Keywords

Translational fMRI: Aims, Challenges, and Opportunities

The cognitive neuroscience of memory and language has been enhanced considerably by fMRI as a technique to probe and map these functions non-invasively. Such studies afford key insights into the neural architecture of these processes, with more recent studies focusing on networks of coherent activity that underlie cognitive operations. However, the translation of these insights into clinical applications has lagged behind, particularly when one considers evidence that goes beyond describing a difference between a particular patient group and the healthy brain controls. The clinical literature is typically hampered by small samples of heterogeneous patient groups, with little replication of findings across centres given adherence to different types of activation paradigms and variation in the clinical outcomes assessed. The emergence of a robust set of data using rsfMRI, where there are no task constraints, has begun to open new avenues for exploring more direct relationships between cognition and alterations in neural networks.

Translational work aims to understand which signal characteristics are meaningful in explaining variance in disease phenotype and severity, likely progression, treatment response, or reorganization capacity. In individuals with mTLE who experience recurrent seizures arising from one MTL, the concern is whether removal of anterior temporal lobe structures, in attempting to eliminate seizures, will come at a significant cost in the form of postsurgical decline. In the domain of language, this is typically seen when the anterior temporal neocortex of the dominant (typically left) hemisphere is excised [1,2,3]. For memory, the functional adequacy of the epileptogenic MTL is key, in that better preoperative capacity is associated with greater decline [4, 5], typically with verbal memory associated with the left MTL and visuospatial memory with the right MTL [6,7,8]. In principle, fMRI is uniquely suited to investigate the functional anatomy of memory and language processes by providing a direct assay of adequacy of the epileptogenic tissue and the risk of potential post-operative morbidity, as well as the opportunity to characterize possible reorganization or compensation [9]. Another setting where fMRI can be an effective probe for functional adequacy of MTL is amnestic mild cognitive impairment (aMCI), a prodromal stage of Alzheimer’s disease (AD) where memory is selectively impaired, in which the entorhinal cortex and hippocampus reveal earliest signs of structural pathology [10,11,12,13]. Indeed, the MTL has been the focus of much of the foundational work using fMRI as a biomarker of diagnosis and disease progression, as well as to predict the potential impact of treatments including pharmacotherapy and cognitive training [14, 15].

Much of the early translational application in these clinical contexts was driven by questions of task-related patterns of activation, with a particular focus on regions of interest. Relevant findings include reports showing less MTL activity during an encoding task for patients with mTLE [16,17,18,19,20] or MCI/AD [21, 22] compared to healthy controls. Similarly, the strength (magnitude, spatial extent) of activation in the left temporal lobe has been shown to predict the degree of decline in word retrieval following left temporal lobectomy in mTLE [23,24,25]. While this approach has been quite successful, there are also important limitations when one considers the long scanning times that may be required to produce reliable data, the concern about capacity for full task engagement in the patient population, the lack of consensus regarding sensitivity and specificity of particular metrics of activation, and the possibility of alternate or more diffuse patterns of organization within the disease group that might undermine the focus on one particular region of interest. Furthermore, there are examples in the literature showing that greater signal does not always signify better function, such as a study demonstrating strong bilateral hippocampal activation in a patient during an episode of transient global amnesia [26] and evidence that enhanced activity of the hippocampus may be a pathologic signal in aMCI [27, 28] . Moreover, we found that despite controls and aMCI patients showing similar task-related activation in the parahippocampal gyrus, activity in quite distinct networks was associated with good memory performance in the two groups [29]. While much effort has been aimed at identifying ideal task activation paradigms for clinical use [30,31,32], there is a growing scientific interest in the newly characterized ‘resting state’ networks that can be discerned when individuals are scanned doing nothing other than ‘mind-wandering’ in the scanner. We will briefly discuss the foundational work with healthy individuals before turning to the disease states.

