top five most significant connections (Z-score > 6) and distance > 3 of directed SCA are visualized. f, Functional flat mapping of eight DMN ROIs. Lines indicate input to DMN voxels from other (non-DMN) voxels. Top five most significant connections (Z-score > 6) and distance > 3 of directed SCA are visualized. See Extended Data Figure 7-1 for directed and undirected SCA of the human DMN and Extended Data Figure 7-2 for max projection of directed and undirected SCAs in the human DMN.

Fig. 7-1

Directed and undirected SCA of the human default mode network. a, Functional flat mapping of eight DMN ROIs. Lines indicate correlations from DMN voxels to other (non-DMN) voxels. Top five most significant connections (T-value > 18.8) and distance > 0 (on UMAP) of undirected SCA are visualized. b, Functional flat mapping of eight DMN ROIs. Lines indicate output from DMN voxels to other (non-DMN) voxels. Top five most significant connections (Z-score > 6) and distance > 0 of directed SCA are visualized. c, Functional flat mapping of eight DMN ROIs. Lines indicate input to DMN voxels from other (non-DMN) voxels. Top five most significant connections (Z-score > 6) and distance > 0 of directed SCA are visualized. Download Fig. 7-1, TIF file.

Fig. 7-2

Max projection of directed and undirected SCAs in the human DMN. a, Undirected SCA of seed voxels in eight DMN ROIs were plotted in the 1st and 2nd rows. The max-projection in each voxel was used on the functional flat mapping. The higher value of two thresholds (P < 0.05 and Bonferroni correction [T > 4.61], and the top 5% of the T-value matrix [T > 18.82] because of 145 retrograde tracer results in the common marmoset) was applied to show the color range. b, Directed SCA results (output) of seed voxels in eight DMN ROIs were plotted in the 3rd and 4th rows. The higher value of two thresholds (P < 0.05 and Bonferroni correction [Z > 4.61], and the top 5% of T-value matrix [Z > 2.33]) was used to show the color range. c, Directed SCA results (input) of seed voxels in eight DMN ROIs were plotted in the 5th and 6th rows. The higher value of two thresholds was applied to show the color range. Download Fig. 7-2, TIF file.

Analysis of the human DMN using graph analytics

Directed and undirected graphs were obtained from the adjacency matrices generated by directed and undirected SCA. Graphs of human DMN voxels were visualized by the SFDP algorithm (Hu, 2005; Fig. 8a,c) and the Nested Stochastic Block Model (Peixoto, 2014; Fig. 8b,d). Results showed interesting relationships among the DMN regions. Firstly, the hierarchical circle graph of undirected SCA (Fig. 8b) did not show a clear separation among DMN regions where large clusters (i.e., DMN-1, DMN-3) contained most of the DMN regions (regional voxels were well clustered into smaller clusters, but for analysis they were combined into one large cluster). Thus, it was difficult to make functional inferences from these clusters. In contrast, the hierarchical circle graph of directed SCA (Fig. 8d) showed three big regional clusters, namely, the PCC–PPC–mPFC, dlPFC R-temporal and dlPFC L-temporal clusters, and regional voxels were mixed into smaller clusters. The SFDP graph also showed a clear separation between PCC–PPC–mPFC and other regions (Fig. 8c). Thus, we think that these clusters might be functionally separated: PCC–PPC–mPFC form the core DMN regions, and dlPFC R-temporal and dlPFC L-temporal are part of the language network, working together with the DMN. We investigate this idea further in the discussion section. For comparison, we also performed the same analysis at the ROI level (Shen 268 parcellation; Shen et al., 2013)—see Extended Data Figure 8-2. This result showed the difference in clustering between voxel-based and ROI-based analyses. The ROI-based analysis showed slightly separated clusters of dlPFC-temp, PCC, and mPFC–PPC as in the voxel-based analysis. However, the PCC cluster was closer to the dlPFC L-temp cluster than to others (Extended Data Fig. 8-2d, red dashed line). Figure 8, c and d, showed a clear separation between PCC–PPC–mPFC and others. Then, which shows the correct result, voxel-based or ROI-based? Extended Data Figure 1-5 shows the difference between voxel-based and ROI-based time-series analysis. Firstly, the ROI-based time-series is not representative of the set of voxel time-series due to their variability (Extended Data Fig. 1-5b). Furthermore, the ROI-based MTG and MTF relationship of R = 0.435 cannot capture the detail of the voxel-level relationships (Extended Data Fig. 1-5c,d). Thus, we believe that the voxel-based result in Figure 8 should be more correct than result shown in Extended Data Figure 8-2, showing the benefit of using our new methodology.

