Large-scale functional neural network correlates of response inhibition: an fMRI meta-analysis

Paradigms such as the stroop, Flanker, Simon, stimulus response compatibility (SRC), antisaccade tasks, stop signal, and Go/NoGo tasks, which require a response to a target stimulus among irrelevant/distracting stimuli, are commonly regarded as paradigms that involve inhibitory action control (Nigg 2000) and are commonly used for the investigation of response inhibition mechanisms (Nee et al. 2007; van Velzen et al. 2014; Swick et al. 2011; Stahl et al. 2014; Nigg 2000). In stop signal and Go/NoGo tasks, increased automatic tendency to initiate a particular motor response is induced through a higher frequency of go trials than inhibition trials (i.e., NoGo or stop). The resultant action bias has to be suppressed when the inhibition signal is presented during NoGo or stop trials. In the Go/NoGo task, participants have to withhold a prepotent but not yet initiated motor response, whereas the stop signal task cancels an already initiated motor response. In the other tasks, which can be organized under the term “incongruency tasks” (Cieslik et al. 2015), a given stimulus dimension interferes with relevant stimuli and/or response information, thereby affecting responses to the relevant information.

These tasks all require cognitive control over a predominant response tendency and the context-dependent initiation of an appropriate behavioral alternative—that is, either to initiate an alternative, nondominant response, or to not respond at all. Given the similarities and differences across these tasks, it is important to explore whether their neural correlates are shared, distinct, or both. To achieve this goal, we categorized these paradigms into three subcategories: inference resolution, action withholding, and action cancellation (Sebastian et al. 2013b; Stahl et al. 2014; Friedman and Miyake 2004; Hasher et al. 2007). Interference resolution is the process of selecting information with regard to its relevance to an ongoing task and suppressing the processing of irrelevant information (Yarkoni et al. 2010). This process is usually tested by the Stroop, Simon, Flanker, SRC, and antisaccade tasks (Stahl et al. 2014; van Velzen et al. 2014). Comparatively, action withholding is commonly assessed using the Go/NoGo task. The stop signal task is conceptualized as an action cancellation task.

Over the past 15 years, there has been a dramatic increase in the number of studies examining neural correlates of response inhibition using functional magnetic resonance imaging (fMRI) (refer to Fig. 1). Based on the results of these studies, several different models of inhibitory control have been put forward (e.g., Aron et al. 2004, 2014; Swick and Chatham 2014; Van Belle et al. 2014; Zandbelt et al. 2013; Hampshire et al. 2010, 2015). Among these models, those proposed by Aron and Hampshire are arguably the most prominent and influential. Aron and colleagues proposed that response inhibition activated a module mainly in the right lateralized frontal brain areas, in which the inferior frontal gyrus (IFG) was the key component. During intact response inhibition, inhibitory neural stop signals may then be sent from these frontal regions to the motor cortices through cortico-striatal–thalamic–cortical projections (Aron 2011; Chambers et al. 2009). Hampshire’s group, on the other hand, suggested that response inhibition was one example of a broader class of control processes. These control processes are supported by the same set of fronto-parietal networks that exert control by modulating local lateral inhibition processes, which occur ubiquitously throughout the cortex (Hampshire et al. 2010). Thus, instead of focusing on how a specific brain region and its connection pathways may support response inhibition, understanding the neural basis of behavioral control may require a more holistic approach that considers how common network mechanisms support diverse cognitive processes (Hampshire and Sharp 2015).