Resting State Networks

The first described resting state network was termed the default mode (DMN) as it characterized a set of midline and lateral cortical regions that were co-activated during the ‘rest’ condition of a vast number of cognitive paradigms [33]. Subsequent work has shown that these regions show robust functional connectivity during introspective thought or ‘mind-wandering’ [34, 35]. Furthermore, the DMN can be fractionated into components that are collectively engaged during particular cognitive operations such as constructing a mental scene from memory and engaging in self-relevant decisions [36]. A number of other networks have been identified, characterized by intrinsic connectivity that can be captured from 6 to 10 min of scanning while participants are at rest, i.e. not engaged in any cognitive task beyond (typically) focusing on a fixation cross. The networks can be identified using seed-based techniques, which extract temporal correlations between a particular seed region and all other brain voxels, or using independent component analysis (ICA), which decomposes voxel-wise temporal correlations into spatially orthogonal components. The latter is the more common data-driven technique that has been used to identify a large number of robust cortical networks in the healthy brain, including those associated with attention (both dorsal and ventral), language, cognitive control, and more basic visual and sensorimotor functions [37,38,39,40]. By exploiting the massive high-quality rsfMRI data from the Human Connectome Project, it has been possible to further understand that large-scale networks can be decomposed into partially segregated subnetworks and how some regions in association cortex can participate in multiple networks to different degrees; such complexity undergirds the flexible behaviour that arises from those network interactions [41, 42].

rsfMRI and Memory in mTLE

There is a moderate record of success in the clinical application of paradigms eliciting focal activation in the MTL as applied to mTLE. These include verbal or visual memory paradigms in which the degree of activation in the epileptogenic MTL correlated with current memory performance or predicted the magnitude of post-operative memory decline [17,18,19, 43]. Several studies have shown that asymmetry of MTL activation during encoding is more strongly correlated with current memory function [20] or with decline in memory following surgery [44] than standard clinical parameters such as hippocampal volume. However, the success at predicting decline following right anterior temporal lobectomy (ATL) has been less successful than for left ATL, and at least one study demonstrated that activation associated with language lateralization was superior to memory-specific activation in predicting verbal memory decline following left ATL [31] . This may reflect relatively poor sensitivity or specificity to memory outcomes of the paradigms, the metrics, or both.

Resting state fMRI may be particularly well-suited to address questions of functional adequacy of the MTL because the hippocampus and parahippocampal gyrus are components of the DMN, which shows considerable overlap with the constellation of brain regions commonly engaged during episodic recollection [45, 46] . Furthermore, such functional connectivity (rsFC) measures may be superior to task-based measures as they eliminate variance or noise associated with different strategies and/or alternate networks being utilized during the activation task by patients in whom some degree of functional reorganization may have already taken place [47]. There have been a number of reports of abnormal connectivity between the MTL and other DMN regions in patients with mTLE [48,49,50,51], but only a few studies investigating the consequence of that disrupted connectivity to functional integrity as indexed by memory performance. In our initial study of this issue [52], we focused on two regions of interest in the DMN, the hippocampus and posterior cingulate cortex (PCC), which is considered one of the main hubs of the DMN and is commonly activated during recall and recognition. Relative to controls, we found reduced connectivity to the epileptogenic hippocampus and increased connectivity to the contralateral one in both left and right mTLE patients. Furthermore, stronger connectivity on the epileptogenic side was associated with better pre-surgical memory and greater postsurgical memory decline, whereas greater connectivity on the contralateral side was protective. We also found an increase in contralateral connectivity after removal of the epileptogenic side and that increase was also correlated with memory preservation suggesting compensatory plasticity. A subsequent study examined functional connectivity throughout 20 nodes of the DMN in relation to performance on clinical memory tests [53]. Although more complex, the pattern of connections was even more strongly correlated with memory measures than the simple two-node solution. Here, better memory was associated with increased posterior and interhemispheric connectivity (i.e. between MTL structures and medial and lateral parietal cortices), whereas poorer memory was associated with a pattern that emphasized stronger long-range posterior-to-anterior intrahemispheric connections. Of interest, a very similar pattern of stronger long-range connectivity was seen in a different cohort of left mTLE patients when they were actively attempting to retrieve personal autobiographical memories [54]. As this alternate pattern of connectivity was seen under different circumstances (resting state and directed memory retrieval) and with distinct analytic techniques (multivariate correlations and structural equation modelling), it may signify a robust change in memory networks in cases of insult to the MTL (Fig. 3.1).

