Fig. 1
Flowchart of the analysis pipeline
The Influence of Parameters on Results
Each part of the analysis pipeline (clustering, labelling, connectivity inference) corresponds to a distinct and autonomous module of the method. In this respect, the influence of each parameter is confined inside the module it belongs to and the impact it has on the whole is exerted only through the output of that specific module. As illustrated in Fig. 2, the structure of the code affords significant flexibility to the user permitting relatively straightforward expansion and modification capacity. For instance, in the connectivity inference stage, instead of using the cross-correlation one could use a multitude of different measures (transfer entropy, Granger causality, DCM, coherence etc.). The replacement of the final module would not affect the previous stages of the pipeline.
Fig. 2
List of the parameters that affect the clustering module (upper panel), labelling module (middle panel), and connectivity module (lower panel) of the analysis pipeline
In the data preprocessing stage all the parameters are involved in the definition of the activation windows of each source. Higher values will result in the exclusion of more sources and in lower number/duration of activation windows. The parameters of the clustering algorithm affect the size of the resulting clusters. Increasing the search radius and lowering the minimum temporal overlap will result in bigger and more disperse clusters. The same holds for the parameters of the cluster merging test. Higher values of the search radius and lower of the minimum cross-correlation will output larger source clusters.
In the labelling phase major role plays, as expected, the cortical parcellation of the employed atlas. The naming rules affect merely the ease with which the label is given to each cluster. In the common network discovery, one can simply count the number of clusters in order to identify the most common brain regions. Alternatively, one could correct the count by dividing with the total number of clusters in every region or even introduce penalties that promote the importance of clusters in specific regions of the brain. At the end, different common networks may appear. The number of nodes in the network may also change and results in the exclusion or inclusion of more brain regions. Returning back at the subject level, the minimum cluster size parameter controls the minimum size of clusters one is willing to keep for the subsequent cross-lagged correlation analysis.
In the calculation of the cross-correlation values an obvious parameter is the number of lags on which the calculation takes place. The sections of the actual time-series selected depend on the rules that control the division of each cluster’s activity-peaks into early, middle and late. In the statistical test (which is essentially a permutation method), the number of repetitions affects the precision of the results. Higher number of re-runs makes the results of test more trustworthy but also increases computation time. The statistical significance level reflects the strictness of the test. Finally, after the connectivity paths are constructed, in identifying the most commonly observable, one can just count or use different scoring schemes that take into account the relative importance of the paths inside each subject. In return, different paths will be considered more important and they will be ranked higher.
3 Results
3.1 Application of the Method and Preliminary Results
Figure 3 displays single-subject average, time-locked MEG recordings collapsed across the 248 magnetometer sensors. In Fig. 4 images in the four left hand columns display average source-current distributions in the form of Minimum Norm Estimates (MNE) at four representative time points, projected on the cortical surface of one participant. These maps, thresholded using a minimum current level that was determined for each participant on the basis of their respective baseline source-current estimates, were input to the agglomerative hierarchical clustering. This procedure identified groups of vertices that displayed spatial proximity and highly synchronous estimated activations. The time-series for each cluster was calculated as the mean of the time-series of all the constituent vertices as illustrated in Fig. 5. It is important to note that clustering was performed on the individual-subject data and group-average maps are for viewing purposes only.
Fig. 3
Averaged MEG epochs recorded from one participant (#008) during performance of the naming task. Stimulus onset is at 0 ms. The shaded region indicates the time window submitted to further analyses (0–400 ms post-stimulus onset)
Fig. 4
Average cortical activations (across participants) derived through MNE at 100 (a), 150 (b), 230 (c) and 320 ms (d) after stimulus onset are shown on the left (upper row) and right hemisphere surface (lower row). (e): Anatomical locations of cortical source clusters (collapsed over time) which were identified through the hierarchical, spatiotemporal clustering procedure. Each cluster is composed of spatially contiguous current sources displaying highly inter-correlated time courses
Fig. 5
Time series of estimated activity in each of the 19 vertices that formed cluster no.18 in participant S#008 located in the right supramarginal gyrus; left hand panel). The average time series for that cluster is shown in the right-hand panel
The locations of activity clusters that appeared consistently across subjects are shown in the composite image in the right-hand column of Fig. 4. The anatomical regions (as defined in the Tzourio-Mazoyer atlas available in Brainstorm [36]) where activity clusters were found consistently in at least 4/6 participants, regardless of the latency of significant activity characterizing each cluster, are listed in Table 1.
Table 1
Anatomical regions in the Tzourio-Mazoyer atlas where activity clusters were found consistently across subjects
Middle frontal (R) | Frontal inferior opercular (R) | Angular (R) | Frontal inferior triangular (L) | Middle occipital (R) |
Inferior temporal (L) | Frontal inferior triangular (R) | Inferior temporal (R) | Calcarine (L) | Fusiform (R) |
Middle occipital (L) | Lingual (R) | Precentral (R) | Calcarine (R) | Middle frontal (L) |
Middle temporal (L) | Supra-marginal (R) | Frontal middle orbital (R) | Lingual (L) | Postcentral (R) |
In the single-subject data, the program identified more than one clusters in several brain areas, while certain clusters showed more than one discrete peaks of activity. In the sample participant displayed in Table 2 (S#008) three spatially distinct clusters were found within the anatomical borders of the left inferior temporal gyrus (middle column). The activity time series of two of these clusters (1 and 2) contained two peaks (at 188 and 259 ms for cluster 1 and at 130 and 314 ms for cluster 2). Accordingly, the final table of temporally restricted, spatially distinct clusters for this participant contained 57 entries (10 in the early time window, 24 in the middle time window, and 23 in the late time window).
