Network subregions and correspondence with prior network definitions
The lateral frontal clusters of co-activation with the IFJ are formed by a set of confluent separate frontal cortical areas that have been well-characterized in the literature: A group of co-activation peaks rostral to the IFJ, predominantly assigned to the middle fontral gyrus in this analysis highly corresponds to an earlier definition of the mid-dorsolateral prefrontal cortex (mid-DLPFC) as discussed as constituent of a network subserving multiple cognitive demands  and other common definitions of dorsolateral prefrontal cortex (DLPFC) in stereotaxic space, e.g. [38, 39]. Another consistent finding was a common co-activation of the IFJ with a set of peak activations in the mid-ventrolateral prefrontal cortex (mid-VLPFC) extending to the anterior insula , partially overlapping with other prior definitions of mid-VLPFC and anterior VLPFC . Co-activations in BA 44 resemble the location of Broca’s area in terms of a coordinate-based definition .
Observed peak coordinates comprise a range of precentral co-activations. One of these peaks is located close to the frontal eye field. However it better corresponds to an adjacent region that has previously been associated with visuomotor hand conditional activity . Thus in summary, the precentral peaks observed are presumably to some extent more directly associated with task execution than the other observed co-activations, as a range of studies included in this meta-analysis adopted task vs. baseline contrasts. The oberved association in our rs-fc analysis (Table 3) point even to a direct association in terms of neuronal activity.
In addition to these lateral frontal areas the network observed in this meta-analysis comprises medial frontal, posterior parietal and inferior posterior cortical areas (Tables 1 and 2, Figure 1). In this respect results are in line with findings based on other approaches to the study of functional brain networks that did however not directly focus on the IFJ: Similar activations of the medial wall have been observed by Duncan et. al. in their analyses focusing on frontal networks subserving multiple cognitive demands . Similarly, mainly fronto-parietal networks are common findings in functional connectivity analyses based on resting-state fMRI acquisition: A set of lateral frontal (including precentral), insular, medial frontal, posterior parietal and inferior posterior foci form the ‘task-positive network’ in a study by Fox et al. . A similar network has been observed with an exploratory approach based on independent component analyses as well in resting-state fMRI data as in data derived from the BrainMap database analyzed without constraints regarding functional areas or specific tasks .
Our results are highly concordant with the involvement of the IFJ in the cognitive control network (CCN) proposed by Cole et. al. . All its components were found to be significantly co-activated with the IFJ: the DLPFC, pre-SMA, anterior insular cortex and PPC as well as matching aspects of the dorsal premotor cortex as far as coordinates are concerned. In contrast, in that study activations in Broca’s area were not tightly coupled with the CCN. Thus, co-activations of the IFJ with Broca’s area in our analysis could be interpreted as an evidence of a relation to additional language processing demands in the tasks included without representing a direct functional connection. Yet rs-fc results support a more direct links in terms of coherent neuronal activity (Table 3).
A recent task-based meta-analysis of cognitive control identified a comparable fronto-parietal CCN including the IFJ. Yet it was labeled as part of the inferior frontal gyrus (IFG) based on its Talairach coordinates (-42, 4 ,30 and 44, 6 32) in that case. The thalamus was also identified in an overall analysis across task-domains, yet it did not survive a formal conjunction analysis of different sub-domains of the construct of cognitive control .
Moreover, as a rather consistent finding we observed parallel activations of the IFJ with the basal ganglia (mainly the lentiform nucleus) and the thalamus. Regarding thalamic activation there might be a pitfall in ALE analyses related to the more spherical structure of its nuclei compared to rather flat cortical areas: Thalamic activations arising from different nuclei may be concatenated to a single cluster or even lead to a common peak location near midline. This is especially problematic as the activations observed here are finally assigned to the medial-dorsal nucleus (MDN), indeed a near-midline structure. However, thalamic co-activations form rather separate sub-clusters in both thalamic hemispheres (see Figure 2). In addition it is the MDN that has in previous studies been closely associated with the prefrontal cortex and higher cognitive functions in contrast to more lateral thalamic nuclei: Connections of the MDN with prefrontal brain areas have been observed using diffusion-MRI based tractography in humans including the DLPFC [44, 45] and primary fMRI functional connectivity analyses . This notion is also supported by animal studies . This applies to the thalamic peaks observed here as well: After conversion of the coordinates into MNI space using the corresponding tool provided in GingerALE [28, 48], the left thalamic peak exhibited a probability of 0.87 and the right thalamic peak of 0.79 to be connected with the pre-frontal cortex (without differentiation of subdivisions) according to a probabilistic human tractography atlas based on diffusion-MRI [49–51]. The probability of direct structural connectivity to the posterior parietal cortex was however nearly non-existent, reflecting the fact that diffusion tractography only detects direct fiber connections and not polysynaptical connectivity in functional networks. Human lesion data suggests, that executive dysfunction may arise from combined lesioning of several thalamic structures including the MDN .