Fig. 3.1
figure 1

Increased long-range intrahemispheric connectivity as a consequence of unilateral hippocampal damage in mTLE. (a): Blue lines indicate increased (relative to controls) long-range connectivity in rsfMRI from posterior medial and lateral parietal cortices to medial prefrontal regions in patients with left mTLE (Data from McCormick et al. [53]). (b): Altered effective connectivity (relative to controls) during autobiographical memory retrieval (Data from Addis et al. [54]). Major differences are increased connectivity (red lines) from retrosplenial to medial prefrontal cortex on the left and reduced connectivity (blue lines) involving left medial temporal regions. Red nodes in both figures indicate regions of reduced structural integrity of medial temporal cortex, and the width of lines indicates connectivity strength. (This figure is reproduced with permission from [115])

Full size image

Another group has investigated networks from both task and resting state fMRI in relation to memory processes. Their first study, using ICA on data collected during memory encoding, documented reduced functional connectivity that was associated with poorer memory performance in a network that included bilateral MTL and extended to bilateral occipital regions as well as left orbitofrontal cortex [50]. Of interest, focal activation in the orbitofrontal region, and not the ipsilateral hippocampus, was strongly correlated with task performance. In a more recent study using seed-based rsFC to investigate different parcellations of thalamo-cortical networks, they reported short-term memory was related to thalamic connectivity with specific regions in frontal and parietal cortices in the contralateral hemisphere, whereas long-term memory was associated with the strength of rsFC between ipsilateral thalamus and entorhinal cortex [55]. A similar reduction in connectivity between ipsilateral temporal neocortex and the DMN was reported for left TLE based on ICA-derived networks in another study that did not directly examine memory performance [56]. Thus, connectivity amongst many cortical regions may need to be considered, as epileptic seizures may promote reorganization outside of canonical memory networks.

It is also important to know whether these functional connectivity measures go beyond other indicators of structural integrity or functional activation in explaining variation in memory performance; this speaks directly to their potential added value in clinical contexts. Of note, studies show wide variation in the correspondence between functional activation and structural measures of hippocampal integrity, ranging from no correlation between measures [43] to a moderately strong relationship [57]. In their study assessing thalamo-cortical connectivity and memory, Voets and colleagues noted that posterior hippocampal volume did add unique information to thalamic-frontal rsFC in explaining short-term memory performance, but this was not the case for the relationship between long-term memory and thalamic-medial temporal rsFC [55]. In both McCormick studies from our lab described above, performance on clinically relevant memory tests was more strongly predicted by rsFC than by structural damage to the MTL or other DMN nodes (assessed by hippocampal volume and grey matter density). Indeed, we have consistently found the influence of structural alterations on episodic memory capacity to be mediated by functional network capacity [20, 53]. These findings accord well with the concepts underlying the human brain connectome that structure undergirds but does not fully determine the brain’s dynamic and flexible functional repertoire [58, 59]. Unfortunately there is no published work that directly compares the relative clinical utility of different fMRI measures such as focal activation, task-related connectivity, and resting state connectivity, but we can offer some unpublished observations from our own work. In the same group of ten patients with left mTLE, we compared focal left hippocampal activation during an autobiographical memory task that strongly engages the hippocampus and PCC [54, 60,61,62], rsFC between left hippocampus and PCC, and connectivity between those nodes during autobiographical recall [47]. When each measure was correlated with a standard clinical measure of verbal memory, we found the best correlate of behaviour to be rsFC (r = 0.78), followed by task-related FC (r = 0.55), followed by hippocampal activation (r = 0.31). We contend that similar comparisons in other clinical settings will be very important in advancing the predictive utility of fMRI.

