A network correspondence toolbox for quantitative evaluation of novel neuroimaging results

Abstract

The brain can be decomposed into large-scale functional networks, but the specific spatial topographies of these networks and the names used to describe them vary across studies. Such discordance has hampered interpretation and convergence of research findings across the field. We have developed the Network Correspondence Toolbox (NCT) to permit researchers to examine and report spatial correspondence between their novel neuroimaging results and multiple widely used functional brain atlases. We provide several exemplar demonstrations to illustrate how researchers can use the NCT to report their own findings. The NCT provides a convenient means for computing Dice coefficients with spin test permutations to determine the magnitude and statistical significance of correspondence among user-defined maps and existing atlas labels. The adoption of the NCT will make it easier for network neuroscience researchers to report their findings in a standardized manner, thus aiding reproducibility and facilitating comparisons between studies to produce interdisciplinary insights.

Fig. 1. Example atlases included in the Network Correspondence Toolbox (NCT).

The NCT is a toolbox that facilitates exploration of network correspondences across multiple functional network atlases as well as quantitative comparison of novel neuroimaging results with multiple atlases. Ten atlases are shown here for illustration purposes. In this example, the Yeo 17-network atlas in fsaverage6 space (center) serves as the reference atlas. All other surrounding atlases in different spaces are projected to the fsaverage6 space to compute Dice overlap coefficients with the reference networks. A full list of currently available atlases in the NCT can be found in https://pypi.org/project/cbig_network_correspondence.

Fig. 2. Hierarchical structures of the network similarity matrix (Dice overlap).

The nested stochastic block model with 15 clusters was applied onto the network similarity matrix where Dice coefficients were calculated between each pair of networks from different atlases. The network similarity matrix is a 230 × 230 matrix since there are 16 atlases with 230 networks. Each block corresponds to a network cluster. Representative network names within each cluster are highlighted here. See Supplementary Fig. 1 for clustering results with atlas associations. We emphasize that the hierarchical clustering is not meant to generate a new consensus atlas; rather it is a way for us to examine convergence and divergence across atlases. Prim primary, med medial, lat lateral, vis visual, assoc association, par parietal, occ occipital, ant anterior, MTL medial temporal lobe, front frontal, ven ventral, attn attention, tem temporal, SM somatomotor, post posterior, SAL salience, subcor subcortical, cing cingulo, operc opercular.

Fig. 3. Spatial topography of the 15 network clusters.

See Supplementary Fig. 1 for atlas contributions and hierarchical relations. Warmer colors indicate greater concordance across atlases. We emphasize that these maps are not meant to be consensus networks, i.e., we are not proposing a new atlas. Instead, these maps are meant to illustrate the convergences and divergences between existing atlases.

Fig. 4. Networks with similar names can show different spatial topographies.

“Salience” networks from four different atlases. These networks are labeled using similar nomenclature across multiple atlases, though they span different, largely non-overlapping anatomical locations.

Fig. 5. Illustration of NCT usage to explore network correspondence between two atlases.

In this example, we explore overlap between the Yeo2011 17-network atlas (reference atlas) and the Gordon2017 17-network atlas. A The user specifies the name of the reference atlas (here, Yeo2011-17) and the name of comparison atlas (here, Gordon2017-17). The Yeo 17-network atlas is in fsaverage6 space, while the Gordon2017 17-network atlas is in fs_LR_32k. In this case, the reference atlas space is fsaverage6. Therefore, the NCT projects the Gordon2017 17-network to fsaverage6 space. B The NCT computes the Dice overlap between networks from Yeo 17-network atlas and Gordon2017 17-network atlas. The NCT also performs a spin test to test whether networks significantly overlap. C The NCT reports the network correspondence between these two atlases using an overlap heatmap plot, where the k-th row and m-th column represents the Dice overlap coefficient between network k of the Yeo 17-network atlas and network m from the Gordon2017 17-network atlas. A high Dice coefficient indicates high overlap between the two networks. Brighter colors indicate higher overlap, darker colors indicate lower overlap. The “*” indicates two networks significantly overlap (p < 0.05). Most Yeo2011 networks overlap with at least one network in the Gordon2017 atlas. The NCT also provides a summary table showing the exact Dice coefficients and p-values (see Supplementary Table 2). The NCT uses network names from the original paper for each atlas. Dors dorsal, attn attention, sal salience, ven ventral, temp temporal, par parietal, lat lateral, vis visual, med medial, SM somatomotor, cing cingulo, operc opercular, ant anterior, MTL medial temporal lobe.

