We detail the structures contributing to a recently reported resting state network in the thalamus and basal ganglia. By using a high field, high BOLD sensitivity experiment design and high resolution analysis, we show that it encompasses not only the thalamus, pallidum, putamen and transverse temporal gyrus - as has been previously noted - but also the substantia nigra and subthalamic nucleus, allowing the basal ganglia circuit to which it corresponds to be identified. Despite its non-observation in most resting-state studies to date [9, 21–23] it was found to be reproducible across subjects and MR measurement parameters and is evident in both the eyes closed and fixation resting state conditions. The network is positively correlated with the left and right lateral fronto-parietal (also called attention or task-positive) networks, and anticorrelated with the Default Mode, whose hemodynamic response it anticipates. The peak frequency and spectral characteristics are similar to those of other RSNs, distinct from physiological fluctuations, and allow it to be classified as an RSN using the unsupervised classifier described. We proceed to identify which of the parallel segregated basal ganglia circuits this network corresponds to based on participating regions.
Because of their historical significance as motor structures, the best studied basal ganglia network is the motor circuit. Cortical input from precentral motor areas, postcentral somatosensory areas, the arcuate premotor area and the supplementary motor areas projects dominantly to the putamen. The putamen sends projections to the interior segment of the pallidum and on to particular thalamic nuclei (the direct pathway) as well as the internal segment of the globus pallidus via the caudolateral substantia nigra to the thalamus (the indirect pathway). Outputs from the thalamus project to the supplementary motor area, the motor cortex and arcuate premotor area, probably in distinct subcircuits . The observed resting state network corresponds well with the motor circuit; the caudate is notably absent, and the putamen constitutes the focus of activation. The activated regions identified as substantia nigra are consistent with coordinates derived from stereotactic electrophysiological studies . The presence of the substantia nigra suggests that the resting state network corresponds either solely to the indirect pathway of the motor circuit, which acts to inhibit movement, or both the indirect and direct pathways. Weak activation in the supplementary motor area is likewise consistent with the motor circuit. Myeloarchitectonics suggest that the thalamic activation corresponds to the ventro-lateral thalamic nucleus, which a prior DTI study has shown to connect dominantly to motor regions . In fact, endogenous BOLD fluctuations in this part of the thalamus have been found to be strongly correlated with motor and premotor areas by Zhang et al. ( - supplementary material) as well as the whole putamen.
As well as there being good agreement between the areas observed in this RSN and the motor circuit, there is disparity between the principle sites of input to the striatum and the other circuits. The caudate provides input to the oculomotor, dorsolateral prefrontal and lateral orbitofrontal circuits and the ventral striatum provides input to the anterior cingulate circuit. The cortical regions to which the circuits send output are the frontal eye fields for the oculomotor circuit and the dorsolateral prefrontal cortex, the lateral orbitofrontal cortex and the anterior cingulate area for those circuits respectively.
No components were identified corresponding to the associative and limbic thalamo-cortical loops . The question arises of whether it is to be expected that an RSN exists for each of the tripartite or Alexander subdivisions of the basal ganglia, and whether basal ganglia RSNs (if present) would be expected to correspond exactly with circuits that have been established from anatomical and afferent projections from the cortex. Looking outside the basal ganglia, we know that whilst RSNs reflect functional task networks, it is not the case that there is an RSN for every functional network defined by task or anatomy. Where there are spontaneous fluctuations in networks at rest, other studies have shown that there is not complete correspondence with the network as defined by anatomy or tasks. The major anatomical afferents from the cortex suggest a functional partitioning of the putamen, with a minor portion classified as associative and limbic, and the majority as being sensorimotor. Zhang et al. found functional connectivity between the motor and premotor cortex and almost the entire putamen , however, as observed here. The temporal cortex also displayed a much weaker correlation with the caudal-ventral putaminal region, a finding not consistent with the classical striate nucleus tripartite subdivision. In addition, when looking at RSNs involving associative cortical area, such as the Default Mode, ventral attention, dorsal attention and executive control networks, only a portion of the caudate nucleus participates, instead of the entire associative subdivision of the striate nucleus.
