Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

1.1 Background

Localization of function within the brain using functional magnetic resonance imaging (fMRI) traditionally has been performed by presenting stimuli or imposing tasks (such as finger tapping or object naming) to elicit neuronal responses (Posner and Raichle 1994; Spitzer et al. 1995). This type of experiment has been very effective, as evidenced by over 28,000 publications (listed in PubMed, as of December, 2013) utilizing task-based fMRI. fMRI detects changes in the blood oxygen level–dependent (BOLD) signal that reflect the neurovascular response to neural activity. Thus, BOLD fMRI is able to identify regions in the brain associated with a given task.

Since the earliest days of fMRI, it has been recognized that the BOLD signal exhibits spontaneous fluctuations (Purdon and Weisskoff 1998). These fluctuations were initially regarded as noise to be averaged out over many trials or task blocks (Triantafyllou et al. 2005). More recent studies have shown that these spontaneous fluctuations reflect the brain’s functional organization. The human brain is a disproportionate consumer of metabolic energy relative to its weight: 20 % of the total energy utilization but only 2 % of body weight (Clarke and Sokoloff 1999). This energy appears to be largely used for signaling (Shulman et al. 2004; Attwell and Laughlin 2001; Ames et al. 1992; Lennie 2003; Raichle and Mintun 2006). Task performance only minimally increases energy consumption in the brain (Raichle and Mintun 2006). Therefore, task-based experiments ignore the overwhelming preponderance of the brain’s activity. That intrinsic brain activity could be utilized for functional localization was first suggested by Biswal and colleagues who demonstrated that BOLD fluctuations observed in the resting state are correlated within the somatomotor system (Biswal et al. 1995). Correlated intrinsic activity currently is referred to as functional connectivity MRI or resting-state fMRI (rsfMRI). The development of these methods has opened up many exciting possibilities for future neurocognitive research as well as clinical applications, including presurgical planning, which is the subject of this chapter. A historical review is given in (Snyder and Raichle 2012).

1.2 Resting-State Networks

The topographies of functionally connected regions across the brain are known as resting-state networks (RSNs; equivalently intrinsic connectivity networks (Seeley et al. 2007). The resting-state fMRI scans are generally acquired while the subject is in a state of quiet wakefulness (Fox and Raichle 2007). The importance of RSNs lies in the fact that their topography closely corresponds to responses elicited by a wide variety of sensory, motor, and cognitive tasks (Smith et al. 2009). Intrinsic activity persists, albeit in somewhat modified form, during sleep (Samann et al. 2010; Larson-Prior et al. 2009) or even under sedation (Mhuircheartaigh et al. 2010). The persistence of the spontaneous fluctuations during states of reduced awareness suggests that intrinsic neuronal activity plays a role in the maintenance of the brain’s functional integrity (Pizoli et al. 2011). Spontaneous BOLD activity has been detected in all mammalian species investigated thus far (Hutchison et al. 2012; Schwarz et al. 2013; Nasrallah et al. 2013), which reinforces the notion that this phenomenon is physiologically important. However, the precise physiological functions of intrinsic activity remain to be elucidated.

Perhaps the most fundamental RSN is the default mode network (DMN) (Fig. 1a), first identified by a meta-analysis of task-based functional neuroimaging experiments performed with positron emission tomography (PET) (Shulman et al. 1997; Gusnard and Raichle 2001). The defining property of the DMN is that it is more active at rest than during performance of goal-directed tasks. The DMN was first identified using rsfMRI by Greicius et al. (2003). This finding has since been replicated many times, using a variety of analysis methods (Smith et al. 2009; Beckmann et al. 2005; De Luca et al. 2006; Power et al. 2011; Yeo et al. 2011; Damoiseaux et al. 2006; van den Heuvel et al. 2008; Lee et al. 2012a). Some investigators have hypothesized that there are two large anti-correlated systems in the brain (Fox et al. 2005; Golland et al. 2008), one anchored by the DMN and the other comprised of systems controlling executive and attentional mechanisms. This dichotomy has been variously referred to as “task-positive” vs. “task-negative” (Power et al. 2011; Lee et al. 2012a; Fox et al. 2005; Chai et al. 2012; Zhang et al. 2011a) and “intrinsic” vs. “extrinsic” (Golland et al. 2008; Doucet et al. 2011). Although the nomenclature associated with the DMN remains controversial (Jack et al. 2012; Spreng 2012), the topography of the DMN is remarkably consistent across diverse analysis strategies.