Fig. 8-1

A simplified diagram of the human default mode network and language network. Based on Fig. 7, 8 and Extended Data Fig. 7-2, this diagram depicts the DMN regions (PPC-PCC-mPFC), language network (dlPFC-temporal) and cerebellum (Crus I & II) and their relationships. Download Fig. 8-1, TIF file.

Fig. 8-2

ROI-based Graph analytics of directed and undirected SCA of the human DMN. a, Graph visualization of an adjacency matrix (DMN ROIs) from undirected SCA. The matrix was acquired by taking a threshold of T > 4.82. A directed graph was visualized by the Scalable Force-Directed Placement (SFDP) algorithm. Node color is the same as Fig. 6. b, Hierarchical circle graph of DMN ROIs based on undirected SCA result was visualized by the Nested Stochastic Block Model. c, Graph visualization of an adjacency matrix (DMN ROIs) of directed SCA. The matrix was acquired by taking a threshold of Z > 3.56. A directed graph was visualized by the Scalable SFDP algorithm. d, Hierarchical circle graph of DMN ROIs based on directed SCA was visualized by the Nested Stochastic Block Model. nine clusters were observed in this analysis. Download Fig. 8-2, TIF file.

Figure 8.
Figure 8.

Graph analytics of directed and undirected SCA of the human DMN. a, Graph visualization of an adjacency matrix (DMN voxels) from undirected SCA. The matrix was acquired by taking a threshold of T > 18.82. A directed graph was visualized by the SFDP algorithm. Node color is the same as Figure 6b; the hierarchical circle graph of DMN voxels based on undirected SCA result was visualized by the Nested Stochastic Block Model. c, Graph visualization of an adjacency matrix (DMN voxels) of directed SCA. The matrix was acquired by taking a threshold of Z > 4.61. A directed graph was visualized by the SFDP algorithm. d, Hierarchical circle graph of DMN voxels based on directed SCA was visualized by the Nested Stochastic Block Model. Three clusters were observed in this analysis. See Extended Data Figure 8-1 for a simplified diagram of the human DMN and language network.

Discussion

Elucidating the relationship between the brain's structural and FC (SC and FC, respectively) is a major issue in the field of neuroscience. Stafford et al. (2014) carried out FC–SC (in this case cellular connectivity) detection for 168 mouse cortical regions and obtained 78.26% sensitivity and 81.69% specificity for the top 1% of SC. Hori et al. (2020) also investigated FC–SC detection for 55 cortical regions of the marmoset brain left hemisphere, achieving an AUC of 0.72. We used the same dataset of cellular connectivity (Majka et al., 2020) and were able to improve on the AUC [0.832 (>0 connections)] for FC–SC detection at the voxel level. There are several reasons why we were able to achieve a higher AUC: our spatial smoothing and nuisance factor removal methods were more optimized than in the previous studies, and awake marmoset data were used instead of anesthetized marmoset or mouse data. Figure 2a in Stafford et al. (2014) showed that a higher SC threshold could achieve a higher detection accuracy, and our results were consistent with it. We found that the stronger cellular connections were more accurately detected by directed and undirected SCA [Fig. 3a,c; AUC, 0.95 (with cell count of >100 at each voxel)]. Our results suggest that the weaker connections may be confused with indirect connections or weaker cellular connections have more interindividual variability (Extended Data Fig. 1-4b) and such connections may not correlate with mixed-effects (group) FC results.

