Figure 1 shows the spatial distribution of activation for the maps in NeuroVault: the of how many times a voxel appeared in a statistical map with a Z or a T greater than 3. The distribution is strickingly non-uniform throughout gray matter. In particular, the regions most represented are the frontal part of the insula and dorsal ACC, that form a well-known cingula-insulate control network associated with salience processing [Seeley 2007, Dissociable Intrinsic Connectivity Networks for Salience Processing and Executive Control]. The other structures highlighted in figure 1 are the IPS, the , --regions sometimes called the "task positive network" [Fox 2005, The human brain is intrinsically organized into dynamic, anticorrelated functional networks]-- as well as the occipital lobe, encompassing the visual cortex. The presence of the latter can be explained by the fact that many experiments rely on visual stimuli. Interestingly, the networks that stand out on this map are largely related to attention and executive control.
Figure 2 shows a similar map computed from a coordinate-based meta analysis database, NeuroSynth, that collects coordinates of activation foci from the literature [Yarkoni ...]. It displays a similar density of activation, with visible attentional networks. However, the visual cortex is much less present, possibly because researchers tend to report in publications coordinates of high-level contrasts that cancel out low-level effects of stimuli. On the opposite, the NeuroVault database contains a variety of contrasts, including task-versus-baseline maps.
From a global perspective, to understand better the spatial distribution of detections, it is also interesting to look at the average activation across all the database. On figure 3 we give the overall average of the statistical maps.
Unlike a simple count of statistically-significant detections, as in a coordinnate-based meta-analysis, this analysis also captures the dominant sign of the activation, accumulating power in regions that may not cross threshold in individual analyses1. The average maps clearly shows regions that responds, on average, by a deactivation, rather than an activation. These span the default-mode network, that was historically discovered in a similar analysis, noting a decrease in activity across a variety of tasks [Shulman 1997, Common blood flow changes across visual tasks: II. Decreases in cerebral cortex].
One challenge to run finer analysis on data hosted on neurovault is that, unlike BrainMap [Fox XXX] or NeuroSynth [Yarkoni XXX], the images do not come with explicit labels describing their content. We use a simple heuristic to assign labels to images: for each image, we look at its meta-data: its name, description, description of the contrast, if available. We assign a label to an image if this meta-data contains a term associated with, where the labels and associated terms are given in table 1.
| Label | Associated terms |
|---|---|
| language |
|
| audio |
|
| motor |
|
| visual |
|
Table 1: labels and associated terms
Figure 4 shows for each label the mean map of all the entries of the database that contains the corresponding terms. We can see that this very rough meta-analysis does capture some meaningful information. Indeed, the average activation related to "auditory" terms highlights very well the auditory cortex. Similarly the "visual" terms light up the occipital cortex. However, all these average maps contain non-specific regions that are not directly related to the terms probed. First, they contain attentional and executive networks, in particular the dorsal ACC and frontal insula. Their presence may be explained from the fact that such functions are recruited in every task performed in the scanner. Second, they contain regions that are only indirectly linked to the label. The language terms highlight the visual word-form area [McCandliss, the visual word form area: expertise for reading in the fusiform gyrus], in the left fusiform cortex. On the map related to motor terms, the motor cortex appears washed out, possibly because it not the focus of the corresponding experiments, but solely appears during the subject's response.
In cognitive neuroimaging, the conventional approach to disentangle various cognitive effects is to craft specific contrasts, revealing the effect of interest in the experiment. Accumulation of data via an image store such as NeuroVault opens the door to additional statistical power, highlighting new phenomena, such as the default mode network. However, as seen from the term-based analysis, crafting contrasts across experiments is challenging without an exact knowledge of not only the cognitive questions probed by the experiments, but also contingent experiment details such as how the stimuli were presented or what we the subject's response. One possible approach to exploit the richness of the database is to rely on data-driven method to unmix different cognitive components. In figure 4, we show the results of an Independent Component Analysis (ICA) on the NeuroVault database. We extract 20 co-activation networks. We use the following strategy to retrieve the cognitive content that these networks capture. First we use the decoding functionality of NeuroSynth [Yarkoni XXX] to associate with each statistical map in Neurovault cognitive loadings: weights for each term in NeuroSynth. The unmixing matrix estimated by ICA is applied to go from these map-level loadings to loadings for each network. On figure 4, we represent with each ICA network alongside with the four terms most heavily loaded.
This analysis is similar to that conducted in [Smith 2011, Correspondence of the brain's functional architecture during activation and rest] on the brainmap database, however [Smith 2011] rely on manually labeled and curated data whereas here we use fully automated extraction of information.
Concluding remarks
We have presented a first analysis of a few hundred brain images accumulated with little-to-none curation across multiple labs and multiple experiments. Analysis of such heterogeneous data has to overcome new challenges, such automatically discovering the cognitive content of the brain maps or teasing out the multiple psychological effects that are intertwined in the database. Our analysis leads to promising preliminary results. We have automatically extracted functional networks, with brain maps and associated cognitive concepts, from a relatively small number of images (BrainMap and NeuroSynth cover respectively 2500 and 6000 papers). While these networks and their labeling are not free of noise, they give a rich overall vision of brain function and it's neural support. We are confident that increased amount of data will lead to discovering new organizational principles of brain function. Uploading the data is convenient, to sharing results with other researchers, and easy, as it does not impose manual tagging or curation. The analysis is fully automated and can scale easily with limited human resources.
Note that doing a principled statistical inference, eg computing an p-value or a posterior from this heterogeneous collection of maps require methodological developments outside of the scope of this article.↩