2.1 Literature Search and Coordinate Extraction

Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocols [21], we searched PubMed for articles using structural MRI to study children and adults with ADHD using the following search terms: (ADHD OR attention-deficit hyperactivity disorder) AND (“magnetic resonance imaging” OR MRI OR “voxel-based morphometry” OR VBM). As of our search date, March 24, 2022, there were 2315 results. We excluded papers that were review articles, that focused solely on specific regions of interest rather than whole-brain analyses, that contained overlapping participant groups (prioritizing the larger sample or, if equal, the more recent paper), or that included participants with comorbidities beyond those common in ADHD (i.e., oppositional defiant disorder, conduct disorder, depression, anxiety, and learning disorders). Medication status was not an exclusion criterion. Inclusion required an ADHD diagnosis based on Diagnostic and Statistical Manual of Mental Disorders (DSM) or International Classification of Diseases (ICD) criteria. Furthermore, included studies reported coordinates of decreased volume in ADHD compared with controls, and not locations of increased volume nor differences in functional MRI (fMRI) activation or functional connectivity differences. Reference lists were reviewed for additional studies, and authors were contacted for coordinates not listed in papers matching our criteria. In total, we identified 38 studies with a total of 40 comparisons. Extracted coordinate sets were converted to Montreal Neurological Institute's 152 brain template (MNI152) space when necessary.

2.2 Anatomic Likelihood Estimate

We then conducted a spatial meta-analysis of the coordinates obtained above using the ALE method, which was first described [22] to assess for convergence across fMRI studies but which has also proven highly useful for structural meta-analyses. We performed an ALE analysis using the GingerALE software [8, 23, 24], as described in Samea et al. 2019 [9]. In short, the ALE method measures the spatial convergence across studies by generating a Gaussian sphere around each reported coordinate and then statistically assessing whether any locations within the overall group are reported more often than by chance. We performed analyses with a cluster-level familywise error rate of p < 0.05 and a cluster-forming voxel-level threshold of p < 0.001. Sample sizes were used to weight each comparison in the analysis in a random-effects model.

2.3 Comparison With Putamen and Default Mode Network

Association maps for the putamen and DMN were obtained from Neurosynth [25] and thresholded based on relative image intensity. We then calculated the overlap between these regions of interest and the 40 maps of locations with decreased volume from ADHD versus control studies to determine how often the putamen and DMN were reported as atrophied in these studies. Separately, we calculated the functional connectivity of the putamen, DMN, and data-driven ADHD network with each of the 40 studies’ patterns of decreased volume using the Adolescent Brain and Cognitive Development (ABCD) study data described below. After conversion to Fisher's z values, we conducted one-sample t tests to determine if correlations between ADHD-related patterns of decreased volume and the putamen, DMN, or our derived ADHD network were significant. Note that whether the 40 studies’ patterns are correlated with our derived network is a circular analysis but was done here to compare against the putamen and DMN values—a comparison that is not circular. Additionally, we conducted a one-way ANOVA with follow-up two-sample t tests to determine how the correlations of the 40 seed maps with each of these three regions of interest compared with one another.

2.4 Coordinate Network Mapping

Next, we built on the ALE analysis with a CNM approach, which considers the connectivity of each study's set of coordinates by leveraging the availability of large-scale functional connectivity data sets of typically developing individuals [26, 27] (Figure 1). These functional connectivity data create the ability to generate normative connectivity maps for a set of brain coordinates, even when those coordinates are taken from clinical populations for which fMRI data are not available [13, 28]. Starting with the same set of coordinates used for the ALE analysis, 40 seed maps were created, modeling each set of coordinates as spheres with a radius of 4 mm. These seed maps were subsequently used to create functional connectivity maps representing normative resting-state connectivity between each sphere location and the rest of the brain. This was accomplished using a pediatric normative connectome, that is, a set of preprocessed resting-state fMRI time courses from 1000 typically developing adolescents from the ABCD study, chosen for low-motion, 1:1 male:female proportions, and all from Siemens scanners, as previously described [28]. Next, we conducted a voxel-wise one-sample t test via permutation testing using the FSL Permutation Analysis of Linear Models (PALM) software with 1000 permutations [29]. The resulting t map was then thresholded with familywise error (FWE) correction (p < 0.05).