Another issue arises when studying the correlation between FC and SC. Honey et al. (2010) made a direct comparison between SC derived from DWI and FC derived from rs-fMRI for the same set of five human subjects. Scatterplots of SC and the corresponding FC, with 998 ROI parcels, showed a strong correlation of r = 0.54 (p << 10−6). Baum et al. (2020) also calculated the FC–SC correlation using DWI and n-back task fMRI data from a sample of 727 human individuals. The Spearman rank correlation was calculated in 400 cortical region parcellations, and the mean coupling correlation coefficient (727 individuals), in each parcellation, ranged from 0 to 0.42. These two studies used different types of FC, namely, rs- and n-back task–state FC, and their results were noticeably different. Basically, the FC–SC correlation may vary depending on FC state, as FC essentially changes depending on task state or rs, and the SC is thought to be invariant. In our study, we investigated the FC–SC correlation between cellular connectivity and FC in the marmoset left cortex at the voxel level. The full scatterplot of cellular connectivity (cell count of >0 at each voxel) and corresponding FC in 145 injections × 9,862 voxels showed a weak correlation of r = 0.282 (Fig. 1d). Because tracer injections were made into a small portion of a region and into various cortical layers, it appears that results were scattered for each region (Fig. 1g). We consider these results (a weak FC–SC correlation and discrepancies in gradient) will be limited by group comparisons. The invasive tracer dataset was obtained from tracer injections in a cohort of 52 marmosets (Majka et al., 2020), and noninvasive FC and gradient results were obtained from the mixed-effects of 48 sessions in four marmosets. Therefore, when comparing between cellular connectivity which has interindividual variability (Extended Data Fig. 1-4) and mixed-effects FC results, the correlation might become low (Fig. 1d). Although invasive tracer injection data is reliable, group comparisons will be affected due to interindividual variability. Next, we calculated the principal gradient of the marmoset left cortex (Fig. 1e), and the relationship between FC–SC correlation and gradient value was investigated (Fig. 1g,h). Although previous studies (Preti and Van De Ville, 2019; Vázquez-Rodríguez et al., 2019; Baum et al., 2020) showed that “structure and function are closely aligned in unimodal cortex, but diverge in transmodal cortex,” the results of voxel-based analysis did not show this tendency. However, we found a strong connection between our marmoset results and human studies. Baum et al. (2020) showed transmodal and strong SC–FC coupling in the PCC region (Brodmann 23, 31 areas), and this result is consistent in the marmoset. On the other hand, many regions (B20, 21, 39, 45, 47 areas) showed transmodal and SC–FC decoupling in the human cortex, consistent with regions extended by evolution (Baum et al., 2020; Fig. 2e). The marmoset has smaller temporal regions (B20, 21) and does not have B39, 45, and 47 areas. An anticorrelation between the gradient and SC–FC coupling is caused by regions extended by evolution, and in the marmoset there was no such relationship like in the human PCC region.

There are several existing directional analysis studies of the human DMN. Miao et al. (2010) and Jiao et al. (2011) used pwGC to investigate the directional influence among DMN regions. Their results suggest that the PCC is strongly targeted from the temporal cortex but is not targeted from the inferior parietal cortex (IPL) or angular gyrus. However, multivariate Granger causality analysis showed strong directional influence from the left IPL to PCC and from the PCC to right IPL (Cao et al., 2012). Dynamic causal modeling analysis showed positive effective connectivity from the PCC to the left and right IPL (Esménio et al., 2019). Our study showed that many DMN regions were highly interconnected (both input and output), and voxels in the temporal lobe had stronger inputs than outputs. Unfortunately, these previous studies showed different trends. Therefore, in this research, we performed a more detailed analysis at the voxel level rather than at the regional level. Since time-series at a regional level are computed by averaging over several hundred or thousand voxels, it is likely that many time-dependent data points would be missed (Extended Data Fig. 1-5). Therefore, in a validation experiment using tracer injection as the ground truth, our directed SCA method was able to achieve a high accuracy of AUC, 0.942 (Fig. 1c), over previous studies. Additionally, previous studies only treated four to eight nodes inside the DMN, but our study comprehensively investigated 24,860 voxels inside and outside the DMN. We believe that our method provides a more accurate and detailed estimation of directional influence.

The temporal lobe, especially in Brodmann areas (BA) 21 and 22 (Fig. 6a), is also involved in the human DMN (Buckner et al., 2008; Yeo et al., 2011). However, BA22 contains Wernicke's area (Friederici, 2011) and is included in the brain's language network (Mineroff et al., 2018; Lipkin et al., 2022). Additionally, the dlPFC region contains part of Broca's area (BA44, 45; Fig. 6a), and Broca's area is included in the language network (Friederici, 2011; Mineroff et al., 2018; Lipkin et al., 2022). Mineroff et al. (2018) showed that language localizer tasks strongly activated the temporal lobe and inferior frontal gyrus (IFG, BA44, 45), but other DMN regions were not activated. On the other hand, in spatial working memory tasks, the DMN regions were strongly suppressed, but IFG was not suppressed, and the temporal lobe was weakly suppressed. Therefore, we can assume that the function of the temporal lobe and dlPFC regions differ from the other DMN regions. Interestingly, Mineroff et al. defined only PCC–PPC–mPFC as the DMN, while the temporal lobe and dlPFC were defined as the language network and several studies have incorporated this definition (Ji et al., 2019; Gordon et al., 2020). Even in our graph analysis of directed SCA (Fig. 8c,d) using rs-fMRI data, a clear separation was observed between PCC–PPC–mPFC and other regions. Thus, we think PCC–PPC–mPFC forms the core DMN, and dlPFC-temporal is part of the language network, working together with the core DMN regions (Fig. 7a). Extended Data Figure 8-1 summarizes this and the results of Extended Data Figure 7-2 in a simplified diagram. In this way, directional, asynchronous analysis using a time lag can capture features that cannot be separated by synchronous analysis using zero-lag (Fig. 8). These complementary techniques can be used together to provide a more in-depth analysis.