Although these early results are promising, there are important complexities that must be addressed. For example, a few studies looking at connectivity and current memory capacity seeded from the hippocampus (rather than PCC as we had done) to explore widespread connectivity throughout the brain. One reported a pattern similar to ours, in that higher contralateral HC-PCC connectivity was associated with better episodic memory, and also found that connectivity between the ipsilateral or contralateral hippocampus and other MTL structures (entorhinal cortex and parahippocampal gyrus) was negatively associated with memory performance [63]. Another study reported that connectivity between the ipsilateral left hippocampus and left PCC/precuneus and inferior parietal lobule (IPL) was negatively associated with memory, whereas better performance was related to stronger connectivity between the epileptogenic hippocampus and contralateral PCC/precuneus and IPL [64]. Finally, in correspondence with Holmes but in contrast to McCormick papers, higher ipsilateral hippocampus to PCC in patients with left TLE was related to poorer verbal memory, whereas contralateral hippocampus connectivity with the medial frontal cortex was associated with better nonverbal memory in patients with right TLE in another study [65]. These findings indicate that not all alterations in connectivity are functionally efficient or compensatory, a lesson we have already learned from reports of increased focal activation, and also support the need for larger studies and especially those with prediction of change as the primary outcome.

rsfMRI and Memory in Other Patients with MTL Damage

Alterations in task-based activation in the medial temporal lobe have been reported in aMCI and AD, although the patterns are not invariably of a decline; for meta-analyses based on activation co-ordinates, see [66, 67]. Similar to the arguments for mTLE, however, assays of functional integrity in aMCI, AD, or other amnestic cases of focal MTL pathology that are based on extent or strength of activation in the MTL can yield incorrect inferences. For example, current data indicate that early aMCI is characterized by hyperactivation in the hippocampus, which appears to be pathological rather than compensatory [28, 68]. That is, there is an inflection point in the relationship between hippocampal activation and memory integrity as one proceeds from normal ageing, through very mild and more severe aMCI to AD [69]. Instead of hippocampal engagement, a more reliable indicator of encoding success and overall memory status in AD was reported to be the degree to which the medial parietal region (PCC, precuneus) was deactivated during memory encoding [70, 71]. These are two key nodes of the DMN, and there is compelling evidence of a global decrease in DMN integrity in aMCI and AD compared to age-matched controls; for reviews, see [72, 73].

Furthermore, a growing set of studies relate these alterations in rsFC to clinical memory impairment. Amnesic patients with bilateral MTL damage show decreased connectivity between PCC and MTL components of the DMN [74]. The degree of rsFC between these regions is correlated with memory performance and separates patients with amnestic MCI from non-amnestic MCI [75]. Other rsFC studies have described more widespread disruptions in connectivity between the hippocampus or PCC and other neocortical and subcortical regions that are correlated with the degree of memory impairments in MCI [76, 77]. As MCI and dementia will be discussed in more depth in another chapter in this volume, the foregoing is intended only to indicate that studies regarding fMRI biomarkers of memory integrity are being pursued in this disorder and that work from mTLE and aMCI may provide fruitful opportunities for cross-fertilization.

Hippocampal Parcellation and Connectivity

There is growing interest in examining connectivity patterns involving anterior and posterior segments of the hippocampus separately. Recent research has demonstrated that these segments have distinct multisynaptic patterns of connectivity with neocortical regions, which likely has considerable import for the type of memory processes they support [78,79,80,81]. Several different characterizations have been offered that distinguish these anterior and posterior networks. One focusses on the granularity of mnemonic contents with the more fine-grained representations involving posterior regions, enabling retrieval characterized by recollection of details rather than being limited to coarser ‘gist’ information [80]. Another model hypothesizes that encoding new information is the province of the anterior hippocampus-dorsal attention network connectivity, whereas retrieval of any sort depends more so on posterior hippocampus-DMN connectivity [82].