Fig. 1
figure 1

Surface plots of resting-state networks as derived from fuzzy c-means algorithm (Lee et al. 2012a) (a) Default mode network. (b) Somatomotor network. (c) Visual network. (d) Language network. (e) Dorsal attention network. (f) Ventral attention network. (g) Frontoparietal control network

Full size image

Primary sensory and motor RSNs include the somatomotor network (SMN), first identified by Biswal and colleagues (Biswal et al. 1995), which encompasses primary and higher order motor and sensory areas (Fig. 1b). The visual (VIS) network spans much of the occipital cortex (Fig. 1c) (Smith et al. 2009; Beckmann et al. 2005; De Luca et al. 2006; Power et al. 2011; Yeo et al. 2011). The auditory network includes Heschl’s gyrus, the superior temporal gyrus, and the posterior insula (Smith et al. 2009). The language network (LAN) includes Broca’s and Wernicke’s areas but also extends to prefrontal, temporal, parietal, and subcortical regions (Fig. 1d) (Tomasi and Volkow 2012; Lee et al. 2012b; Hacker et al. 2013).

RSNs involved in attentional and cognitive control include the dorsal attention network (DAN) and the ventral attention network (VAN) (Seeley et al. 2007; Power et al. 2011; Yeo et al. 2011; Corbetta and Shulman 2002; Fox et al. 2006a). The DAN (Fig. 1e) includes the intraparietal sulcus and the frontal eye field and is recruited by tasks requiring control of spatial attention. The VAN (Fig. 1f), which includes the temporal-parietal junction and ventral frontal cortex, is involved in the detection of environmentally salient events (Corbetta and Shulman 2002; Fox et al. 2006a; Astafiev et al. 2006). The frontoparietal control network (FPC) (Fig. 1g), which includes the lateral prefrontal cortex and the inferior parietal lobule, is associated with working memory and control of goal-directed behavior (Vincent et al. 2008; Power and Petersen 2013). Finally, the cingulo-opercular network (CON), also known as the salience network (Seeley et al. 2007) or the core control network (Dosenbach et al. 2006), includes the medial superior frontal cortex, anterior insula, and anterior prefrontal cortex. The CON is thought to enable the performance of tasks requiring executive control (Power et al. 2011; Power and Petersen 2013; Dosenbach et al. 2006).

2 Methods

2.1 Overview of Processing Methods

Resting-state fMRI methodology is currently dominated by two complementary strategies, spatial independent components analysis (SICA) (Beckmann et al. 2005) and seed-based correlation mapping (Biswal et al. 1995). Both strategies depend on the fact that spontaneous neural activity is correlated (coherent) within widely distributed regions of the brain. Both strategies yield highly reproducible results at group level (Damoiseaux et al. 2006; Shehzad et al. 2009). SICA decomposes rsfMRI data into a sum of single components, each component corresponding to a spatial topography and a time course. In contrast, seed-based correlation mapping is computed by voxelwise evaluation of the Pearson correlation between the time courses in a targeted region of interest and all other voxels in the brain (Fox et al. 2009). The principal advantage of SICA is that it provides a direct means of separating artifact from BOLD signals of neural origin, although this separation typically requires observer expertise. The results obtained using SICA may vary substantially depending on processing parameters (e.g., number of requested components). Thus, SICA can be difficult to use in the investigation of targeted RSNs, especially in single subjects. In contrast, targeting of selected RSNs is intrinsic to seed-based correlation mapping. However, the principal difficulty in using seed-based correlation mapping is the exclusion of non-neural artifacts, which is typically accomplished using regression techniques (Fox et al. 2009; Jo et al. 2010; Vincent et al. 2006).

SICA and seed-based correlation mapping both represent strategies for assigning RSN identities to brain voxels. Since SICA makes no a priori assumptions regarding the topography of the obtained components, this method exemplifies unsupervised classification. In contrast, seed-based correlation mapping depends on prior knowledge and so exemplifies supervised classification. For additional discussion of the distinction between supervised vs. unsupervised methodologies, see (Hacker et al. 2013). Below, we present results obtained by two unsupervised methods, SICA and c-means clustering, and two supervised methods, conventional seed-based correlation mapping and RSN mapping using a trained multilayer perception (MLP) classifier.