Our hybrid simulations suggest that ICA is capable of separating circuits which overlap to some extent, but which also have non-overlapping elements and different temporal behaviour (Additional File 3). Further evidence of this is provided by the fact that some brain regions (such as the thalamus) are present in multiple networks. The most likely explanations for not finding RSNs corresponding to the associative and limbic thalamo-cortical circuits is, therefore, that they either do not show spontaneous low-frequency BOLD fluctuations or that these fluctuations are below the sensitivity of this study.
Di Martino et al. have reported a resting state functional connectivity analysis focussing on the striatum  independent of our preliminary reporting of these results . A gender-mixed group of subjects were studied while fixating on the word "Relax" in a study with a sensitivity likely to be equivalent to that here; applying a similar, relatively short TE EPI protocol and comparable number of total image volumes (6895 c.f. 7800 here). The connectivity results obtained for seeds in the dorsal and caudal putamen in the Di Martino et al. study is similar to the RSN observed by us , although the involvement of the substantia nigra (which contributes substantially to the motor circuit attribution) was not reported. Also, the dorsal putamen seeds indicated correlated activity in the anterior cortex cingulate, which relate to executive function, indicating some mixing of fluctuations relating to motor and executive control circuits. The basal ganglia RSN reported here was recently noted as an incidental finding by Damoiseaux et al. , adding independent verification of these results.
The conditions under which this resting state data were acquired were similar to those used in a number of previous studies which also applied group ICA approaches [9, 21–23]. The question then arises as to why this network was not reported in those studies, but has been observed only here and in the most recent publication by Damoiseaux et al. . The basal ganglia resting state network was one of the weakest identified in this study, measured in terms of the percentage of total variance in the data it explains. The sensitivity of this study is likely to be higher than that in the studies cited due to the high field strength (3 T), relatively large number of subjects (26) and short repetition time (1 s). The echo time of 28 ms is quite short and well matched to the T2* of basal ganglia structures at 3 T , yielding optimal BOLD sensitivity. T2* is shorter in the basal ganglia than is typical in the cortex due high iron concentration . While it is likely that, other than reference , previous studies were not sufficiently sensitive to detect this network, it is also possible that it was simply overlooked in the wealth of components.
Previous studies using functional connectivity analyses have demonstrated an anticorrelation relationship between the Default Mode and the dorsal attention network in the resting state [5, 11, 42, 43], which has been interpreted as indicating an interplay between modes thought to reflect stimulus-independent thought and goal-driven activity. These anticorrelation findings have been recently called into question, however, as global signal subtraction performed as a pre-processing step to reduce the influence of physiological noise and scanner drift itself introduces anticorrelations [27, 28]. The approach taken here is not subject to these problems. We calculated the correlation between independent component time courses, with no prior global signal subtraction. The most significant results were as follow. The anticorrelation finding reported previously between the Default Mode and the dorsal attention network [5, 11, 42, 43] was reproduced. The basal ganglia network was also found to be anticorrelated with the Default Mode, and correlated, to a similar degree, with the lateralised attention networks. Although we and others have shown that RSNs possess similar frequency characteristics, it must be the case that they are at most weakly coupled, or they would not be separable in functional connectivity analyses . Correlation values observed here between component time courses are correspondingly low - in the range of 0.19 to 0.27 in magnitude - consistent with those observed in other studies . Despite the fact that they are weak, their consistency across subjects is such that these results are highly significant. There have been suggestions that the thalamus, with involvement of the basal ganglia, may be responsible for instigating the task-independent deactivation of the Default Mode observed when subjects are posed cognitively demanding tasks . This would be consistent with findings by Uddin et al., which have shown that correlations between homologous RSN structures in the cerebral hemispheres are preserved in a patient with complete commissurotomy, indicating that functional connectivity can be mediated by subcortical structures . The hypothesis that the basal ganglia RSN represents the controller of fluctuations in the Default mode is supported by the results of the functional network connectivity analysis. In 18 out the 20 randomly composed groups of 13 subjects, the time course of the basal ganglia independent component preceded that of the Default Mode. Although this demonstrates that the hemodynamic response in the basal ganglia RSN consistently precedes that in the Default Mode, latency differences in the hemodynamic response functions  in both networks would have to be analysed and corrected for  before concluding that activation in the basal ganglia network precedes deactivation in the Default Mode. Possible differences between the precedence of neuronal activation and the measured MR response arising from hemodynamic shape and latency effects  preclude testing order hypotheses with this and other approaches such as Granger Causality Modeling.