Although the histopathology of TLE is distributed along the longitudinal axis of the hippocampus [83], there is evidence that greater structural abnormalities on MRI are seen in the anterior half and there is better chance of seizure freedom in that setting [84,85,86]. Thus, it may be important to isolate contributions from the two segments in assessing functional capacity in this population. One recent study found similar general patterns of heightened connectivity amongst MTL regions for anterior and posterior segments, whereas only the latter region showed reduced connectivity to another cortical target (PCC) that was associated with defective memory in patients with left mTLE [63]. Of interest, our own work with healthy controls has also demonstrated different connectivity patterns, with the posterior hippocampus-to-neocortex networks showing a greater involvement with the type of relational memory processes that are particularly impaired in mTLE patients [87, 88]. There has not yet been much work on this anterior-posterior segmentation concept in the MCI/AD populations. In one study with no behavioural correlates, it was reported that reductions in spatial extent of these networks were found in aMCI patients compared to controls, with the most striking appearing in the left posterior hippocampus and right anterior MTL [89]. Clearly, this is an important avenue for future investigations in these populations, marrying rsfMRI to well-designed cognitive assays of clinically meaningful memory processes.

rsFC and Language in TLE

Resting state fMRI is also emerging as a tool for mapping the language network, composed of the inferior frontal gyrus, superior temporal gyrus, supramarginal gyrus, inferior parietal lobule, and premotor cortex. While these regions have been interrogated extensively using task-based fMRI [90], researchers have also found reliable rsFC amongst these regions [91, 92]. Here, we review the literature on rsfMRI and language network activity, discussing its findings and applications in healthy and disease populations. We will also discuss the similarities and differences between pre-surgical rsFC network characterization and traditional methods of pre-surgical language mapping, noting the potential advantages of each.

Language networks have been extracted from rsfMRI data using both seed-based [92] and ICA-based [91] approaches . Tomasi and Volkow [92] examined 970 healthy individuals collected from 22 research sites around the world. The authors used a standard seed-to-voxel approach with two regions of interest – Broca’s and Wernicke’s areas . Both seeds demonstrated robust connectivity to language network regions including inferior frontal gyrus (IFG, comprised of pars orbitalis, triangularis, and opercularis), superior temporal gyrus, inferior parietal cortex, middle frontal gyrus, inferior temporal cortex, superior frontal cortex, caudate, putamen, and cerebellum. This network connectivity was reliable across all sites and both seeds, although Broca’s area was significantly more connected to anterior language regions, while Wernicke’s area was significantly more connected to posterior regions. Furthermore, they found leftward lateralization of Broca’s area and posterior Wernicke’s area (angular gyrus) connectivity which supports lateralization of language to the left hemisphere as is expected in the general population. A subsequent study using the same procedures found these networks to be highly reliable within subjects across time intervals ranging from 45 min to 16 months [93]. This level of consistency in the network provides reassurance that low-frequency fluctuations at rest are related to intrinsic, stable network properties. In addition, there is evidence that connectivity in this network is associated with language performance. Reduced resting state connectivity was observed between posterior temporal regions related to reading (fusiform gyrus, inferior temporal gyrus, middle temporal gyrus, and superior temporal gyrus) and the left IFG in dyslexic readers compared to nonimpaired readers [94]. Furthermore, increased functional connectivity between the fusiform gyrus and left IFG was associated with better reading fluency.

The mounting evidence that rsfMRI can identify reliable networks and clinically relevant language functional capacity provides confidence that it may provide important information for pre-surgical language mapping for individuals with TLE. Prior to epilepsy surgery, it is imperative to establish hemispheric dominance and, occasionally, fine-grain localization of eloquent cortex. Traditionally this was established with invasive tests such as intracarotid injection of an anaesthetic agent such as sodium amobarbital (Wada procedure, [95]) or electrical stimulation mapping (ESM, [96]). More recently, fMRI using activation tasks has become a standard of practice in many epilepsy surgical centres, as it has been shown to have high concordance with Wada results [97,98,99,100,101]. Concordance with ESM has not been found as consistently [102], although using multiple language tasks does improve agreement [103]. Results from task-based fMRI are reliant on several factors such as the selection a threshold for activation [103, 104], calculation of laterality indices [105, 106], the type of task demands and control conditions [107,108,109,110], and the ability of the participant to perform the task, features that are not constraints for rsfMRI.