The cerebellum has long been known to be involved in the human DMN (Krienen and Buckner, 2009; Fig. 6d). Gordon et al. (2020) divided the DMN into nine subnetworks, and they showed that the left dlPFC, left PPC, and right Crus I-II constituted one subnetwork, and the right dlPFC, right PPC, and left Crus I-II constituted another subnetwork. Krienen and Buckner (2009) showed bilateral seed regions in the dlPFC correlated with the opposite hemisphere regions in Crus I-II of the posterior cerebellum. Additionally, a reading task activated the language network (dlPFC-temporal) and Crus I-II regions (Lesage et al., 2017). Our results were very consistent with these prior works, and the strong output was observed from the PPC and dlPFC voxels to Crus I-II on the opposite left and right hemispheres (Fig. 7e). However, we could not observe strong inputs from cerebellar voxels to any DMN and language network voxels (Fig. 7f), and they did not have reciprocal connections as other DMN regions often show. This unidirectional nature is also supported from anatomical perspective: cortical output directly affects the cerebellar cortex via the Pons; however, the output of the cerebellar cortex is sent to the dentate nucleus and then to the thalamus (Clark et al., 2020). Thus, we conclude that Crus I-II could be the output target of the DMN and language networks but may not be part of them.

Finally, our directed SCA and functional flat mapping toolbox is implemented as open-source code and downloadable from https://github.com/takuto-okuno-riken/dirsca. We believe that our analysis toolbox will be useful for future research into elucidating the functional properties of the brain.

Footnotes

  • This research was supported by the program for Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/MINDS) from the Japan Agency for Medical Research and Development, AMED, Grant Numbers JP15dm0207001 and JP24wm0625408 to T.O., JP23wm0625001 to H.O., and JP24bm1223008 to J.H. Data were provided (in part) by the Human Connectome Project, WU-Minn Consortium (Principal Investigators, David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research and by the McDonnell Center for Systems Neuroscience at Washington University.

  • The authors declare no competing financial interests.

  • Correspondence should be addressed to Takuto Okuno at takuto.okuno{at}riken.jp.