2.2 General Preprocessing

Preprocessing of fMRI data varies across laboratories. The following describes the procedures used in our laboratory (Shulman et al. 2010). Briefly, these include compensation for slice-dependent time shifts, elimination of systematic odd-even slice intensity differences due to interleaved acquisition, and rigid body correction for head movement within and across runs. The fMRI data are intensity scaled (one multiplicative factor applied to all voxels of all frames within each run) to obtain a mode value of 1,000 (Ojemann et al. 1997). This scaling facilitates assessment of voxelwise variance for purposes of quality assessment but does not affect computed correlations. Atlas transformation is achieved by composition of affine transforms connecting the fMRI volumes with the T1- and T2-weighed structural images. Head movement correction is included in a single resampling to generate a volumetric time series in 3 mm3 atlas space.

2.3 Preprocessing in Preparation for Seed-Based Correlation Mapping

Additional preprocessing in preparation for seed-based correlation mapping includes the following: (1) spatial smoothing (6 mm full-width half- maximum Gaussian blur in each direction), (2) voxelwise removal of linear trends over each run, (3) temporal low-pass filtering to retain frequencies <0.1 Hz, and (4) reduction of spurious variance by regression of nuisance waveforms derived from head motion correction and extraction of the time series from regions of noninterest in white matter and cerebrospinal fluid. In our lab, this step includes regression of the global signal averaged over the whole brain (Fox et al. 2009; Buckner et al. 2005). A consequence of global signal regression (GSR) is that all subsequently computed correlations are effectively partial correlations of first-order controlling for widely shared variance. As global signal regression currently is a contentious maneuver, this topic is considered further in the next section.

2.4 Global Signal Regression

Global signal regression (GSR) prior to correlation mapping is a highly effective means of reducing widely shared variance and thereby improving the spatial specificity of computed maps (Fox et al. 2009; Aguirre et al. 1998; Macey et al. 2004). Some part of the global signal undoubtedly is of neural origin (Scholvinck et al. 2010). However, much (typically most) of the global signal represents non-neural artifact attributable to physical effects of head motion (Friston et al. 1996; Yan et al. 2013; Power et al. 2012; Satterthwaite et al. 2012) and variations in the partial pressure of arterial carbon dioxide (Wise et al. 2004). In the absence of GSR, all parts of the brain appear to be strongly positively correlated (Chai et al. 2012; Fox et al. 2006a; Lowe et al. 1998; Joel et al. 2011). GSR causes all subsequently computed correlation maps to be approximately zero-centered; in other words, positive and negative values are approximately balanced over the whole brain (Fox et al. 2009). Thus, GSR unambiguously does negatively bias all computed correlations, although iso-correlation contours, i.e., map topographies, remain unchanged. This negative bias has caused some experts in the field research to criticize GSR on the grounds that it induces artifactual anti-correlations (Anderson et al. 2011; Murphy et al. 2009). This objection to GSR has largely been dissipated following the demonstration that some parts of the brain appear to be truly anticorrelated in the resting state, as demonstrated using SICA (Liao et al. 2010; Zuo et al. 2010). More recent objections to GSR focus on the possibility that it can distort quantitative functional connectivity differences across diagnostic groups (Saad et al. 2012). This objection to GSR, however, is irrelevant in the context of using rsfMRI for purposes of RSN mapping in individuals.

2.5 Seed-Based Correlation Mapping

Seed-based correlation mapping is one of the most widely adopted techniques for studying co-fluctuations in intrinsic neuronal activity, or functional connectivity (Shehzad et al. 2009; Cordes et al. 2000). The high adoption rate of the seed-based approach is partly attributable to the simplicity of its implementation and to the ease with which the results can be interpreted. Biswal et al. used this method to first demonstrate the feasibility of using fMRI to detect spatially distributed networks (Biswal et al. 1995).