The previous studies have indicated common regions activated by response inhibition tasks, including those correlates comprising the fronto-parietal network, specifically the IFG, pre-SMA, and parietal regions. For example, Sebastian et al. (2013b), using hybrid tasks combining Simon, Go/NoGo, and stop signals, demonstrated that areas such as the IFG and the bilateral parietal regions showed activation across all tasks. Nonetheless, unique neural networks for the three subcategories of processes underpinning response inhibition have also been suggested (Rubia et al. 2001; Sebastian et al. 2013b; Chevrier et al. 2007; van Velzen et al. 2014; Sebastian et al. 2013a). For example, the Simon task (interference resolution) was found to stimulate the pre-motor and parietal regions to a greater extent than action cancellation (stop signals), whereas action cancellation was found to elicit stronger activation in the bilateral posterior inferior frontal gyrus/insula and the right striatum than action withholding (Go/NoGo) (Sebastian et al. 2013b). Although the above findings provide important preliminary evidence of the potential common and distinct neural underpinnings of different response inhibition tasks, the majority of studies have either explored only a single response inhibition task in small samples with limited statistical power or were challenging to compare, because they employed various different stimuli (e.g., visual, auditory). We are, therefore, hindered from drawing complete conclusions on the common and distinct neural correlates of the subcategory of processes for response inhibition (Swick et al. 2011; Logan et al. 2015). An understanding of the common and unique neural correlates of the subcategories of cognitive processes associated with response inhibition provides critical insight into early detection, diagnostic accuracy, and treatment targets of clinical disorders afflicted by dysfunctional response inhibition (Aron 2011; Chambers et al. 2009). Meta-analytic pooling of all related studies is a powerful statistical tool that combines data sets from a collection of similar studies to obtain a more accurate and robust estimate of the effect size of a given phenomenon (Fox et al. 2014), and thus may serve as a promising approach to studying the common and distinct neural correlates of the three subcategories of processes associated with response inhibition.

Furthermore, investigation of network correlated activations may provide system-level interpretations as to what neural mechanisms underlie response inhibition (Damoiseaux et al. 2006; Smith et al. 2009; Erika-Florence et al. 2014). Indeed, studies have suggested that response inhibition is correlated with the connections between the cingulo-opercular network (salience network) and fronto-parietal network, and that default mode network activity might be actively downregulated by an inhibitory module within the fronto-parietal network (Leech et al. 2011). Aberrant network interactions are proposed to lead to deficits in response inhibition (Hampshire and Sharp 2015; Jilka et al. 2014). A large-scale network approach will help to clarify these speculations and provide system-level understanding about the working principle of response inhibition (Leech and Sharp 2014). Given that response inhibition is subserved by a large-scale distributed system that includes the bilateral cortical and subcortical regions (Swick et al. 2011), conceiving of a large-scale network rather than the classical region-based functional-anatomy analysis will be more suitable for this purpose (Woo et al. 2014; Hampshire and Sharp 2015).

In summary, the overall objective of this study is to provide a quantitative summary of the major findings across studies on response inhibition, on the basis of previously reported neural network parcellations (Yeo et al. 2011; Choi et al. 2012). Specifically, we aimed to (1) examine activation patterns across all studies using multilevel kernel density analysis (MKDA); (2) analyze the brain functional network correlates responsible for interference resolution, action withholding, and action cancellation; and (3) identify the common and primary neural network correlates for each subcategory through conjunction and subtraction analyses, respectively.

Two online citation indexing services—PubMed and Web of Science—were searched. Keywords including “fMRI” with “response inhibition”, “inhibitory control”, “interference resolution”, “action withholding”, “action cancellation”, “stop signal”, “stopping”, “go nogo”, “action restraint”, or “countermanding” were used to retrieve relevant literature published prior to Dec. 31, 2015 (for detailed search results, please see Table S1 of the Supplementary Materials). We also searched the BrainMap database using Sleuth (http://​brainmap.​org/​) within the imaging modality of “fMRI” and the behavioral domain of “action inhibition” and obtained 106 papers. All articles were pooled into a database, and redundant entries were eliminated, yielding 4092 reports. We applied the following exclusion criteria to eliminate articles that were not directly relevant to this study: (1) nonoriginal studies (e.g., review articles), (2) studies that did not report results in standard stereotactic coordinate space (either Talairach or the Montreal Neurological Institute, MNI), (3) studies that were purely based on region of interest (ROI) analysis (e.g., using anatomical masks or coordinates from other studies), (4) analyses applying methods other than nonlinear modeling, e.g., multi-variate pattern analysis (MVPA), (5) studies with sample size below five and/or age range of participants outside 18–65 years, (6) studies on atypical populations whose brain functions may have deviated from those of healthy adults and those in which results for healthy controls were not reported separately, (7) pharmacological or training-related studies, and (8) single-sex studies. A total of 225 full articles were included in the current meta-analysis (Demographic data see Table S2 of Supplementary Materials). The detailed searching and selection procedures are shown in Fig. 2.