References

  1. ↵
    1. Avants BB,
    2. Tustison NJ,
    3. Song G,
    4. Cook PA,
    5. Klein A,
    6. Gee JC
    (2011) A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage 54:2033–2044. https://doi.org/10.1016/j.neuroimage.2010.09.025 pmid:20851191
    OpenUrlCrossRefPubMed
  2. ↵
    1. Baum GL, et al.
    (2020) Development of structure–function coupling in human brain networks during youth. Proc Natl Acad Sci U S A 117:771–778. https://doi.org/10.1073/pnas.1912034117 pmid:31874926
    OpenUrlAbstract/FREE Full Text
  3. ↵
    1. Beckmann CF,
    2. Smith SM
    (2004) Probabilistic independent component analysis for functional magnetic resonance imaging. IEEE Trans Med Imaging 23:137–152. https://doi.org/10.1109/TMI.2003.822821
    OpenUrlCrossRefPubMed
  4. ↵
    1. Behzadi Y,
    2. Restom K,
    3. Liau J,
    4. Liu TT
    (2007) A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage 37:90–101. https://doi.org/10.1016/j.neuroimage.2007.04.042 pmid:17560126
    OpenUrlCrossRefPubMed
  5. ↵
    1. Bonakdarpour B,
    2. Parrish TB,
    3. Thompson CK
    (2007) Hemodynamic response function in patients with stroke-induced aphasia: implications for fMRI data analysis. Neuroimage 36:322–331. https://doi.org/10.1016/j.neuroimage.2007.02.035 pmid:17467297
    OpenUrlCrossRefPubMed
  6. ↵
    1. Buckner RL,
    2. Andrews-Hanna JR,
    3. Schacter DL
    (2008) The brain’s default network anatomy, function, and relevance to disease. Ann N Y Acad Sci 1124:1–38. https://doi.org/10.1196/annals.1440.011
    OpenUrlCrossRefPubMed
  7. ↵
    1. Cao R,
    2. Zhu D,
    3. Huang Q,
    4. Wang W,
    5. Ruan Z
    (2012) Dynamic causality analysis on default mode network. AICI 2012, LNAI 7530, 42–49.
  8. ↵
    1. Chuang K-H,
    2. Lee H-L,
    3. Li Z,
    4. Chang W-T,
    5. Nasrallah FA,
    6. Yeow LY,
    7. Singh KK
    (2019) Evaluation of nuisance removal for functional MRI of rodent brain. Neuroimage 188:694–709. https://doi.org/10.1016/j.neuroimage.2018.12.048
    OpenUrlCrossRefPubMed
  9. ↵
    1. Ciric R, et al.
    (2017) Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. Neuroimage 154:174–187. https://doi.org/10.1016/j.neuroimage.2017.03.020 pmid:28302591
    OpenUrlCrossRefPubMed
  10. ↵
    1. Clark SV,
    2. Semmel ES,
    3. Aleksonis HA,
    4. Steinberg SN,
    5. King TZ
    (2020) Cerebellar-subcortical-cortical systems as modulators of cognitive functions. Neuropsychol Rev 31:422–446. https://doi.org/10.1007/s11065-020-09465-1
    OpenUrl
  11. ↵
    1. Esménio S,
    2. Soares JM,
    3. Oliveira-Silva P,
    4. Zeidman P,
    5. Razi A,
    6. Gonçalves ÓF,
    7. Friston K,
    8. Coutinho J
    (2019) Using resting-state DMN effective connectivity to characterize the neurofunctional architecture of empathy. Sci Rep 9:2603. https://doi.org/10.1038/s41598-019-38801-6 pmid:30796260
    OpenUrlCrossRefPubMed
  12. ↵
    1. Friederici AD
    (2011) The brain basis of language processing: from structure to function. Physiol Rev 91:1357–1392. https://doi.org/10.1152/physrev.00006.2011
    OpenUrlCrossRefPubMed
  13. ↵
    1. Friman O,
    2. Borga M,
    3. Lundberg P,
    4. Knutsson H
    (2004) Detection and detrending in fMRI data analysis. Neuroimage 22:645–655. https://doi.org/10.1016/j.neuroimage.2004.01.033
    OpenUrlCrossRefPubMed
  14. ↵
    1. Gămănuţ R,
    2. Kennedy H,
    3. Toroczkai Z,
    4. Ercsey-Ravasz M,
    5. Van Essen DC,
    6. Knoblauch K,
    7. Burkhalter A
    (2018) The mouse cortical connectome, characterized by an ultra-dense cortical graph, maintains specificity by distinct connectivity profiles. Neuron 97:698–715.e10. https://doi.org/10.1016/j.neuron.2017.12.037 pmid:29420935
    OpenUrlCrossRefPubMed
  15. ↵
    1. Gordon EM, et al.
    (2020) Default-mode network streams for coupling to language and control systems. Proc Natl Acad Sci U S A 117:17308–17319. https://doi.org/10.1073/pnas.2005238117 pmid:32632019