Pearson product-moment correlation is the most widely used measure of functional connectivity (Biswal et al. 1995; Greicius et al. 2003; Fox et al. 2005; Lowe et al. 1998; Cordes et al. 2000; Xiong et al. 1999). Seed-based analyses require prior knowledge of the locations of regions of interest, and these can be obtained from previously determined atlas coordinates or from task-based fMRI data. For instance, a simple motor paradigm may be used to generate data involving the motor network. The activation data is then analyzed, and the voxel that is associated with the strongest activation is used as a “seed” region to study the resting-state data. Once the coordinates of the seed region have been identified, the resting-state time courses from the rest of the brain are compared with this region, and a correlation map is generated. An example of multiple RSNs using the seed-based approach is presented in Fig. 2 (Zhang and Raichle 2010).

Fig. 2
figure 2

Examples of multiple resting-state networks generated using a seed-based approach (blue circles in the figure) (Zhang and Raichle 2010). Six of the major networks are illustrated: visual, sensorimotor, auditory, default mode, dorsal attention, and frontoparietal executive control. The scale numbered 0–7 indicates the relative correlation strength

Full size image

2.6 Independent Component Analysis

Unsupervised data-driven approaches are of interest to researchers looking to analyze resting-state data without a priori assumptions. SICA is the most widely used data-driven approach to analyze resting-state data (Goldman et al. 2003; Beckmann and Smith 2004; Greicius et al. 2004). SICA decomposes resting-state fMRI data (time×space) into spatial components that are maximally independent. Each spatial component is associated with a particular time course. The components are useful for differentiating noise data from physiological data as well as identifying statistically independent systems. Comparison studies between seed-based correlation maps and spatial patterns determined by SICA have found similar spatial patterns (Beckmann et al. 2005; Rosazza et al. 2012).

Although the SICA approach eliminates the need for a priori seed identification, the user is required to choose the initial number of components as well as to select which components represent noise and which represent functional networks. Some studies have aimed to automate this process and use SICA as a method for identifying and eliminate noise within the BOLD signal (Starck et al. 2010; Thomas et al. 2002; Tohka et al. 2008).

2.7 Clustering Algorithms

Another method used to analyze rsfMRI data makes use of clustering algorithms. Clustering algorithms attempt to group items that are alike on the basis of a set of relevant characteristics to the problem of interest. Voxels can be grouped on the basis of similarity of their BOLD time courses by using some distance metric, such as a Pearson correlation. One example of a clustering algorithm is hierarchical clustering (Salvador et al. 2005; Cordes et al. 2002), which builds a dendrogram (a treelike structure) of all members. Other examples of clustering algorithms are the K-means (Golland et al. 2008) and Fuzzy c-means (Lee et al. 2012a) clustering algorithms. In these algorithms, all voxels are assigned membership to one or more of several clusters on the basis of their distances from the cluster centers, which, in turn, are calculated from an average of their members. Clustering algorithms iteratively update memberships and cluster centers until convergence is achieved (Lee et al. 2012a) (Fig. 1). Other variations on clustering algorithms include spectral-based clustering (Bellec et al. 2010) and graph-based clustering (van den Heuvel et al. 2008).

2.8 RSN Mapping Using a Trained Multilayer Perceptron (MLP)

One technique for mapping the topography of known RSNs in individuals uses a Multilayer perceptron (MLP) (Hacker et al. 2013). Perceptrons are machine learning algorithms that can be trained to associate arbitrary input patterns with discrete output labels (Rumelhart et al. 1986). For example, perceptrons can be trained to read handwritten digits, e.g., zip codes on addressed letters (LeCun et al. 1989). Here, an MLP was trained to associate seed-based correlation maps with particular RSNs. Running the trained MLP on correlation maps corresponding to all voxels in the brain generates voxel-wise RSN membership estimates. Thus, RSN mapping using a trained MLP exemplifies supervised classification. An example of the RSN produced by the MLP algorithm in three subjects is presented in Fig. 3. It is critical to note that our MLP assigns RSN membership to rsfMRI correlation maps. This application of machine learning must not be confused with other methodologies in which classifiers have been trained to assign diagnostic labels to patients on the basis of their resting-state functional connectivity patterns (Abdulkadir et al. 2011). Our implementation of MLP-based RSN mapping utilizes the same preprocessing steps described above in connection with seed-based correlation mapping and fuzzy c-means clustering (Sects. 2.2, 2.3, and 2.4).