2.5 Comparison With Acute Lesions

We conducted further permutation-based two-sample t tests (using PALM software) to compare the results of our ADHD analysis with lesion networks from four independent control data sets. In each data set, MNI coordinates were reported for regions with decreased volume in participants with the disorder(s) of interest than in control participants. The first data set came from a previous ALE meta-analysis of 193 voxel-based morphometry studies reporting decreased volume in psychiatric disorders: schizophrenia, bipolar disorder, depression, substance use disorder, obsessive-compulsive disorder, and anxiety disorders [30]. The second data set was derived from 72 published neuroimaging studies of neurodegenerative disorders (Alzheimer's disease, behavioral variant frontotemporal dementia, corticobasal syndrome, and progressive nonfluent aphasia) [16]. The third data set came from a meta-analysis of 159 studies of neurodegenerative disorders, namely mild cognitive impairment, Alzheimer's disease, Parkinson's disease, and Huntington's disease [31]. The final data set included 135 lesions from patients who were part of the Washington University Stroke Project [32].

2.6 Control Comparison With Random Decreased Volume/Atrophy Studies

To assess the specificity of our results to ADHD, we conducted additional permutation-based two-sample t tests to compare our identified set of coordinates with a separate and random sample (n = 160 studies) of volume decrease coordinates from the BrainMap database, which includes a range of studies on typically and atypically developing populations [33]. Using these new studies, we replicated the CNM approach and constructed a map of commonly affected networks. We computed the spatial correlation between regions of volume decrease in ADHD and in these other studies.

2.7 Control Comparison With Random Coordinates

Finally, using a binary gray matter mask, we randomly generated 236 coordinates split into 40 groups, matching the number of ADHD coordinates. Using these new seed regions, we replicated the CNM approach and constructed a map of commonly affected networks. We then compared this map with the ADHD map using permutation-based two-sample t tests. Finally, we computed spatial correlations between regions of significance from the random analysis and the ADHD analysis.

3.1 Participant Demographics

Our literature search identified 38 studies (Figure 2), including 40 experiments, that directly compared voxel-wise brain volumes, for example, voxel-based morphometry [34], of people with ADHD and typically developing controls. These results represent data from 2952 participants (Table 1). The median age of participants at the time of study was 12.63 years, and the ratio of males to females was 2002:950, consistent with observed M:F ratios of ADHD diagnosis [1]. Due to strict inclusion and exclusion criteria across these studies, few participants were included with comorbidities; however, some participants with comorbidities common to ADHD were included in some studies. These included anxiety (3.5% of the ADHD group and 0.9% of the control group) and oppositional defiant disorder (3.2% of the ADHD group and 0.15% of the control group). Additionally, some participants were medicated, while others were medication naïve. Table 1 contains a detailed breakdown of participant demographics.

3.2 Association With the Default Mode Network and Putamen

To test the hypotheses presented by the existing ADHD literature, we assessed whether the significant coordinates for each of the identified ADHD experiments were near or connected to a priori putamen and DMN maps obtained from Neurosynth [25]. The putamen was chosen as an a priori region because it was the only region implicated in ADHD in the Samea et al. meta-analysis [9], and the DMN was chosen as an a priori network because it has been the most consistently implicated in ADHD [35]. Coordinates from 14 of the 40 experiments (35%) had an overlap of at least 1% of the voxels in the putamen and/or DMN masks. Only five (12.5%) had an overlap of at least 5% with either mask, and only two (5%) had a 20% overlap with the putamen mask (no experiments reached this threshold for the DMN mask). While the regions obtained from the 40 independent experiments had a significant functional connection with the putamen (r = 0.141, t = 4.316, p = 0.0001), this correlation was weak and explained only 1.99% of the variance. There was no significant correlation between these regions and the DMN (r = 0.061, t = 1.513, p = 0.138).

3.3 Data-Driven Meta-Analysis Using Anatomic Likelihood Estimate

We then performed an ALE meta-analysis to identify the spatial convergence of regions reported across experiments. The original coordinates occurred in widespread regions across the brain (Figure 3), and ALE analysis yielded no significant clusters. Visual inspection indicates bilateral patterns of decreased volume across a range of cortical and subcortical structures.

