REVIEW ARTICLE

A few thoughts on brain ROIs

Tianming Liu

Published online: 10 May 2011# Springer Science+Business Media, LLC 2011

Abstract Quantitative mapping of structural and functionalconnectivities in the human brain via non-invasive neuro-imaging offers an exciting and unique opportunity tounderstand brain architecture. Because connectivity alter-ations are widely reported in a variety of brain diseases,assessment of structural and functional connectivities hasemerged as a fundamental research area in clinicalneuroscience. A fundamental question arises when attempt-ing to map structural and functional connectivities: how todefine and localize the best possible Regions of Interests(ROIs) for brain connectivity mapping? Essentially, whenmapping brain connectivities, ROIs provide the structuralsubstrates for measuring connectivities within individualbrains and for pooling data across populations. Thus,identification of reliable, reproducible and accurate ROIs iscritically important for the success of brain connectivitymapping. This paper discusses several major challenges indefining optimal brain ROIs from our perspective andpresents a few thoughts on how to deal with those challengesbased on recent research work done in our group.

Keywords Region of interests . Diffusion tensor imaging,functional MRI . Brain connectivity

Introduction

Segregation and integration is a general principle of thebrain’s functional architecture (Friston 2009; Ashburner etal. 2004). It is widely believed that the brain’s function is

integrated via structural and functional connectivities(Biswal 2010; Sporns et al. 2005; Van Dijk et al. 2010;Hagmann et al. 2010; Friston et al. 2003). Therefore,analysis of brain connectivity is of significant importance tobrain imaging and neuroscience (Biswal 2010; Sporns et al.2005; Van Dijk et al. 2010; Hagmann et al. 2010; Friston etal. 2003; Bullmore and Sporns 2009). Essentially, whenmapping brain connectivity, Regions of Interests (ROIs)provide the structural substrates for measuring connectivi-ties within individual brains and for pooling data acrosspopulations. Identification of reliable, reproducible andaccurate ROIs is critically important for the success ofbrain connectivity mapping. Therefore, a fundamentalquestion arises when attempting to measure brain connec-tivities: how to define and localize the best possible ROIs?Current approaches for identifying ROIs in brain imagingcan be broadly classified into four categories (Li et al.2009a). The first is manual labeling by experts based ontheir domain knowledge (Sobel et al. 1993). While widelyused, this method is vulnerable to inter-subject and intra-subject variation and its reproducibility may be low. Thesecond method is to cluster ROIs from the brain imageitself and is data-driven (Zang et al. 2004; Hyvärinen andOja 2000; Beckmann et al. 2005; Calhoun et al. 2004).However, these data-driven approaches are typically sensi-tive to the clustering parameters used, and their neurosci-ence interpretation is not clear. The third one is to predefineROIs in a template brain, and warp them to the individualspace using image registration algorithms (Shen andDavatzikos 2002). The accuracy of these atlas-basedwarping methods is limited due to the remarkable variabil-ity of neuroanatomy across different brains. The fourthmethod uses task-based fMRI paradigms to identifyactivated brain regions as ROIs. This methodology isregarded as the benchmark approach for ROI identification.

T. Liu (*)Department of Computer Science and Bioimaging ResearchCenter, The University of Georgia,Athens, GA 30602, USAe-mail: tliu@cs.uga.edu

Brain Imaging and Behavior (2011) 5:189–202DOI 10.1007/s11682-011-9123-6

However, task-based fMRI itself has limitations, and issubject to a variety of variables that might affect theaccuracy of detected ROIs (Jo et al. 2008; Geissler et al.2005; White et al. 2001). For instance, a few recent studies(Jo et al. 2008; Geissler et al. 2005; White et al. 2001)reported that the locations of detected fMRI activationscould be significantly shifted due to spatial smoothing,which is commonly used in popular fMRI analysis toolkitssuch as FSL (http://fsl.fmrib.ox.ac.uk/fsl/feat5/), SPM(http://www.fil.ion.ucl.ac.uk/spm/) and AFNI (http://afni.nimh.nih.gov/afni) in order to increase signal-to-noise ratioand facilitate aggregating data across subjects. As shown inFig. 1a, because the local maximum of the ROI was shiftedby 4 mm due to spatial smoothing (Li et al. 2010a), itsstructural profile (Fig. 1b) was significantly altered. Inshort, existing ROI identification approaches, includingtask-based fMRI, are inadequate to accurately localize ROIs(Li et al. 2010a).

Why identification of brain ROIs is so challenging?

So far, ROI identification is still an open, urgent problem inmany brain imaging applications including brain connec-tivity mapping. From our perspective, the major challengescome from three critical reasons: 1) The boundaries

between cortical regions are unclear (Van Essen and Dierker2007; Handbook of Functional Neuroimaging of Cognition);2) The individual variability of cortical anatomy, connectionand function is remarkable (Brett et al. 2002); 3) Theproperties of ROIs are highly nonlinear. For instance, a slightchange of the size or location of a ROI might dramaticallyalter its structural and functional connectivity profiles(Fig. 1b) (Li et al. 2010a).