Comparison of MACM and resting-state results
Meta-analytic results were mostly confirmed using an analysis of intrinsic BOLD signal fluctuations in a presumably independent, publicly available dataset. However, there were some distinct differences: The VLPFC was less clearly identifiable in the resting-state analysis. It was present in the correlation maps but it was not marked as a distinct local maximum. The definition of local maxima in the rs-fc analysis was however constrained by a distance criterion (8 mm). As the VLPFC is wedged in Broca’s area in the left hemisphere and the anterior insula as well as the DLPFC in both hemispheres it might have been missed for that reason.
The ACC has been extensively studied as a region crucial for cognitive control processes . Therefore its identification in the resting-state analysis is in line with previous findings. More inferior parts of the cerebellum identified in the rs-fc analysis might have been missed in the MACM analysis because this inferior region is not usually covered in many functional neuroimaging studies. Superior and middle temporal locations were only identified quite inconsistently when comparing different analysis strategies and can therefore not be considered a verified finding in this study.
Finally there were some pre- and postcentral areas of significant functional connectivity in the analysis of resting-state data that were not observed in the MACM analysis.
In contrast to prior findings in resting-state fMRI analyses based on spatial independent component analyses (ICA)  the network observed here appears more interhemispherically connected and additionally overlaps with a fronto-insular component. This may be related to a possible advantage of the meta-analytic connectivity modeling approach adopted here: classical definitions of functional connectivity are based on the analysis of a tight temporal coupling of neurophysiological events . In contrast, functional connectivity in terms of MACM can be interpreted as remote brain areas cooperating in dealing with a task without necessarily exhibiting highly temporally correlated activity. Thus if two rather independent networks in terms of direct structural connectivity or classical functional connectivity are parallel recruited due to comparable task demands, these networks can potentially be identified as one coherent network by MACM . This clear differentiation is however limited by our seed-based rs-fc analysis: Though exhibiting a certain degree of asymmetry, resting-state networks were not limited to the seed’s hemisphere. The main difference might thus arise from different analysis strategies of rs-fc analyses (with spatial ICA emphasizing spatial independence of networks).
There are different possible explanations for the fact that more regions were connected to the IFJ in the analysis of the resting-state dataset compared to the MACM results: Statistical power of both approaches is most likely different. In addition to some baseline-comparisons the BrainMap database contains coordinates from many well-controlled fMRI contrasts to delineate specific behavioral processes by including associated functions (e.g. stimulus-perception and motor responses) in control conditions. The additional correlations in the resting-state data may therefore also represent meaningful and necessary connections of the actual CCN components to brain areas relevant for direct interaction with the environment.
Though the results of both methods are comparable the occasional differences of resting-state and MACM results point to the critical fact that current converging methods applied in the study of complex brain networks may oversimplify the actual functional organization of the human brain as they may not optimally account for the internal organization of such networks and their complex interdependences. It is a notable finding in this context that analyzing fMRI data with an increased temporal resolution using temporal ICA Smith et al. recently reported temporally-independent functional modes of spontaneous brain activity that overlapped with each other and networks known from conventional (seed-correlation or spatial ICA) analyses .
As stated in the introduction the IFJ has been studied as a specific brain area in task-based fMRI and meta-analyses limited to a few task domains. Results can be summarized as an involvement of the IFJ in three main component processes of cognitive control (task switching, inhibitory control and working memory) [1–8].
The functional significance of similar fronto-parietal networks as observed in this study has explicitly or implicitly been assessed in numerous often highly specific task-based studies. As reported above, a recent meta-analysis has accumulated such findings based on the BrainMap database : In that meta-analysis cognitive control was operationalized as initiation, inhibition, working memory, flexibility, planning and vigilance. Therefore for each of these sub-domains a set of established tasks (like Flanker, Go/No-Go, Antisaccade, Simon and Stroop tasks for inhibition) was defined and the studies included in the final analyses were restricted to those using these a priori defined tasks. A core network was observed in a conjunction analysis of flexibility, inhibition and working memory that highly overlapped with the CCN observed in the MACM analysis reported here. Thus results point to an involvement of the CCN in all of these functions in a rather unspecific way and to a connection of the IFJ with these other regions in this context.