3.4 Data-Driven Meta-Analysis Using Coordinate Network Mapping

Next, we examined the overlap among normative functional connectivity maps from each experiment and found that the regions from 85% of the studies were connected to the right anterior cingulate cortex (ACC) and right thalamus at t > 5 (corresponding FWE-corrected p < 0.05), although these regions were extremely small, comprising four voxels total (Supporting Information S1: Figure 1). Relaxing our inclusion threshold to 80% of studies demonstrating a connection, t > 5 identified connections to 400 voxels in the ACC and several subcortical regions (Supporting Information S1: Table 1). To remove the potential impact of choosing thresholds, we also used normative functional connectivity maps for each experiment to perform a permutation-based one-sample t test of the experiment-level ADHD versus control contrasts, in effect performing a fixed-effects ANOVA of the set of studies [29]. This identified 13 clusters with greater positive connectivity to ADHD patterns compared with typically developing cohorts. After eliminating clusters smaller than 200 cubic mm in volume [18], seven significant clusters remained, with the largest cluster (67704 cubic mm) having peaks in the left and right thalamus, bilateral accumbens, hippocampus, caudate, pallidum, and insular cortex, and to the subcallosal cortex, paracingulate gyrus, ACC, orbitofrontal cortex, cerebellum, and brainstem (Supporting Information S1: Table 1; Figure 4A). There were also connections to locations within the cerebral white matter.

3.5 Comparison to Known Brain Networks and A Priori ADHD Results

Visually, these regions appear to localize to two well-established, previously described networks: the reward network, traditionally consisting of the ventral tegmental area/ventral striatum, ACC, orbital and medial prefrontal cortices, amygdala, and dopaminergic midbrain [36] (Figure 4B); and the cingulo-opercular network, recently renamed as the action mode network [37], comprising the dorsal ACC/medial superior frontal cortex and the anterior insula/frontal operculum [38] (Figure 4C). As further verification for these visual findings, we processed our map through the Neurosynth Decoder [25], which compares any neuroimaging map with the entire database of Neurosynth meta-analyses. This analysis found that the maximum correlation between our CNM map, showing decreased volume in ADHD, and any term in the database was 0.324. The correlation with the term “reward” was 0.288. There is no singular term for the cingulo-opercular/action mode network (AMN) in Neurosynth, but correlations for regions within this network ranged from r = 0.054 (frontal operculum) to 0.239 (anterior cingulate). Purely for the sake of comparison, we also calculated the connectivity of 1) our identified network, 2) the previously mentioned putamen region of interest, and 3) the DMN mask to each of the 40 experimental regions. As expected, an ANOVA of these three groups was significant (F(2,117) = 3.285, p = 0.041) and post hoc two-sample t tests demonstrate that the CNM-derived ADHD network was more strongly correlated to the 40 originating regions than the putamen (t = 1.051) or DMN (t = 2.535). Please note that this is a circular analysis and is included purely for comparison with the results summarized above.

3.6 Control Experiment 1: ADHD Study-Related Networks of Decreased Volume Versus Networks of Decreased Volume in Previously Characterized Clinical Cohorts

3.6.1 Psychiatric Disorders

Since attentional difficulties are common in psychiatric disorders [39], we assessed the specificity of our network by comparing the connectivity patterns from regions of decreased volume in ADHD with regions of decreased volume in a range of psychiatric disorders (n = 212 experiments across 193 studies). Disorders included in the control group were schizophrenia (including schizoaffective and schizophreniform disorders), bipolar disorder, major depressive disorder, substance use disorder, obsessive-compulsive disorder, and several anxiety disorders [30]. However, the resulting two-sample statistical map was both nonsignificant and strongly resembled the one-sample map of the control disorders (Figure 5A), indicating a much higher level of consistency across multiple distinct psychiatric conditions than among ADHD studies only.

3.7 Control Experiment 2: ADHD Study-Related Coordinates Versus Random Study-Related Decreased Volume Coordinates

Given the surprising lack of specificity of our ADHD-related networks against other clinical cohorts, despite the strong internally consistent distribution described above (Figure 4A), we next sought to investigate whether the network identified here represents a unique finding or is simply a reflection of patterns of decreased volume in general. We statistically compared the results from our set of 40 ADHD studies to those from a set of 40 randomly chosen studies showing decreased volume from the BrainMap [33] database. First, we conducted a one-sample PALM analysis of these random studies, revealing a map that strongly resembled the ADHD map, comprising regions of the basal ganglia, thalamus, hippocampus, and frontal regions like the orbitofrontal cortex (see Supporting Information S1: Table 2 for a full list of peak coordinates). Notably, however, the ACC was largely absent in this map. A subsequent two-sample t test analysis comparing this random study map with the ADHD map revealed no significant differences in connectivity (Figure 5E), and the one-sample t map of 40 random studies had a strong spatial correlation with the one-sample t map of 40 ADHD studies (r = 0.850).