Fig. 6. Illustration of NCT usage to explore network correspondence between user-defined input data and a set of atlases.

In this example, we explore the overlap between a single-dimension input dataset and 4 atlases: the Yeo2011 17-network atlas (“Yeo2011-17”, the Gordon2017 17-network atlas (“Gordon2011-17”), the Glasser2016 360-ROI atlas with Ji2019 12 Cole-Anticevic networks (“Glasser2016-360 + Ji2019-12”), and the Shen2013 268-ROI atlas with 8 networks (“Shen2013-268-8”). A The user provides the input data together with a config file specifying the name, data space (e.g., fs_LR_32k, fsaverage6, FSLMNI2mm), and data type (“Metric” if the input data contains floating values; “Hard” if the input data contains binary values). In this example, the input data is in fs_LR_32k space and contains binary values. The data type is “Hard”. The user also provides an atlas list indicating which atlases to include. B The NCT reads the input data and projects the atlases in the atlas list to the input data space. C The NCT computes the Dice overlap between the input data and networks from the atlases in the list. The NCT also performs spin tests to test whether the input data and networks significantly overlap. D The NCT reports the network correspondence summary using a “Network Clock” plot, a “Network Radar” plot, and a summary table for single-dimensional data. The “Network Clock” provides a visual comparison of the Dice overlap across networks from different atlases. Different colors represent different atlases. The lollipop bars represent the Dice overlap coefficients. Networks significantly overlapping with the input data (p < 0.05) are indicated by the network names. A larger font size represents a larger Dice coefficient. The “Network Radar” shows the Dice overlap across networks within each atlas. Networks significantly overlapping with the input data (p < 0.05) are indicated by “*”. The NCT also provides a summary table showing the exact Dice coefficients and p-values across different atlases (see Supplementary Table 3). The NCT uses network names from the original paper for each atlas. Dors dorsal, attn attention, par parietal, ven ventral, sal salience, subcor subcortical, cing cingulo, operc opercular, vent ventral, multi multimodal, orbit orbital, front frontal.

Fig. 7. Illustration of usage of the NCT to explore network correspondence between user-defined input data and multiple atlases: HCP Example 1.

In this example, we explore overlap between the HCP working memory task contrast (2BK vs. 0BK) and networks from 8 atlases. The user provides the input data together with a config file specifying the name, data space, data type, and the threshold which is used to threshold the input data. The user also provides an atlas list indicating which atlases to include. The NCT thresholds the data based on the config file and projects the atlases in the atlas list to the input data space. The NCT computes the Dice overlap between the thresholded data and all networks from atlases in the list. The NCT also performs spin tests to test whether the thresholded input data and networks significantly overlap. The NCT reports the network correspondence summary using a “Network Clock” plot (middle), a “Network Radar” plot (top right), and a summary table showing the exact Dice coefficients and p-values for all atlases (see Supplementary Table 3). The “Network Clock” provides a visual comparison of Dice overlap across networks from different atlases. Different colors represent different atlases. The lollipop bars represent the Dice overlap coefficients. Networks significantly overlapping with the thresholded input data (p < 0.05) are indicated by the network names. Larger font sizes represent larger Dice coefficients. The “Network Radar” shows the Dice overlap across networks within each atlas. Networks significantly overlapping with the thresholded input data (p < 0.05) are indicated by “*” in the “Network Radar”. WM working memory, BK back, dors dorsal, attn attention, par parietal, ant anterior, sal salience, LECN left executive control network, RECN right executive control network, DMN default mode network, ven ventral, subcor subcortical, mot motor, vis visual, cog cognition, occ occipital, front frontal, cing cingulo, operc opercular, med medial, lat lateral, ant anterior, MTL medial temporal lobe, SM somatomotor, post posterior.