    OpenUrlAbstract/FREE Full Text
  16. ↵
    1. Granger C
    (1969) Investigating causal relations by econometric models and cross-spectral methods. Econometrica 37:424–438. https://doi.org/10.2307/1912791
    OpenUrlCrossRef
  17. ↵
    1. Harris JA,
    2. Oh SW,
    3. Zeng H
    (2012) Adeno-associated viral vectors for anterograde axonal tracing with fluorescent proteins in nontransgenic and cre driver mice. Curr Protoc Neurosci 59:1.20.1–1.20.18. https://doi.org/10.1002/0471142301.ns0120s59
    OpenUrlCrossRefPubMed
  18. ↵
    1. Holmes AP,
    2. Friston KJ
    (1998) Generalisability, random effects & population inference. Neuroimage 7:S754. https://doi.org/10.1016/S1053-8119(18)31587-8
    OpenUrlCrossRef
  19. ↵
    1. Honey CJ,
    2. Thivierge J-P,
    3. Sporns O
    (2010) Can structure predict function in the human brain? Neuroimage 52:766–776. https://doi.org/10.1016/j.neuroimage.2010.01.071
    OpenUrlCrossRefPubMed
  20. ↵
    1. Hori Y,
    2. Schaeffer DJ,
    3. Gilbert KM,
    4. Hayrynen LK,
    5. Clery JC,
    6. Gati JS,
    7. Menon RS,
    8. Everling S
    (2020) Comparison of resting-state functional connectivity in marmosets with tracer-based cellular connectivity. Neuroimage 204:116241. https://doi.org/10.1016/j.neuroimage.2019.116241
    OpenUrlCrossRefPubMed
  21. ↵
    1. Hu Y
    (2005) Efficient and high quality force-directed graph drawing. Math J 10:37–71.
    OpenUrl
  22. ↵
    1. Iriki A
    (2017) The marmoset MRI standard brain. Available at: https://brainatlas.brain.riken.jp/marmoset/modules/xoonips/detail.php?id=004
  23. ↵
    1. Ji JL,
    2. Spronk M,
    3. Kulkarni K,
    4. Repovs G,
    5. Anticevic A,
    6. Cole MW
    (2019) Mapping the human brain's cortical-subcortical functional network organization. Neuroimage 185:35–57. https://doi.org/10.1016/j.neuroimage.2018.10.006 pmid:30291974
    OpenUrlCrossRefPubMed
  24. ↵
    1. Jiao Q,
    2. Lu G,
    3. Zhang Z,
    4. Zhong Y,
    5. Wang Z,
    6. Guo Y,
    7. Li K,
    8. Ding M,
    9. Liu Y
    (2011) Granger causal influence predicts BOLD activity levels in the default mode network. Hum Brain Mapp 32:154–161. https://doi.org/10.1002/hbm.21065 pmid:21157880
    OpenUrlCrossRefPubMed
  25. ↵
    1. Krienen FM,
    2. Buckner RL
    (2009) Segregated fronto-cerebellar circuits revealed by intrinsic functional connectivity. Cereb Cortex 19:2485–2497. https://doi.org/10.1093/cercor/bhp135 pmid:19592571
    OpenUrlCrossRefPubMed
  26. ↵
    1. Lesage E,
    2. Hansen PC,
    3. Miall RC
    (2017) Right lateral cerebellum represents linguistic predictability. J Neurosci 37:6231–6241. https://doi.org/10.1523/JNEUROSCI.3203-16.2017 pmid:28546307
    OpenUrlAbstract/FREE Full Text
  27. ↵
    1. Lipkin B, et al.
    (2022) Probabilistic atlas for the language network based on precision fMRI data from >800 individuals. Sci Data 9:529. https://doi.org/10.1038/s41597-022-01645-3 pmid:36038572
    OpenUrlCrossRefPubMed
  28. ↵
    1. Majka P, et al.
    (2020) Open access resource for cellular-resolution analyses of corticocortical connectivity in the marmoset monkey. Nat Commun 11:1133. https://doi.org/10.1038/s41467-020-14858-0 pmid:32111833
    OpenUrlCrossRefPubMed
  29. ↵
    1. Marcus DS,
    2. Harwell J,
    3. Olsen T,
    4. Hodge M,
    5. Glasser MF,
    6. Prior F,
    7. Jenkinson M,
    8. Laumann T,
    9. Curtiss SW,
    10. Van Essen DC
    (2011) Informatics and data mining tools and strategies for the human connectome project. Front Neuroinform 5:4. https://doi.org/10.3389/fninf.2011.00004 pmid:21743807
    OpenUrlCrossRefPubMed
  30. ↵
    1. Margulies DS, et al.
    (2016) Situating the default-mode network along a principal gradient of macroscale cortical organization. Proc Natl Acad Sci U S A 113:12574–12579. https://doi.org/10.1073/pnas.1608282113 pmid:27791099