3.8 Control Experiment 3: ADHD-Related Coordinates Versus Random Gray Matter Coordinates

As an additional synthetic control, we also compared our ADHD-related coordinates with an equal number of random gray matter coordinates split into 40 groups. A one-sample analysis of these random “synthetic” studies revealed a small number of seemingly randomly placed significant coordinates, with positive connections in regions of left cerebral white matter and negative connections in the brainstem, cerebellum, caudate, and bilateral white matter. The FWE-corrected two-sample t test of ADHD networks of decreased volume versus these purely random coordinates only identified one small cluster in the brainstem, indicating a stronger connection to the random coordinates than the ADHD coordinates. As with the analysis above using actual studies, the one-sample t map of random gray matter coordinates was still correlated, although to a lesser degree, to the one-sample t map of the ADHD networks (r = 0.666a) (Figure 5F).

4.1 Focal Patterns of Decreased Volume in ADHD Have Been Frequently Reported Yet Are Highly Variable Across Studies

To date, multiple studies have reported distinct regions of decreased volume associated with ADHD [6, 7, 9]. However, these regions exhibit significant spatial heterogeneity, with no single region consistently implicated across studies [7]. As described above, the most commonly reported region (i.e., the right putamen) appeared in only five of the 40 experiments included.

Furthermore, the ALE approach did not identify any singular regions as consistently atrophied in ADHD. ALE is a well-validated technique that has been useful in numerous meta-analyses, finding convergence of regions of decreased volume in neuropsychiatric disorders (e.g., migraine [41] and Tourette syndrome [42]). As the field expands to examine complex neuropsychiatric phenotypes, we expect there will be more cases where ALE produces a limited result, even when appropriately powered, as many of these conditions may represent network phenomena.

CNM may be a more appropriate approach for complex phenotypes as it is more likely to find convergence in disorders reflecting network dysfunction rather than impairment of individual regions [12]. Since ADHD is associated with a spectrum of behavioral phenotypes [1], frequent comorbidity with other neuropsychiatric disorders [1, 43], and likely neuroanatomical variability from patient to patient, ADHD is a useful test of CNM's ability to determine whether patients all have in common an alteration of a consistent brain network, even if via remote functional effects, that is, diaschisis [44].

4.2 ADHD-Associated Volume Decreases Are Connected With the Reward and Action Mode Networks

Our localization (Figure 4A) has face validity and biological plausibility as a neuroanatomical substrate for ADHD. Functional neuroimaging studies have long implicated the mesolimbic pathway in reward processing [45]. While not a diagnostic criteria of ADHD, reward processing is a commonly accepted feature of the ADHD phenotype, with more pronounced attention difficulties in activities with lower immediate reward than highly rewarding activities [46]. However, given the correlational nature of traditional imaging studies, it is unclear whether these decreased volumes represent a cause or downstream effect of persistent atypical reward processing.

Multiple regions in our ADHD network also fall within the AMN, which overlaps with the “salience” network and has been proposed to be a core circuit involved in planning, sustained attention, and goal maintenance [37, 38]. As such, it is eminently plausible that this network is affected in ADHD. Moreover, the AMN is strongly negatively correlated with the DMN [37], which may align with prior observations that found decreased connectivity within the DMN in ADHD [7, 11].

Interestingly, our results also converge with other methods that leverage the network structure of the brain. One recent analysis compared focal structural brain measurements with neuropsychiatric symptoms and found three distinct patterns of decreased volume associated with (1) multiple psychiatric disorders, (2) autism spectrum disorder–related symptoms, or (3) ADHD-related symptoms [4]. While our map has mild similarity with this third map (Figure 6) (r = 0.411), there is a high degree of focal overlap with specific regions identified in the ACC, bilateral putamen, thalamus, and medial orbitofrontal cortex.

4.3 The Lack of Specificity in the ADHD Network Map May Reflect the Brain's Large-Scale Network Structure

Despite the biological plausibility of our ADHD network, we describe that numerous regions positively connected to coordinates representing decreased volume consistently emerged across data sets (Figure 5). Large clusters encompassed areas of the ACC, bilateral putamen, caudate, hippocampus, and brainstem, with smaller clusters across frontal and subcortical regions and the cerebellum. These regions were observed both in disorder-associated coordinates and in purely random gray matter coordinates, albeit to a lesser extent. While these regions are biologically plausible as an ADHD network, the lack of statistical specificity with other conditions indicates that the ADHD results are not nearly as consistent as those seen in other previously studied conditions. Furthermore, the lack of statistical difference between random gray matter coordinate-derived maps and the ADHD maps suggests that either the identified networks are centrally located within the brain's overall connectome and will often appear in CNM analyses in general—a bias of the method—or represent a null finding indicating that these regions are nearly random in location and the ADHD literature has yet to capture any true signal. This suggests a heightened likelihood for these regions to be functionally connected to other brain areas in any individual, indicating that our network may simply reflect the large-scale network structure of the brain. If these networks are highly connected, they are likely to emerge even in random coordinates because of their general functional relevance across the brain. As such, we cannot reject the hypothesis that any set of coordinates could result in a similar network by chance. Notably, this is the first time a large CNM study has found a significant and consistent within-cohort signal that has failed to demonstrate specificity to a disorder when compared with control groups, including purely random coordinates. This result likely stems from the heterogeneity of the original coordinates and may speak to the difficulty in both studying ADHD and conceptualizing ADHD as a consistent diagnosis.