As with memory, there are only a few instances of this newer technique being applied in the pre-surgical mapping context. In one, language networks were extracted from resting state scans using a machine learning algorithm that was trained to classify voxels as being a part of one of seven canonical resting networks [111]. The patients also underwent ESM for language, and the electrode placement was co-registered to anatomical space from the MRI scans. They found strong agreement between ESM and rsFC language network classification. Similarly, moderate to strong concordance has also been found between resting state language network patterns and task-based fMRI activation patterns [112, 113]. Furthermore, greater left lateralization, determined by task activation, is associated with stronger connectivity between left IFG and other neocortical regions in the left hemisphere in controls and TLE patients [114].

While these findings are encouraging, more work is needed to validate that the information obtained from rsfMRI is informative for surgery if it is to replace invasive procedures or task-based fMRI . A large-scale comparison of rsfMRI laterality and Wada or other fMRI metrics has yet to be performed, and importantly, few studies have related rsFC to pre-surgical language performance or postsurgical language changes . We offer relevant observations from our own clinical series in which TLE patients had a resting state scan as well as a language activation task (involving naming to description compared to fixation control). Correlations with performance on a standard clinical measure (Boston Naming Test ) were strongest for task-based functional connectivity between left inferior frontal and left middle temporal gyri (r = 0.50), weaker for rsFC between those regions (r = 0.25) and weakest for the task activation laterality index (r = 0.16); note the latter two values are previously unpublished data. However, using newer analytic techniques with larger language networks, we have begun to identify reasonably strong relationships between network integrity, as a function of similarity to controls, and post-operative changes in naming performance [115]. As with memory, there are compelling reasons to develop a strong clinical grounding for rsFC in language processes. Any patient may be incapable of performing our simple language tasks adequately or in the manner we designate, and thus our task-based patterns of activation may include regions that are engaged for the fMRI task but not truly indicative of language capacity.

Newer Metrics: Graph Theory and Dynamic Functional Connectivity

The number of interregional correlations that can be generated in a typical whole-brain rsfMRI study can yield a bewildering array of findings (e.g. 96 regions of interest will generate 4560 correlations). Graph theory is a mathematical technique that enables one to characterize complex patterns of pairwise correlations between objects (which could be brain regions, airport locations, genes, or people in a social network). Graphs are made up of vertices (nodes) which are connected by edges (connections), and vertices can interact through direct connections or indirectly across paths that are composed of multiple edges. There are a variety of metrics that have been applied to the study of brain organization, both at the level of structural and functional connectivity, aimed at characterizing typical features as well as those associated with development and pathology. A full accounting of the technique and specific metrics is well beyond the scope of this chapter, but the excellent work of Olaf Sporns, who carried out the foundational work and continues to explore its explanatory power in neurosciences, is recommended [116,117,118,119]. There are two principles or modes of brain connectivity that are key to cognition, and several metrics have been articulated to capture these. One is segregation, which refers to how neural elements form separate discrete clusters that can enable specialized local processing, and the second is integration, which refers to the capacity of the network as a whole to be interconnected to enable coordinated activity across neural populations. The manner in which individual nodes may contribute to the network’s integrity and information flow is captured by influence measures. The most efficient interplay of segregation and integration is reflected in a ‘small-world’ organization based on functional connectivity patterns characteristic of the healthy primate brain [120, 121].