We extracted the following information from each study: authors, year of publication, sample size, experimental design, paradigms and task contrasts (including the frequency of the presented inhibitory stimuli), field strength of the MRI scanner, and cluster coordinates in the MNI or Talairach space (Table S2 in Supplementary Materials). As the MNI/Talairach coordinate bias associated with reference frame (position and orientation) and scale (brain size) can be substantially reduced using the best-fit tal2icbm transform (Lancaster et al. 2007), coordinates that were not reported in the MNI space were transformed using the Lancaster transformation.

We conducted meta-analyses using the MKDA (Wager et al. 2007) toolbox (http://​wagerlab.​colorado.​edu) to identify brain regions activated during response inhibition. Peak effect coordinates from each study were convolved with a spherical kernel (r = 5 mm) and threshold to obtain an indicator map, with a value of one indicating a significant effect in the neighborhood and a value of zero indicating no significant effect. The density of the effect was computed by averaging the indicator maps weighted by the study sample size, and the resulting density maps showed the proportion of studies in which activation was observed within 5 mm of each voxel. The family wise error (FWE) rate was estimated with a Monte Carlo simulation to correct for multiple comparisons, and a natural null hypothesis was that the “activated” regions are randomly distributed throughout the brain. Thus, the reported meta-analytic results in this study represent consistently activated regions across studies: regions in which significant activations were observed in the local neighborhood by more studies than would be expected by chance (p < 0.05, FWE corrected across the entire brain). The MKDA was performed for characterizing brain activation patterns. First, we identified those brain regions that showed significant convergence across 225 studies comprising 3453 foci from 323 contrasts. Then, three additional MKDA analyses were conducted for the specific activations produced by the three subcategories: interference resolution, action withholding, and action cancellation. For interference resolution, 817 foci from 68 contrasts were included. The analyses for action withholding consisted of 1932 foci from 147 contrasts, and the analyses for action cancellation included 1523 foci from 108 contrasts. The same statistical analyses and thresholding approaches were applied for all meta-analyses in this study. Three subtraction analyses were conducted to capture the selectively or preferentially activated brain regions for the different classifications of responses: interference resolution vs. action withholding, interference resolution vs. action cancellation, and action withholding vs. action cancellation.

To examine the related neural network correlates, voxels significantly activated by response inhibition and its sub-processes were overlaid onto seven commonly referenced brain functional network correlates covering the cerebral cortex and striatum (Yeo et al. 2011; Choi et al. 2012): the fronto-parietal network (FPN), dorsal attention network (DAN), ventral attention network (VAN), somatomotor network (SMN), visual network (VN), affective network (AFN), and default mode network (DMN). Chi-square tests were performed to contrast the proportions of the activated voxels in the seven network correlates distributed among interference resolution, action withholding, and action cancellation.

Action withholding (147 experiments of Go/NoGo tasks) and action cancellation (108 experiments of stop signal tasks) included approximately two times the number of experiments as interference resolution (68 experiments). Random selections of 68 contrasts from action withholding and action cancellation were performed to test the effect of the experiment numbers. We repeated the MKDA analysis using the same settings, and Chi-square analyses were performed to compare the activated voxels in the seven network correlates to those of our main analyses and the random selected studies.

The all-inclusive analysis of the 225 studies showed significant activations of three large clusters in the right hemisphere: the frontal cortex, the angular gyrus, and the supplementary motor area (Fig. 3a and Table S3 in Supplementary Materials). More precisely, cluster activations in the frontal cortex included the insula, inferior frontal gyrus, middle frontal gyrus, and superior frontal gyrus. These activations extended to the subcortical regions (the thalamus and pallidum). Furthermore, activation of the angular gyrus cluster extended to the superior temporal gyrus, while activation of the supplementary motor area extended to the middle frontal gyrus. In addition, several areas in the left hemisphere, including the insula, putamen, and middle frontal gyrus were consistently activated across all response inhibition tasks. By overlapping these clusters with the seven functional networks, we discovered that the activated areas were primarily distributed in the fronto-parietal network (36%), ventral attention network (27%), dorsal attention network (18%), and default mode network (13%) (Table S9 in Supplementary Materials) (Fig. 4a).