Fig. 8. Illustration of NCT usage to explore network correspondence between user-defined input data and multiple atlases: HCP Example 2.

In this example, we explore the overlap between the HCP social task contrast (theory of mind vs. random) and 8 atlases. The user provides the input data together with a config file specifying the name, data space, data type, and the threshold which is used to threshold the input data. The user also provides an atlas list indicating which atlases to include. The NCT thresholds the data based on the config file and projects the atlases in the atlas list to the input data space. The NCT computes the Dice overlap between the thresholded data and all networks from atlases in the list. The NCT also performs spin tests to test whether the thresholded input data and networks significantly overlap. The NCT reports the network correspondence summary using a “Network Clock” plot (middle), a “Network Radar” plot (top right), and a summary table showing the exact Dice coefficients and p-values for all atlases (see Supplementary Table 4). The “Network Clock” provides a visual comparison of Dice overlap across networks from different atlases. Different colors represent different atlases. The lollipop bars represent the Dice overlap coefficients. Networks significantly overlapping with the thresholded input data (p < 0.05) are indicated by the network names. Larger font sizes represent larger Dice coefficients. The “Network Radar” shows the Dice overlap across networks within each atlas. Networks significantly overlapping with the thresholded input data (p < 0.05) are indicated by “*” in the “Network Radar”. Dors dorsal, attn attention, post posterior, multi multimodal, temp temporal, par parietal, med medial, front frontal, ven ventral occ occipital, lat lateral, vis visual, cing cingulo, operc opercular, ant anterior, MTL medial temporal lobe, SM somatomotor.

Fig. 9. Illustration of NCT usage to explore network correspondence between user-defined input data and multiple atlases: UKB Example 1.

In this example, we explore overlap between UKB ICA component 5 and 8 representative atlases. The UKB ICA z-stat maps with 21 good components (https://www.fmrib.ox.ac.uk/ukbiobank/group_means/rfMRI_GoodComponents_d25_v1.txt) were thresholded by FSL melodic mixture-modeling threshold 0.6. The component 16 corresponds to the cerebellum and was further excluded, resulting in 20 thresholded UKB ICA maps. The 20 thresholded UKB ICA maps can be found in Supplementary Fig. 1. The NCT also provides a summary table showing the exact Dice coefficients and p-values for all atlases (Supplementary Table 5). Med medial, vis visual, lat lateral, prim,primary, cing cingulo, operc opercular, SM somatomotor, dors dorsal, attn attention, par parietal, front,frontal, post posterior, ant anterior, MTL medial temporal lobe.

Fig. 10. Illustration of NCT usage to explore network correspondence between user-defined input data and multiple atlases: UKB Example 2.

In this example, we explore overlap between UKB ICA component 3 and 8 representative atlases. The UKB ICA z-stat maps with 21 good components (https://www.fmrib.ox.ac.uk/ukbiobank/group_means/rfMRI_GoodComponents_d25_v1.txt) were thresholded by FSL melodic mixture-modeling threshold 0.6. The component 16 corresponds to the cerebellum and was further excluded, resulting in 20 thresholded UKB ICA maps. The 20 thresholded UKB ICA maps can be found in Supplementary Fig. 1. The NCT also provides a summary table showing the exact Dice coefficients and p-values for all atlases (Supplementary Table 6). Cing cingulo, operc opercular, dors dorsal, attn attention, mot motor, vis visual, assoc association, sal salience, ven ventral, post posterior, front frontal, par parietal; SM somatomotor, ant anterior, MTL medial temporal lobe, lat lateral, med medial.