Although RSNs show a maximum in frequency spectra in the range 0.01 - 0.04 Hz, this does not accurately reflect the frequency distribution of underlying neuronal fluctuations. The intrinsic autocorrelation of BOLD fMRI data has 1/f behaviour in the frequency domain , and low frequencies are cut off by sampling over a finite duration, leading to a low frequency peak. In fact, when the hemodynamic response function is deconvolved from RSN time courses prior to frequency analysis, the spectra of RSN components are essentially flat up to 0.1 Hz . The analysis of RSN frequencies in this work serves two purposes, neither of which relate to the absolute frequencies observed in the spectra. The first is that differences between the compositions of RSN spectra can be observed. The frequency spectrum of the basal ganglia network is most similar in composition to that of the Default Mode, reinforcing the connection between the two networks. The second is that RSN spectra may be distinguished from components of physiological origin because their frequency distributions reflect the intrinsic convolution with the hemodynamic response function. This is the basis for the classifier used here, which can reliably distinguish RSNs from physiological components using frequency characteristics alone . The basal ganglia RSN clusters clearly with other RSNs.
It is likely that the component which consists solely of the caudate (Figure 3) reflects activity relating to the inhibition of ocular saccades. As such, caudate activity is a task-specific response rather than an RSN. The caudate component is apparent in the fixation study, in which there was a point fixation condition, while no corresponding activity is apparent in the eyes closed study. The role of the caudate in saccadic eye movements is well documented [55–57]. Fluctuations in caudate activity in this capacity (which are a pre-requisite of their identification in an ICA) may arise due to phasic preparation of reflexive saccades and their voluntary inhibition. Ultimately controlled by the superior colliculus, saccades can be generated and inhibited by the input of the caudate nucleus, via efferent projections to the substantia nigra pars reticulata . The oculomotor basal ganglia-thalamocortical circuit, includes, as origins of input to the caudate, frontal eye fields, dorsolateral prefrontal cortex and posterior parietal cortex (Brodmann's areas 8, 9&10 and 7, respectively)  all of which areas are apparent in this component. Frontal eye field neurons are known to fire during passive fixation  as was the condition in the fixation study. The oculomotor network yielded from the dorsal caudate seed in the Di Martino et al. analysis  is similar to the caudate component identified here in the fixation study. Our hypothesis that this is a task-related response to the suppression of ocular saccades is consistent with the use of a fixation condition in that work, and it not being observed in the many previous studies which have used the eyes closed condition [9, 21–23, 32].
Another area of application of this network is as a candidate marker for neuro- and psychopathologies involving the basal ganglia. In a parallel with attempts to use Default Mode activity as a diagnostic marker for Alzheimer's disease , deficits in the basal ganglia RSN may offer a marker for one or more of the diseases in which the basal ganglia are known to play a role. Parkinson's disease is a basal ganglia disorder characterized by the degeneration of dopaminergic neurons in the striatum. This has been shown to have opposing effects on the direct and indirect pathways in the striatum , with hyperkinesia as a result. Functional connectivity in the basal ganglia-cortical circuit has been demonstrated in this condition via synchronous oscillations between local field potentials in the basal ganglia and cortical EEG , the transmission to the basal ganglia of motor cortex electrostimulation in the monkey and rat  and oscillatory high-voltage spindles). Dopaminergic lesion in a rodent model for Parkinson's has been shown to lead to an increase in oscillatory synchronisation in the basal ganglia and increase in frequency and duration of high voltage spindle events). Similarly, excessive synchronisation of subthalamic nucleus neurons has been confirmed as a cause of movement slowing in Parkinsonism . Dopamine-dependent changes in the functional connectivity between the basal ganglia and cortex have likewise been demonstrated . The discovery of the fMRI manifestation of this functional connectivity will allow dysfunction in this system to be probed non-invasively, even if patients are not able to perform motor tasks appropriately.