Activation patterns common to the three subcategories of processes of response inhibition were derived by overlapping their MKDA maps. The common regions included the right inferior frontal gyrus, which extended to the insula, right median cingulate, and paracingulate gyri, and the right superior parietal gyrus (Fig. 5). The network analysis showed that the common activated areas were mostly distributed in the ventral attention network (61%), fronto-parietal network (26%), default mode network (8%), and dorsal attention network (4%).

Chi-square analyses revealed that the network distributions of interference resolution versus action withholding/action cancellation were significantly different (interference resolution vs. action withholding, χ ² = 19.93, df = 6, p = 0.003; interference resolution vs. action cancellation, χ ² = 22.37, df = 6, p = 0.001). Furthermore, we observed that interference resolution, which was related to greater activation in the ventral and dorsal attention networks (Fig. 6a, b), involved the left supplementary area, the left precentral gyrus, and the left superior parietal gyrus. On the other hand, action withholding engaged the fronto-parietal network, including regions of the right middle frontal gyrus, bilateral insula, the right triangular part of the inferior frontal gyrus, and subcortical areas such as the left putamen and right pallidum (Table 4; Fig. 6c, d).

Action cancellation compared with action withholding revealed a significant activation in the ventral attention and fronto-parietal networks, in which the primary correlates were the right inferior frontal gyrus, the right supplementary motor area, and the right superior parietal gyrus (Fig. 6f). Brain areas in bilateral insula and the right caudate showed greater activation during action withholding than action cancellation (Fig. 6e).

The results obtained in this study were proven replicable under different validation schemes. (1) Using the leave-one-out cross-validation procedure, we tested the effects of excluding paradigms on interference resolution and found that the neural networks of interference resolution were primarily distributed in the ventral attention network and the dorsal attention network regardless of which paradigm was excluded (Table 5: Table S9 and Fig. S1 in Supplementary). (2) When examining the NoGo vs. Go contrast in relation to action withholding, activation of the correlates of the ventral attention network and the fronto-parietal network was observed (detailed brain areas are listed in Table S7 in Supplementary Materials). (3) With regard to the contrast condition of the stop signal tasks, we found no significant differences in the distribution of neural networks, including all studies and Stop versus Go contrasts (Table 4 and Table S8 in Supplementary Materials). (4) The evaluation of the number of experiments when contrasting the three subcategories indicated no significant differences among the real contrasts and the randomly selected 68 contrasts for action withholding and action cancellation (Table 4). The activated brain areas and network distributions are displayed in Tables S4, S5, and S9 in the Supplementary Materials.

The fronto-parietal network is characterized as a set of cortical areas that are mutually activated when performing a wide variety of cognitively demanding tasks (Fedorenko et al. 2013; Duncan 2010). In this study, we found that across all 225 studies with 323 experiments, areas in the dorsal lateral frontal cortex, pre-supplementary motor area, and temporal parietal junction were consistently activated during all response inhibition tasks (Fig. 3a), i.e., activation was mainly distributed in the fronto-parietal network. Notably, similar results were also detected when input contrast counts were taken into account (Tables S5, S9), which demonstrated that the neural network involved in response inhibition was stable and robust. These findings are in line with the findings of previous studies on the neural correlates of working memory, sustained attention, and reasoning. Nee et al. (2013) reported that there was widespread bilateral fronto-parietal network activation during various types of working memory tasks. Similarly, consistent involvement in a very similar network was also observed in a study of the neural correlates of sustained attention (Langner and Eickhoff 2013). By synthesizing semantic and visuospatial analogy tasks, Hobeika et al. (2016) observed that there were domain-oriented regions in the inferior and middle frontal gyri. Thus, domain-oriented regions within the fronto-parietal network are widely involved in cognitive control components including response inhibition, working memory, sustained attention, and reasoning (Miyake et al. 2000; Chan et al. 2008).