Fig. 3
figure 3

Single subject, voxel estimation of resting-state networks using trained multilayer perceptron (MLP) in three subjects. The results are from best (a), median (b), and worst (c) performers as determined by RMS error and demonstrate high quality results in individual subjects even in the worst case. MLP output was converted to a percentile scale and sampled onto each subject’s cortical surface (Hacker et al. 2013)

Full size image

Figure 4 demonstrates the degree to which the MLP captures individual variability, by showing that, in each subject, the location of the central sulcus in the cortical surface segmented using FreeSurfer (Fischl 2012) is highly correlated with the location of the SMN centroid calculated by the MLP. Detailed quantitative evaluation of the MLP performance is given in (Hacker et al. 2013). MLP performance was also compared to alternative RSN estimation schemes such as dual regression and linear discriminant analysis and was found to provide improved “area under the curve estimation” with better orthogonal estimates of RSN membership.

Fig. 4
figure 4

Multilayer perceptron (MLP) somatomotor network (SMN) validation results derived from five individuals selected to represent the correspondence between SMN variability and anatomical variability in the central sulcus. The plot shows the correlation between the Talairach y-coordinate of the centroid of the MLP SMN and the y-coordinate centroid of the central sulcus traced over the anatomy (as determined by the FreeSurfer program) in a large validation dataset (Hacker et al. 2013)

Full size image

In summary, the MLP accurately generates RSN topography estimates in individuals consistent with previous studies, even in brain regions not represented in the training data, and can be used for generating individual RSN maps. These findings are important to future applications because they demonstrate that this approach can reliably and effectively map multiple RSNs across individual subjects.

3 Application to Presurgical Planning

3.1 Introduction

Multiple studies have demonstrated that maximal resection of a brain tumor while sparing nearby eloquent cortex leads to improved outcomes, with minimal morbidity (Keles et al. 2001, 2006; Lacroix et al. 2001; McGirt et al. 2009; Sanai et al. 2008). Historically, neurosurgeons have been concerned with the localization of the motor and language system on the assumption that these parts of the brain (“eloquent” cortex) instantiate critical functionality. However, a broader understanding of brain function suggests that all parts of the brain contribute to important functionality (Yeo et al. 2011; Lee et al. 2012a; Golland et al. 2008; Hacker et al. 2013). Thus, improved functional mapping of multiple RSN beyond motor and language could lead to further improvements in patient outcomes.

Multiple prior publications have explored the use of rsfMRI in presurgical planning. An early case report example of this technique was used to localize the motor cortex in a patient with a brain tumor (Shimony et al. 2009). Kokkonen et al. (2009) similarly compared motor task data to resting-state data and showed that the motor functional network could be localized on the basis of resting-state data in 8 tumor patients, as well as 10 healthy control subjects.

Resting-state fMRI may also be used for presurgical planning in patients with epilepsy. The higher spatial resolution offered by rsfMRI over electroencephalography could provide a distinct advantage in mapping epileptic foci or networks. Seed-based methods were used by Liu et al. (2009) to successfully locate sensorimotor areas by using rsfMRI in patients with tumors or epileptic foci close to sensorimotor areas. They found agreement between rsfMRI, task-based fMRI, and intraoperative cortical stimulation data. In another study from the same laboratory, Stufflebeam and colleagues (Stufflebeam et al. 2011) were able to localize areas of increased functional connectivity in five of six patients that overlapped with epileptogenic areas identified by invasive encephalography. Zhang et al. (2011b) used graph methods and a pattern classifier applied to rsfMRI data to identify subjects as either having medial temporal lobe epilepsy or as normal controls. Using data from 16 patients with intractable medial temporal lobe epilepsy and 52 normal controls, they achieved an average sensitivity of 77.2 % and a specificity of 83.86 % in classification. Bettus et al. (2010) reported that increases in basal functional connectivity were a specific marker of the location of the epileptogenic zone in 22 patients with mesial temporal lobe epilepsy. Weaver et al. (Weaver et al. 2013) studied four patients with focal epilepsy along with 16 control subjects to determine whether the seizure focus could be found using the functional patterns near the epileptogenic zone. By averaging voxel homogeneity across regions of interest in comparison to other regions, they were able to accurately identify the epileptic focus. Tie et al. (2013) employed ICA data evaluation on a training group of 14 healthy subjects to identify the language network from rsfMRI data. The result of that analysis was then used to identify the language network in a second group of 18 healthy subjects on an individual level. The authors further propose an automated system for determining the language network in individual patients using SICA. A more detailed presentation of our experience using rsfMRI for presurgical mapping with both seed-based and MLP approaches is given in the next sections.