In contrast to these previous meta-analytic approaches providing information about the IFJ  or the CCN  the analysis reported in this article is conceptually different (1) in that it is not a priori limited to the context of cognitive control and (2) in that it starts from the IFJ as a previously defined specific location in the brain and therefore adds specificity to the knowledge about this set of connections. Although our analysis aimed at studying connectivity, this framework can also be used to explore functional meanings of the IFJ and the CCN using an analogous approach: The BrainMap database contains structured information (hierarchical meta-data) about behavioral aspects represented in the reported contrasts. Lancaster et al. have reported an automated behavioral analysis based on these meta-data that allows ROI-based searches . Queries regarding the IFJ and the whole network observed here (Additional file 1: Table S2) give a rough estimation of functions associated with the specific coordinate definitions and network maps in this study. They show a statistically significant association with cognitive processes including (working) memory, inhibition and attention but among others also language processing (left hemisphere) and perceptive processes presumably involved in some of the chosen fMRI paradigms.
There is evidence in the meta-analytic results on CCN functions by Niendam et. al.  that in addition to the rather unspecific involvement of the CCN core regions additional areas are recruited in a sub-domain-specific manner. This finding is compatible with the recent meta-analytic and task-based finding that within the frontal cortex the IFJ is generally involved in the cognitive control subdomain of switching / flexibility but together with other lateral and medial frontal regions which are recruited more specifically [7, 8]. This in turn is in some way reminiscent of the assumption of a hierarchical organization of the rostro-caudal axis of the frontal lobes .
Taken together but potentially limited by the power to detect the involvement of certain brain regions in the different approaches reported the findings seem to support the notion that the IFJ is rather specifically involved in brain systems playing an important role in cognitive control compared to other aspects of brain function.
A potential limitation of the MACM approach may arise from the fact that, unlike for example in resting-state fMRI approaches to functional connectivity, results, though including several different functional domains, may be influenced by the overall distribution of tasks in the BrainMap database and correspondingly the distribution of tasks adopted by the whole functional neuroimaging community. However, as discussed above, our results are in line with recent literature regarding IFJ connectivity and a different approach to MACM  adopted in the NeuroSynth project (http://www.neurosynth.org/) yields qualitatively comparable results regarding IFJ connectivity, thus at least we suppose that there is no specific bias regarding the BrainMap database and studies included.
Recently it has been argued that the IFJ is functionally dissociated from the directly adjacent posterior part of the inferior frontal gyrus (pIFG) . We have not directly observed this dissociation in terms of different peaks in our analysis. The used cuboid-shaped ROIs were based on prior literature regarding IFJ location in stereotaxic space. However, irregularly shaped ROIs might better conform to the IFJ as a functional brain area and help clarify this issue.
The selection of coordinates most exactly representing the IFJ in stereotaxic space is still a matter of debate. For the meta-analysis we aimed at high specificity regarding the IFJ as a distinct functional brain region without accidentally including other functionally defined areas in our seed. We therefore selected a relatively small ROI based on a large body of literature aggregated in a review based on a number of single studies and meta-analytical data . Yet there is recent data that suggests that the IFJ may be located more medial in some subjects . The meta-analysis might thus have missed some studies reporting IFJ peaks just outside the ROI borders. We addressed this issue by using two different ROI definition strategies in the confirmatory analysis based on resting-state functional connectivity. Results were not qualitatively different regarding the identification of the main constituent brain areas of the CCN) when comparing the more lateral and more medial ROI in each hemisphere.
Though the main finding of this study is a considerably specific connection of the IFJ with brain regions previously characterized as a network engaged in cognitive control from a systems neuroscience perspective, it cannot be reliably deduced from these results that the CCN is indeed ‘only’ involved in cognitive control from a classical neuropsychological perspective. This would be some kind of problematic reverse inference . Though the additional analysis of BrainMap meta-data reported shows a significant association of the IFJ as well as the whole network with ‘cognitive’ tasks, a disjunctive analysis of ‘cognitive control’ in a strict sense does not seem to be feasible in this framework, especially due to the fact that the above-mentioned definition of cognitive control partially overlaps with different categories in the BrainMap meta-data. The behavioral analysis in this framework is also limited by the fact that the set of studies which the analysis is based on overlaps with previous function-guided meta-analyses and also includes, as a minority, studies that have originally led to the fundamental assumption that this specific brain location in the inferior frontal junction area is involved in cognitive control.