    OpenUrlAbstract/FREE Full Text
  31. ↵
    1. McInnes L,
    2. Healy J,
    3. Melville J
    (2018) UMAP: uniform manifold approximation and projection for dimension reduction. arXive, 1802.03426.
  32. ↵
    1. Miao X,
    2. Chen K,
    3. Li R,
    4. Wen X,
    5. Yao L,
    6. Wu X
    (2010) Application of granger causality analysis to effective connectivity of the default-mode network. IEEE/ICME Int. Conf. on Comp. Med. Eng.
  33. ↵
    1. Mikl M,
    2. Mareček R,
    3. Hluštík P,
    4. Pavlicová M,
    5. Drastich A,
    6. Chlebus P,
    7. Brázdil M,
    8. Krupa P
    (2008) Effects of spatial smoothing on fMRI group inferences. Magn Reson Imaging 26:490–503. https://doi.org/10.1016/j.mri.2007.08.006
    OpenUrlCrossRefPubMed
  34. ↵
    1. Mineroff Z,
    2. Blank IA,
    3. Mahowald K,
    4. Fedorenko E
    (2018) A robust dissociation among the language, multiple demand, and default mode networks: evidence from inter-region correlations in effect size. Neuropsychologia 119:501–511. https://doi.org/10.1016/j.neuropsychologia.2018.09.011 pmid:30243926
    OpenUrlCrossRefPubMed
  35. ↵
    1. Muta K, et al.
    (2022) Effect of sedatives or anesthetics on the measurement of resting brain function in common marmosets. Cereb Cortex 33:5148–5162. https://doi.org/10.1093/cercor/bhac406 pmid:36222604
    OpenUrlPubMed
  36. ↵
    1. Oh S, et al.
    (2014) A mesoscale connectome of the mouse brain. Nature 508:207–214. https://doi.org/10.1038/nature13186 pmid:24695228
    OpenUrlCrossRefPubMed
  37. ↵
    1. Okano H
    (2021) Current status of and perspectives on the application of marmosets in neurobiology. Annu Rev Neurosci 44:27–48. https://doi.org/10.1146/annurev-neuro-030520-101844
    OpenUrlCrossRefPubMed
  38. ↵
    1. Okano H,
    2. Miyawaki A,
    3. Kasai K
    (2015) Brain/MINDS: brain-mapping project in Japan. Philos Trans R Soc Lond B Biol Sci 370:20140310. https://doi.org/10.1098/rstb.2014.0310 pmid:25823872
    OpenUrlCrossRefPubMed
  39. ↵
    1. Okuno T,
    2. Ichinohe N,
    3. Woodward A
    (2024) A reappraisal of the default mode and frontoparietal networks in the common marmoset brain. Front Neuroimaging 2:1345643. https://doi.org/10.3389/fnimg.2023.1345643 pmid:38264540
    OpenUrlCrossRefPubMed
  40. ↵
    1. Okuno T,
    2. Woodward A
    (2021) Vector auto-regressive deep neural network: a data-driven deep learning-based directed functional connectivity estimation toolbox. Front Neurosci 15:764796. https://doi.org/10.3389/fnins.2021.764796 pmid:34899167
    OpenUrlCrossRefPubMed
  41. ↵
    1. Parkes L,
    2. Fulcher B,
    3. Yücel M,
    4. Fornito A
    (2018) An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI. Neuroimage 171:415–436. https://doi.org/10.1016/j.neuroimage.2017.12.073
    OpenUrlCrossRefPubMed
  42. ↵
    1. Peixoto TP
    (2014) Hierarchical block structures and high-resolution model selection in large networks. Phys Rev X 4:011047. https://doi.org/10.1103/PhysRevX.4.011047
    OpenUrl
  43. ↵
    1. Penny W,
    2. Friston K,
    3. Ashburner J,
    4. Kiebel S,
    5. Nichols T
    (2007) Statistical parametric mapping: the analysis of functional brain images. London: Elsevier.
  44. ↵
    1. Preti MG,
    2. Van De Ville D
    (2019) Decoupling of brain function from structure reveals regional behavioral specialization in humans. Nat Commun 10:4747. https://doi.org/10.1038/s41467-019-12765-7 pmid:31628329
    OpenUrlCrossRefPubMed
  45. ↵
    1. Pruim RH,
    2. Mennes M,
    3. van Rooij D,
    4. Llera A,
    5. Buitelaar JK,
    6. Beckmann CF
    (2015) ICA-AROMA: a robust ICA-based strategy for removing motion artifacts from fMRI data. Neuroimage 112:267–277. https://doi.org/10.1016/j.neuroimage.2015.02.064
    OpenUrlCrossRefPubMed
  46. ↵
    1. Shen X,