3.2 Preoperative Sensorimotor Mapping in Brain Tumor Patients Using Seed-Based Approach

Zhang and colleagues (2009) described our initial experience with rsfMRI brain mapping for presurgical planning of brain tumor resections in four patients. All tumors were in critical spatial relationship to the primary motor and somatosensory cortices, thus necessitating accurate localization prior to surgery to minimize postoperative deficits. Each of the patients was scanned using rsfMRI and using task-based fMRI while performing a block design finger-tapping task. fMRI in each patient included four 7-min runs (28 min total). The rsfMRI data previously acquired from a group of normal controls (N = 17) were also used for comparison. Data processing for both resting-state and task-based fMRI was standardized (Zhang et al. 2008).

Preprocessing was done using the above-described methods. Correlation maps were generated using 6 mm radius spherical seed regions of interest (ROIs). The seed regions for all the normal brains were placed in the left sensorimotor cortex. In the tumor patients, the seed was placed in the hemisphere contralateral to the tumor at coordinates taken from an independent group of subjects that performed a button-press task (Fox et al. 2006b; Zacks et al. 2001). Electrocortical stimulation mapping (ECS) was performed on three of the four tumor patients, and these data, in additional to the task-based fMRI, were used for comparison with the resting-state data.

The 17 control brains were mapped using the seed region in the left sensorimotor cortex. The correlations of resting state activity to the seed region were recorded for each of the other voxels in the brain. The group average was used as a control to show the distribution of the sensorimotor network in a healthy brain. To confirm the reproducibility of this method in individual subjects, the full rsfMRI data set in each subject (28 min) was divided into the four separate scans (7 min) and a separate analysis was performed on each segment. The somatomotor cortex was consistently activated in the same region over the four scans in all control subjects. The four tumor patients were also mapped individually following the placement of the seed regions on the contralateral side of the brain. In the following paragraphs, we discuss two of the four patients from this paper:

Figure 5 shows a glioblastoma multiforme that was diagnosed in the right hemisphere of a 45-year-old man. ECS during surgery revealed that motor cortex was displaced anteriorly. Figure 5a shows the results of resting-state fMRI, which confirmed the anterior displacement of the motor cortex. Figure 5b shows the results of the task-based fMRI that also demonstrated an anterior shift in the motor cortex ipsilateral to the tumor. However, the task-based result was unreliable, as the second trial showed likely artifactual activation in the posterior part of the tumor. Resting-state fMRI was more consistent.

Fig. 5
figure 5

Comparison of resting state and task-related functional magnetic resonance imaging (fMRI) mapping in a 45-year-old with a diagnosis of glioblastoma multiforme (Case 1). (a) Resting-state correlation mapping × 2 trials shows correlated activity with a distribution resembling the activation from trial 1 but not trial 2 of the task. (b) Finger-tapping fMRI × 2 trials. Activity within the tumor (blue arrows) was seen in trial 2 but not in trial 1. All colored regions represent activations in the somatomotor network. All images are displayed left-on-left (neurologic convention) (Zhang et al. 2009)

Full size image

In another case, shown in Fig. 6, a 64-year-old man developed focal motor seizures secondary to a mass in the left hemisphere (Fig. 6a). Finger-tapping fMRI showed atypical response topography including activation in the right parietal cortex in addition to the expected activation of the somatomotor area (Fig. 6b). Seed-based (Fig. 6c) correlation mapping rsfMRI showed the somatomotor RSN without parietal involvement. Correlation mapping with a seed in the right parietal cortex matched the topography of the DAN (Fig. 6d). Our interpretation of this result is that, during task-based fMRI, the patient had to strongly focus his attention in order to complete the task, which accounts for the prominent activation in the attentional network. This case illustrates the potentially increased specificity of resting-state compared to task-based fMRI. The findings of the rsfMRI were consistent with the intraoperative ECS.