Table 2
List of spatially distinct and temporally restricted clusters of activity located in the left inferior temporal gyrus for participant S#008
Inferior temporal (L)—1.1 | 188 ms | ||
Inferior temporal (L)—1 | Inferior temporal (L)—1.2 | 259 ms | |
Inferior temporal (L) | Inferior temporal (L)—2 | Inferior temporal (L)—2.1 | 130 ms |
Inferior temporal (L)—3 | Inferior temporal (L)—2.2 | 314 ms | |
Inferior temporal (L)—3.1 | 354 ms |
In the current implementation of the pipeline, autocorrelations between successive time-series segments from the same cluster (indicative of self-loops) or between clusters located in the same anatomical region (indicative of feedback loops) were not considered. Thus for each participant a total of 47–106 clusters were identified and 400–2,800 directed interdependencies (“pathways”) were calculated. Less than 15 % of the possible pairwise interdependencies met the statistical test (at p < .05) and exceeded the r = .70 threshold. The consistency with which these potential functional pathways between clusters appeared across subjects was subsequently assessed by representing each cluster by their respective anatomical area.
The final output of the algorithm returns the connectivity pathways in Fig. 6, which represents a hypothetical network of brain regions each containing at least one cortical patch that showed strong, directed, and reliable interdependence with at least one other brain region in the first 400 ms following presentation of a picture to be named. The actual time-series of clusters that participate in this hypothetical network in a given participant can then be visualized. For instance, Fig. 7 shows the entire time series of the three clusters of activity that contributed to the hypothetical pathway linking the left lateral occipital region, the right lingual gyrus, and the left inferior temporal gyrus in participant S#009.
Fig. 6
Upper panel. Significant associations between regional activation time courses in single-subject analyses which were found consistently in the majority of participants (≥4/6 cases). Each box represents a single source-cluster peaking during the time window shown at the bottom of the panel. Lower panel. Clusters demonstrating significant interdependencies in pairwise cross-lag analyses are marked with oval yellow shapes. g gyrus, L left hemisphere, R right hemisphere, MFG middle frontal gyrus, MTG middle temporal gyrus, ITG inferior temporal gyrus, SMG supramarginal gyrus, IFG inferior frontal gyrus
Fig. 7
Time-series for three clusters forming a hypothetical path (left Lateral Occipital region → right Lingual gyrus (r = .77 at 8 ms time lag) → left inferior temporal gyrus (ITG; r = .72 at 5 ms time lag) in participant S#009. Boxes indicate the sections of each time series on which cross-correlations were computed
4 Discussion
4.1 Discussion of Preliminary Results and Future Directions
By using a method that does not require a priori assumptions regarding the spatiotemporal profiles of activation in individual participants, activation loci associated with object naming were found in the following areas: medial and lateral occipital cortex, fusiform and lingual gyri, the posterior portion of the superior and middle temporal gyri (MTG), the anterior, middle-inferior temporal lobe (ITG), motor cortex, and the inferior (IFG) and middle frontal gyri (MFG). Meta-analytic data revealed hemodynamic activation foci in the majority of these regions. One notable exception was the anterior ITG where several fMRI studies have failed to find significant activation (e.g., [37–39]). This may have been due to increased susceptibility of fMRI signals to distortion by neighboring air-filled cavities. Conversely, activation sites in the anterior and/or middle ITG are common in MEG studies (e.g., [30, 38]). Large-scale voxel-based lesion studies have stressed the key role of this region for semantic processing [40], while anterior temporal atrophy is considered as one of the hallmarks of semantic dementia [41]. Results from hemodynamic studies that succeeded in revealing anterior ITG activation sites show largely overlapping activations for words and pictures, suggesting that perhaps this region maintains amodal semantic representations [42, 43]. Such stored semantic memory entries may result, over time, by convergence of inputs from posterior inferior temporal and occipitotemporal regions as well as inputs from Wernicke’s area. The present results are in line with more recent proposals to revise the classical Broca–Wernicke model [1], which implicates mainly the temporoparietal cortex, the IFG, the angular gyrus, and the posterior portion of the inferior temporal gyrus in naming. Recent reviews and meta-analyses call for further revisions to this model by stressing additional roles of the temporoparietal and inferior frontal areas in the executive control over semantic processing [4].
Connectivity analyses identified a subset of these areas demonstrating reliable (p < .05) and substantial (r > .70) pair-wise interdependencies at time lags between 5 and 80 ms. Although both the extent (cortical surface) and peak source-current were greater in the left hemisphere, several right hemisphere regions were also identified in the connectivity analyses. Regions where dynamic activation patterns appeared to be “affected” by prior activation in primary and association visual cortices (Lateral Occipital, Lingual and Fusiform gyri in Fig. 6) were located in the MTG, anterior ITG, and MFG. The left anterior ITG emerged as one of the sites of converging late activation receiving input primarily from occipitotemporal (lingual gyrus) and right anterior ITG. The IFG (located in the orbital portion of the frontal operculum) emerged as a second converging site, receiving inputs from occipitotemporal cortex (lingual gyrus) through MTG. The former functional association is likely subserved by the inferior longitudinal fasciculus and the latter by fibers in the deep layers of the inferior fronto-occipital fasciculus [44].
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