By overlapping the neural activity pattern of each category of response inhibition tasks, three main clusters of activation were observed at the (1) right IFG extending to insula, (2) right median cingulate and paracingulate gyri, and (3) right superior parietal gyrus (Fig. 5). The right IFG, besides detecting changes in stimulus features (Sharp et al. 2010; Dodds et al. 2010), facilitates infrequent action-related events by activating nondominant but relevant responses while inhibiting automatic but irrelevant actions at the same time. Along this line of thought, lesion studies have also indicated that stop signal task performance worsens as the size of an inferior frontal gyrus lesion increases (Aron et al. 2003), which supports an inhibitory role of the inferior frontal gyrus in resolving conflicts during response execution. The right anterior insula has been proposed to represent a hub that controls brain activity across different tasks and stimulus modalities to initiate and adjust cognitive control mechanisms (Cai et al. 2014). The previous research has demonstrated a linear relationship between the neural activity of the anterior insula and task performance across three response inhibition tasks (Flanker, Go/NoGo etc.) (Wager et al. 2005), which were included in the current meta-analysis. With respect to the right median cingulate and paracingulate gyri (MCG), studies have revealed that the MCG is the key region for proactive rather than reactive action control, indicated by increased neural activity for endogenous action selection (Aron 2011). Notably, higher MCG activity was observed when subjects were required to perform a dual task, e.g., deciding which hand will perform the action and when to give the response (Hoffstaedter et al. 2013, 2014). Moreover, the MCG has also been implicated in performance monitoring via conflict detection in information processing, reallocation of attentional resources according to task-relevant information, and corresponding action formation (Badzakova-Trajkov et al. 2009). Friedman and Miyake 2004 indicated that subjects might experience several different mental processes during response inhibition tasks: maintaining the stimulus, which requires appropriate response in working memory; detecting conflicts for stimuli inconsistent with the goal; overcoming preponderant tendencies; and choosing correct responses.

Based on the previous findings and the current results on the common areas involved in different response inhibition tasks, we argue that when performing these tasks, individuals need to maintain task requirements across trials and goal-corresponding action inhibition through engagement of the right IFG. In contrast, the right insula is required to overcome preponderant tendencies, notice the salient stimuli, and coordinate various control mechanisms. Finally, the right MCG is engaged in monitoring conflicts so as to promote task-relevant actions. Overall, we propose that the right inferior frontal gyrus, insula, and MCG may comprise the core neural network of the supervisory attentional control system needed to implement a nondominant, context-dependent behavior against a competing behavioral alternative (Alexander and Brown 2010; Cieslik et al. 2015).

Our findings show that distinct neural correlates and hence networks were indicated for each of the three subcategories of processes: inference resolution, action withholding, and action cancellation. Interference resolution was found to draw on the ventral attention network, while fronto-parietal network was implicated in action withholding/cancellation.

In this meta-analysis, the experimental paradigms classified as evoking interference resolution were those that required conflict resolution and inhibition of response tendencies for successful responding (Nee et al. 2007). These paradigms induce an automatic attention reorientation and response preparation in the direction of the dominant but task-irrelevant stimuli. Participants then need to actively reorient attention to the nondominant spatial location to initiate an adequate response. This process may require enhanced consideration of the objective and engagement of resources for conflict management. Thus, inference resolution, relative to action inhibition, may draw upon significant coherent activation in the ventral attention network, especially the left pre-supplementary motor area (pre-SMA) and the left superior parietal gyrus. Pre-SMA is associated with monitoring and selecting appropriate motor response output (Nachev et al. 2008; Iannaccone et al. 2015) and the superior parietal gyrus has an essential role in re-directing attention by promoting attention allocation to nonspatial properties of stimuli (Mevorach et al. 2009; Wang et al. 2015). Overall, stronger ventral attention network activation during interference resolution compared to action inhibition may indicate that the former is to a greater extent dependent on response selection processes modulated by goals and conflicts (Nee et al. 2007).

Action withholding encompasses future action selection and inhibition, whereas action cancellation goes beyond this to demand inhibition of an ongoing response. This is induced by presenting Go and NoGo signals at the same time point in their respective trials for withholding or presenting stop signals with a delay after a Go signal for cancellation of an already initiated response. The inhibitory load is likely to be higher in cancellation than withholding (Schachar et al. 2007). Studies have suggested that the time difference in presenting the NoGo and stop signals induces distinct activation patterns (Swick et al. 2011; Rubia et al. 2001), such as a greater extent of activation in the right inferior and superior frontal gyri. This notion is partially supported by the results of our subtraction analysis between action withholding and action cancellation (Table 4). The activation differences observed between these two processes suggest that action withholding and cancellation may interact at different times within the action generation or action inhibition process, thus jointly influencing motor response inhibition (Sebastian et al. 2013a; Dambacher et al. 2014; Cieslik et al. 2015).