Fig. 6
figure 6

Magnetic resonance image (MRI) of a 64-year-old man who presented with focal motor seizures (Case 2). (a) Structural MRI revealed a tumor in the left parietal cortex that invades territory near the central sulcus (neurologic convention). The green circle represents the location of the ipsilateral hand response to cortical stimulation. (b) Task-related activity was seen bilaterally in the frontal lobe. In addition, a large band of activity appeared in the right parietal cortex, not consistent with the pattern of activity from the sensorimotor network. (c) Resting-state correlation mapping using a seed in the right (unaffected) hemisphere (blue circle) showed ipsilateral correlations anterior to the tumor as well as a region of activity in the midline of the parietal cortex. Note absence in the correlation mapping results of the parietal activity seen in the task-related map. (d) Parietal activation seen during the task-evoked scan is revealed to represent a separate resting-state network, the dorsal attention network, which is normally dissociated from the sensorimotor network (seed: blue circle). All images are displayed left-on-left (neurologic convention) (Zhang et al. 2009)

Full size image

These cases show that the seed-based correlation mapping using resting-state data is consistent with task-based fMRI but, in some cases, more reliable.

3.3 Preoperative Mapping of Functional Cortex Using MLP

Mitchell and colleagues reported the application of MLP-based RSN mapping to presurgical planning in six patients with intractable epilepsy and seven patients with brain tumors (Mitchell and Hacker 2013). Patients with epilepsy underwent electrocorticographic monitoring to localize the epileptogenic zone of seizure onset and to perform functional mapping with ECS. Patients with tumors underwent intraoperative ECS mapping prior to resection of the tumor mass.

For preoperative rsfMRI analysis in the tumor patients, lesions were manually segmented using T1- and T2-weighted images. The MLP was trained and applied de novo in each tumor patient, omitting tumor voxels. To determine the probability that an electrode covers a portion of a RSN, electrode MRI co-registration was used with the results of the MLP analysis, and gray matter voxels located within 30 mm of the electrode were averaged with a factor that was inversely proportional to the square of their distance from the electrode.

For the seizure-monitoring patients, electrodes were segmented on the basis of a CT image co-registered to the patient’s MRI using methods similar to those previously described (He et al. 2007; Hermes et al. 2010). Electrodes imaged in the post-grid implantation CT typically are displaced inward relative to the cortical surface imaged on preoperative MRI because of traction from dural over-sewing and postsurgical edema. This inward displacement was corrected by projecting electrode coordinates outwards to the brain surface along a path normal to the plane of the grid.

Electrodes were classified as over eloquent cortex using ECS mapping. Motor regions were defined by the presence of induced involuntary motor movements. Language sites were defined by speech arrest during the stimulation.

For comparison of MLP-based RSN mapping to ECS mapping, an electrode was classified as positive or negative in the MLP results according to the probability of it belonging to the appropriate RSN (motor or language). These probabilities were then plotted against the ECS results to generate receiver–operator characteristic (ROC) curves. These ROC curves were averaged, and the area under the averaged curve (AUC) was used as a measure of the agreement between the MLP and ECS methods.

The resulting RSN maps for the tumor patients are shown in Fig. 7. Structural images in column 1 reveal the extent of the lesion in each patient. Columns 2, 3, and 4 show axial, coronal, and sagittal views of the MLP results displayed in a winner-take-all format. Significant asymmetry across the midline can be seen in networks near the tumor, which is consistent with previous findings (Zhang et al. 2009). Networks were preserved in the presence of a tumor, though they were often shifted with respect to their normal anatomic position. Despite these distortions, there was good agreement between the ECS and MLP results.