Another possible explanation for the lack of specificity seen here relates to the broader issue of specificity in ADHD diagnosis. The association of both the reward and action mode networks with ADHD is plausible; however, reward processing and sustained attention deficits are associated with a variety of disorders, including autism, depression, and schizophrenia [39, 47]. Thus, it may be unsurprising that we could not demonstrate specificity of our findings to ADHD when compared with control groups with other neuropsychiatric or neurodegenerative conditions [15, 30-32]. Supporting this theory, a recent meta-analysis linked disruptions in the salience network with various common psychopathologies [48], and early ADHD diagnosis has even been associated with an increased risk of later neurogenerative disorders [49]. Involvement of the observed networks, albeit complicated by the network bias described above, may still reflect shared attention and executive function symptoms across various neuropsychiatric conditions. While specific patterns of impairment may exist in different disorders, the literature suggests that most common psychopathologies, including ADHD, are consistently associated with executive function impairments across a wide range of tasks [50, 51].

In fact, given the frequent under- and over-diagnosis, neuropsychiatric comorbidity, and phenotypic heterogeneity [2] of ADHD, studying individual symptom profiles may be necessary to identify brain regions or networks associated with ADHD. This bottom-up approach aligns with the National Institute of Mental Health Research Domain Criteria initiative, which aims to classify mental disorders based on neurobiological substrates and dimensions of observable behavior beyond the symptoms specified by current categorical diagnoses [52]. While categorical diagnoses have practical clinical utility, research on symptom profiles may help refine the diagnostic criteria of ADHD to better resolve the heterogeneous biological reality.

4.4 Limitations

Several limitations of the current study must be acknowledged. Due to a lack of control data sets with coordinates representing volume increases, any findings of increased volume across the ADHD studies were excluded. An exploratory one-sample t test with permutations of these increase coordinates found no significant voxels when FWE-corrected. With a less stringent false detection rate correction, peak coordinates appeared in the cerebellum, fusiform gyrus/inferior temporal gyrus, brainstem, postcentral gyrus, thalamus, and cerebral white matter (see Supporting Information S1: Table 3 for details). Additionally, due to the absence of participant-level coding, differentiation between children and adults or among the three ADHD subtypes was not feasible. It is possible that the networks involved in ADHD change throughout the lifetime, and/or that different networks are associated with inattentive, hyperactive/impulsive, and combined ADHD. Nonetheless, the current analysis identifies commonalities across these dimensions, and the diversity of our sample would only hinder our ability to detect a consensus network. Additionally, consistent with the prevalence of ADHD, there was a high proportion of males in this analysis. Most studies used matched cohorts to mitigate sex-related effects; however, future work leveraging participant-level coding (mega-analysis) may be able to address ADHD-related sex differences. Of note, repeating our analyses with a male-only connectome did not alter the results. Finally, a small number of participants had comorbidities common to ADHD, including learning disorders, oppositional defiant disorder/conduct disorder, anxiety, mood disorders, and tic disorders [43]. Therefore, it is possible there would be some overlap between the networks involved in ADHD and those involved in other disorders. While not ideal, the inclusion of participants with comorbidities should only bias against identifying consistent effects here.

4.5 Future Directions

Future meta-analyses examining regions of increased volume and/or increased and decreased fMRI activation may provide a more complete picture of network involvement in ADHD. However, leveraging fMRI studies to examine long-term traits are confounded by the specific tasks performed compared with structural source data used here. It would also be beneficial to study differences in brain networks between age groups, gender, and ADHD subtype. Given the possibility that the lack of specificity in our results was due to heterogeneity within individuals diagnosed with ADHD, future studies would be enhanced by prospective data collection with in-depth, objective ADHD symptom-specific characterization. Finally, studies that directly modulate brain activity or connectivity of specific brain regions or networks, via direct invasive or noninvasive neuromodulation or fMRI neurofeedback, may help disambiguate which regions are more directly and causally associated with impaired attention.