Graph metrics are now being used to characterize differences between the healthy adult brain and patients with MTL damage in mTLE [122,123,124,125,126] and MCI/AD [127,128,129,130]. To date, there have been few studies that have attempted to link alterations in graph characteristics to specific cognitive functions in these patient populations. Based on rsfMRI collected preoperatively, Doucet and colleagues examined characteristics of regions selected for their known involvement in specific cognitive domains (nine regions for verbal episodic memory, six for nonverbal episodic memory, ten for working memory/executive functions, and seven for expressive language) and their relationship to neuropsychological outcome in each domain following anterior temporal lobe resection [131]. They focused on properties of segregation (local efficiency), integration (distance), and influence or centrality (participation) for each of the nodes in an attempt to identify characteristics that might be selective to the functional domain and demonstrated strong utility of these measures as shown by significant prediction of outcome across most regression models. Considering language in left TLE, they found that integration amongst canonical subregions in the left inferior frontal cortex (pars orbitalis, pars triangularis, pars opercularis) was the most significant predictor of outcome but also that integration of the healthy (contralateral to side of seizure onset) hippocampus with other cortical regions was associated with better post-operative outcome. Surprisingly, despite the well-established involvement of the hippocampus in memory, they found that integration of the non-epileptogenic hippocampus with other neocortical regions was the most consistent predictor of cognitive outcome for all domains except for episodic memory. A somewhat complex pattern of findings characterized graph properties and verbal episodic memory, with increased integration and reduced centrality of the left inferior frontal cortex associated with better outcome for patients with left TLE. For patients with right TLE, the predictors of good verbal memory outcome were higher participation in the healthy left middle temporal cortex together with lower participation of the epileptogenic middle temporal cortex. The finding that no metrics involving either hippocampus was critical to predicting episodic memory outcome is somewhat at odds with our findings using simple correlation of DMN nodes [52]. These preliminary findings indicate that there is a substantial work ahead to determine the best methods for characterizing functionally relevant patterns of connectivity in mTLE and other clinical contexts.

Caveats and Limitations

While there is considerable promise to the clinical utility of rsfMRI, there are several limitations to the technique and concerns that need to be addressed. First, there are multiple analytic approaches that can be taken, and researchers and clinicians must decide which is most appropriate for their needs. A seed-based approach is easy to interpret but requires a priori selection of seed regions from the literature that will capture critical networks and may lead one astray if there has been substantial reorganization in the patient population of interest. For example, the selection of a damaged hippocampus as seed will likely not lead to identifying the effective memory network in mTLE. ICA approaches are data driven and will extract networks without a priori definition of seeds but can sometimes split networks and require selection of components of interest following analysis. A semiautomated approach has been proposed for identification of these networks based on training templates [91], and machine learning techniques can enable identification of atypical networks as long as there is sufficient homogeneity within the patient population of interest. Finally, issues of thresholding that are inherent to task-based fMRI can potentially be an issue in rsFC analyses. Arbitrary cut-offs such as p < 0.05 have the potential to eliminate meaningful voxels in single-subject analysis (e.g. fail to incorporate a region of meaningful connectivity into a language map), but without any form of thresholding, the signal will be noisy and uninterpretable. Examination of connectivity between two regions of interest in order to predict postsurgical change may avoid this but is again heavily reliant on proper ROI selection.

In addition to the techniques for characterizing the rsFC networks, one needs to be concerned about reliability of these measures, particularly at the individual subject level. Certainly, concerns have been raised about test-retest reliability in activation paradigms, particularly for complex tasks that can induce different strategies [132, 133]. There is evidence that rsFC is more reliable than activation magnitude or extent during a task [134, 135], but it is important to note that variables such as participant age, scan length, and other acquisition parameters can have modest influences [136, 137]. In addition, the extent to which the pathology in mTLE or MCI/AD or use of medications in these conditions may influence (either globally or locally) the neurovascular coupling that is crucial for observing the BOLD effect is just beginning to be investigated [138,139,140,141]. A recent review of limitations in clinical application can be found in [142]. A final crucial step in advancing the clinical utility of rsfMRI, particularly for the individual patient, is establishing sensitivity, specificity, and positive predictive value for these metrics. Fortunately the barrier to conducting resting state studies is considerably lower than with activation tasks in terms of amount of data required, the complexity of task influences, and the possibility of pooling data from multiple centres, making it more likely that we will make rapid progress on some of these issues.

Conclusions

Overall, there is very good evidence that rsfMRI is developing into a useful clinical tool for mapping language networks and characterizing functional integrity of memory networks in clinical populations and for providing predictions following surgical or other interventions. At present, however, considerable work still needs to be done in terms of comparisons to relevant ‘gold standards’ in clinical practice, determining ideal analytic strategies and decision algorithms, and relating network properties to the behaviours we are most interested in measuring and predicting.