Fig. 7
figure 7

Resting-state network (RSN) maps produced by the multilayer perceptron (MLP) for seven tumor patients. The seven networks, LAN language, SMN somatomotor network, VIS visual, DAN dorsal attention network, VAN ventral attention network, FPC frontoparietal control, and DMN default mode network, were mapped in the area of the tumor using the winner-take-all format (Mitchell and Hacker 2013)

Full size image

Figure 8 demonstrates a high degree of qualitative overlap between the location of the motor and language networks as compared to the ECS results in patients with epilepsy. The positive motor ECS electrodes were centered in the precentral gyrus. The MLP-mapped motor areas encompassed both the pre- and postcentral gyri. The positive language ECS electrodes were centered in the pars opercularis of the inferior frontal gyrus (IFG), approximately in the Brodmann area (BA) 44. The MLP language-positive regions were in the pars triangularis of the IFG, which corresponds to BA 45. The anteriorly shifted MLP-based localization of the language cortex (BA 45 vs. 44) suggests the possibility that the definition of eloquent cortex should be expanded. Quantitative comparisons were performed with an ROC analysis which yielded an AUC of 0.89 for the motor network and an average AUC of 0.76 for the language network. These findings demonstrate that MLP-based mapping can identify RSNs in the presence of distorted anatomy.

Fig. 8
figure 8

Comparison of electrocortical stimulation (ECS) and multilayer perceptron (MLP) results for the motor and language cortices in six patients with epilepsy. Colored triangles are ECS positive and black circles are ECS negative. In the left column, the high ECS sensitivity method was employed to classify motor electrodes as ECS positive (red triangles) and compared to the MLP results (light blue). In the middle column, the high ECS specificity method was employed to classify motor electrodes. In the right column, the high ECS sensitive method was used to classify language electrodes as ECS positive (green triangles), with the MLP results displayed in orange (Mitchell and Hacker 2013)

Full size image

Loci in MLP maps outside the appropriate RSN, but eloquent as determined by ECS, are defined as MLP false negatives. Minimization of MLP false negatives is critical to reduce surgical morbidity, since resection of a false negative area could lead to a clinical deficit. Figure 9 illustrates the results of an analysis undertaken to minimize MLP motor false negatives. This analysis showed that the probability of a MLP false negative could be reduced to less than 2 % by expanding the “no-cut” zone by 15 mm around the contour corresponding to an 85 % likelihood of belonging to the motor RSN.

Fig. 9
figure 9

The method employed to define a “no-cut” area in patients with epilepsy, in which the probability of damage to motor cortex is substantial. (a) To define the area, several multilayer perceptron (MLP) thresholds (70th, 75th, 80th, 85th percentiles) were used to classify electrodes as covering motor cortex, and the “no-cut” zone was expanded around each of the motor electrodes. The probability of a missed motor electrode, which could result in motor deficits, was plotted against the radius of expansion. (b) Visualization of the method performed at the 85 % and at a radius of expansion of 15 mm. Red triangles mark the motor cortex as determined by ECS that were missed by the MLP method (Mitchell and Hacker 2013)

Full size image

In summary, MLP-based RSN mapping robustly identified all networks in all patients, including those with distorted anatomy attributable to mass effect. When the ECS positive sites were analyzed, rsfMRI had AUCs of 0.89 and 0.76 for motor and language, respectively. MLP false negatives were minimized by including a 15 mm safety margin around the edge of the motor RSN. These findings demonstrate that MLP-defined RSNs are able to identify eloquent cortices.

4 Conclusion

This chapter presents an introduction to rsfMRI and RSNs. We briefly covered RSN imaging methods and several common analysis techniques. We presented our early experience using this technique for the localization of the motor cortex using seed-based correlation mapping. We then presented MLP-based RSN mapping in detail, as this is our current method of choice for simultaneously mapping multiple RSNs to provide the surgeon with an accurate map of the resting state architecture of the brain prior to surgery. Finally, we reviewed our experience using the MLP-based technique in patients with epilepsy and brain tumors. This experience suggests how MLP-based RSN mapping should be applied to minimize surgical morbidity.

As these results demonstrate, rsfMRI is a promising technique in the context of presurgical planning with the objective of decreasing morbidity while maximizing complete resection of pathological tissue. However, the methodology is still in its early stages of development. Further research is necessary to make these tools more accurate and available in the operating room. Additional research is needed to explore the differences between rsfMRI and ECS mapping and to better understand the consequences of disrupted RSNs other than the motor and language systems. Related engineering development should incorporate the presurgical MRI results into intraoperative neuronavigation systems, including the rsfMRI results in conjunction with white matter fiber bundle anatomy derived from diffusion imaging. An additional possibility is real-time intraoperative rsfMRI during surgery using MRI systems that currently are becoming more common.