    2. Tokoglu F,
    3. Papademetris X,
    4. Constable RT
    (2013) Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. Neuroimage 82:403–415. https://doi.org/10.1016/j.neuroimage.2013.05.081 pmid:23747961
    OpenUrlCrossRefPubMed
  47. ↵
    1. Skibbe H, et al.
    (2022) The brain/MINDS marmoset connectivity atlas: exploring bidirectional tracing and tractography in the same stereotaxic space. bioRxiv.
  48. ↵
    1. Škoch A,
    2. Bučková BR,
    3. Mareš J,
    4. Tintěra J,
    5. Sanda P,
    6. Jajcay L,
    7. Horáček J,
    8. Španiel F,
    9. Hlinka J
    (2022) Human brain structural connectivity matrices–ready for modelling. Sci Data 9:486. https://doi.org/10.1038/s41597-022-01596-9 pmid:35945231
    OpenUrlCrossRefPubMed
  49. ↵
    1. Stafford JM, et al.
    (2014) Large-scale topology and the default mode network in the mouse connectome. Proc Natl Acad Sci U S A 111:18745–18750. https://doi.org/10.1073/pnas.1404346111 pmid:25512496
    OpenUrlAbstract/FREE Full Text
  50. ↵
    1. Van Essen DC,
    2. Smith SM,
    3. Barch DM,
    4. Behrens TE,
    5. Yacoub E,
    6. Ugurbil K
    , for the WU-Minn HCP Consortium (2013) The WU-Minn human connectome project: an overview. Neuroimage 80:62–79. https://doi.org/10.1016/j.neuroimage.2013.05.041 pmid:23684880
    OpenUrlCrossRefPubMed
  51. ↵
    1. Vázquez-Rodríguez B,
    2. Suárez LE,
    3. Markello RD,
    4. Shafiei G,
    5. Paquola C,
    6. Hagmann P,
    7. Van Den Heuvel MP,
    8. Bernhardt BC,
    9. Spreng RN,
    10. Misic B
    (2019) Gradients of structure–function tethering across neocortex. Proc Natl Acad Sci U S A 116:21219–21227. https://doi.org/10.1073/pnas.1903403116 pmid:31570622
    OpenUrlAbstract/FREE Full Text
  52. ↵
    1. Vos de Wael R, et al.
    (2020) Brainspace: a toolbox for the analysis of macroscale gradients in neuroimaging and connectomics datasets. Commun Biol 3:103. https://doi.org/10.1038/s42003-020-0794-7 pmid:32139786
    OpenUrlCrossRefPubMed
  53. ↵
    1. Wang Q, et al.
    (2020) The Allen Mouse Brain common coordinate framework: a 3D reference atlas. Cell 181:936–953.e20. https://doi.org/10.1016/j.cell.2020.04.007 pmid:32386544
    OpenUrlCrossRefPubMed
  54. ↵
    1. Whitfield-Gabrieli S,
    2. Nieto-Castanon A
    (2012) Conn: a functional connectivity toolbox for correlated and anticorrelated brain networks. Brain Connect 2:125–141. https://doi.org/10.1089/brain.2012.0073
    OpenUrlCrossRefPubMed
  55. ↵
    1. Yen CC,
    2. Papoti D,
    3. Silva AC
    (2018) Investigating the spatiotemporal characteristics of the deoxyhemoglobin-related and deoxyhemoglobin-unrelated functional hemodynamic response across cortical layers in awake marmosets. Neuroimage 164:121–130. https://doi.org/10.1016/j.neuroimage.2017.03.005 pmid:28274833
    OpenUrlCrossRefPubMed
  56. ↵
    1. Yeo BT, et al.
    (2011) The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 106:1125–1165. https://doi.org/10.1152/jn.00338.2011 pmid:21653723
    OpenUrlCrossRefPubMed
  57. ↵
    1. Yushkevich PA,
    2. Piven J,
    3. Hazlett HC,
    4. Smith RG,
    5. Ho S,
    6. Gee JC,
    7. Gerig G
    (2006) User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31:1116–1128. https://doi.org/10.1016/j.neuroimage.2006.01.015
    OpenUrlCrossRefPubMed
Back to top

In this issue

Email
Print
View Full Page PDF
Citation Tools
Respond to this article
A Novel Directed Seed-Based Connectivity Analysis Toolbox Applied to Human and Marmoset Resting-State FMRI
Takuto Okuno, Junichi Hata, Chino Kawai, Hideyuki Okano, Alexander Woodward
Journal of Neuroscience 6 November 2024, 44 (45) e0389242024; DOI: 10.1523/JNEUROSCI.0389-24.2024
Twitter logo Facebook logo Mendeley logo

Jump to section

Keywords

Responses to this article

Respond to this article

Jump to comment:

No eLetters have been published for this article.