Further supporting the notion of time sequence difference between action withholding and action cancellation, Sebastian et al. (2013b) used a hybrid response inhibition task to study functional and spatial segregation and the specialization of underlying neural sub-processes of response inhibition and found that neural activity levels in the fronto-parietal network follow a quantitative progression: action cancellation > action withholding > interference resolution. Three stages of general information processing exist: stimulus identification, response selection, and response execution, or the motor stage. The neural resource required increased gradually when progressing through these three stages of general information processing (Schank 2014; Marois and Ivanoff 2005). Herein, the progressive increase of neural activations by interference resolution, action withholding, and action cancellation may imply that response inhibition may contain subcomponents that interact in a sequential fashion; action withholding is an intermediate process within the sequence of interference resolution, action withholding, and action cancellation (Sebastian et al. 2013a, b).

Notably, action inhibition compared to interference resolution also engages neural networks in the striatum areas, such as the putamen and pallidum. The role of subcortical areas in response inhibition has extensively been discussed (see reviews: Aron 2011; Deffains et al. 2016; Aron et al. 2016). Lesion studies also suggest that lesions of the medial striatum in rodents lead to overall longer stop signal reaction time (Eagle and Robbins 2003). In line with these animal findings, patients with damaged basal ganglia are slower to stop their responses than controls (Rieger et al. 2003). Furthermore, Go/NoGo tasks also elicit striatal activation. Many functional and structural MRI studies have pointed to a fronto-striatal “circuit” underlying response inhibition in the Go/NoGo paradigm (Durston et al. 2003; Wessa et al. 2007). In a neurophysiological experiment, striatal activity was recorded, while monkeys were performing a Go/NoGo task. The authors found that the striatum could be important for preparing to stop a response (under working memory) and then implementing inhibitory control over a sustained period (Apicella et al. 1992). Thus, the neural activity pattern in the striatum plus the frontal areas comprises a cortical circuit related to the control of response inhibition, specifically response execution (Aron 2011).

We acknowledge that our meta-analysis is subject to limitations. The first one is related to the bias of synthesizing different tasks into the components of interference resolution. Our meta-analysis included studies adopting a mixture of inhibitory control tasks, such as the Stroop, Simon, and Flanker tasks, with the consequence of increased heterogeneity of study paradigm and designs. We adopted the leave-one-out cross-validation method to test homogeneity and determined that all brain activation areas showed highly replicable network distribution profiles (Table S9; Figure S1 in Supplementary Materials). Nevertheless, future meta-analyses including additional studies with common stimulus–response incompatibility tasks are needed. The second limitation is related to the use of the Go/NoGo task for action withholding. Go/NoGo can be classified as a simple task (the NoGo stimulus was always the same) or a complex task (the NoGo stimulus changed depending on context) that may require more frequent updating of stimulus–response association in working memory (Simmonds et al. 2008). As such, simple and complex Go/NoGo tasks may not be perfectly matched across the included studies. In line with this, it has been suggested that the activity pattern in the fronto-parietal network may be derived from the high demand placed on attentional or working memory resources (Simmonds et al. 2008; Criaud and Boulinguez 2013). However, such suggestions were derived from meta-analyses that included less than 30 experiments and thus may not be sufficiently robust (Eickhoff et al. 2016b). We compared high (simple) and low (complex) frequencies of NoGo stimulus (i.e., equal to 50% and less than 50%, respectively) and found that the network correlate distributions of activated areas did not differ significantly (Tables S6, S7, respectively, in the Supplementary Materials). Third, the use of the data generated from stop signal tasks to assess action cancellation drew more on proactive than reactive control for inhibiting inappropriate responses (Aron 2011; Vink et al. 2014; Van Belle et al. 2014; Vink et al. 2015). To control for this confounding effect, we have conducted a supplementary Chi-square test. The result of this test demonstrated that the network correlate distributions between all contrasts and those contrasts that included only the Stop–Go tasks did not differ significantly (detailed condition comparisons of each stop signal task and the results are presented in Tables S2 and S9 in the Supplementary Materials). Nonetheless, despite the robust findings in this study, the heterogeneity of the action cancellation tasks should be a point of concern for future research.