Boundaries between cortical regions

In vivo parcellation of the cerebral cortex from brainimaging data has been extensively studied in the past twodecades for definition of brain ROIs at the scale of gyral orsulcal regions. Current approaches to cortical parcellationcan be broadly classified into two categories: model-drivenor data-driven methods. In model-driven methods, the atlas-based warping method is widely used. For instance, theFreeSurfer system (Fischl et al. 1999) and the HAMMERalgorithm (Shen et al. 2002) can be used to parcellate asubject’s cortex by transforming an expert-labeled atlas tothe subject’s space. In data-driven approaches, people haveused geometrical (Fischl et al. 1999; Shen et al. 2002; Li etal. 2009b), morphological (Van Essen and Dierker 2007),connectional (Behrens et al. 2003), or functional (Cohen etal. 2008) features to guide the parcellation procedure. Anessential issue in cortical parcellation approaches is whatfeatures or attributes should be used to define the corticalregion boundaries, no matter if it is model-driven or data-driven. For instance, cortical geometric folding patternsmight be good for sulcal or gyral parcellation (Li et al.2009b, 2010d). Figure 2 demonstrates 6 examples of gyralparcellation via folding pattern based segmentation andatlas guided recognition (Li et al. 2010d). It is even possibleto further parcellate a sulcus into finer-granularity sulcalbanks (Li et al. 2010b) based on geometric shape features,as shown in Fig. 3.

However, it is unclear how the boundaries defined bygeometric shape features, e.g., in Figs. 2 and 3, correlatewith cytoarchitectural or functional boundaries. In addition,the brain ROIs defined at the scale of gyral or sulcal regionsmight contain multiple functionally inhomogeneous sub-units. Hence, definition of fine-granularity brain ROIswithin functionally meaningful and homogeneous regionsis much desired (Zhu et al. 2011). Recently, we attemptedto develop a novel hybrid feature based on fiber shape andconnectivity patterns for fine granularity gyrus parcellation(Zhu et al. 2011). The underlying premise is that eachbrain cytoarchitectonic area has a unique set of extrinsicinputs and outputs, called the ‘connectional fingerprint’(Passingham et al. 2002), which is crucial in determiningthe function that each brain area performs (Passingham et

Fig. 1 a Local activation map maxima (marked by the cross) shift ofone ROI due to spatial volumetric smoothing in a working memorytask (Faraco et al. 2011). The top one was detected using unsmootheddata while the bottom one used smoothed data (FWHM: 6.87 mm). bThe corresponding fibers for the ROIs in (a). The ROIs are presentedusing a sphere (radius: 5 mm)

190 Brain Imaging and Behavior (2011) 5:189–202

al. 2002). We formulated the gyrus parcellation as avertices clustering problem, as illustrated in Fig. 4. Bothfeature similarity and vertex adjacency are used to definethe distance between vertices during the clustering(Passingham et al. 2002). Our results show that theprecentral gyrus can be reasonably parcellated into 3 similarbroad areas (Fig. 4i) on both hemispheres across differentsubjects, achieving finer granularity of regions than thetraditional Brodmann areas. The parcellation results werepartly verified via motor task-based fMRI studies(Fig. 4j) (Passingham et al. 2002).

However, it should be noted that the data-drivenclustering approach in (Passingham et al. 2002) is stillsensitive to the clustering parameters used. Though theaffinity propagation clustering approach (Frey and Dueck2007) we used in (Passingham et al. 2002) can automati-cally find the optimal cluster numbers and centers, it is stilldependent on a few key parameters that need to be fine-tuned for specific dataset. For instance, there are manypossibilities of clustering the similarity matrix in Fig. 4hinto various numbers of cluster centers, resulting indifferent brain ROIs. In summary, it is an open questionhow to define the possible best boundaries in the cerebralcortex to optimally define brain ROIs.

Individual variability of cortical anatomy and function

The human cerebral cortex is a highly convoluted andcomplex anatomy composed of sulci and gyri. Evidenceshows that cortical folding pattern can be described frommulti-scale perspective (Li et al. 2010c). Curvature, as avery local descriptor of folding pattern, has been used tostudy the cortex folding patterns for its simple computationand effectiveness (Cachia et al. 2003). In addition, there areseveral meso-scale folding pattern descriptors that weredeveloped recently. For example, we proposed a parametricrepresentation of cortical folding patterns via polynomialmodels (Zhang et al. 2009). This parametric foldingdescriptor classifies meso-scale cortical surface patchesinto eight primitive patterns, including peak, pit, ridge,valley, saddle ridge, saddle valley, flat, and inflection.Recently, we proposed a gyrus-scale folding patternanalysis technique via cortical surface profiling (Li et al.2010c). This approach combines advantages of a parametricmethod (achieving compact representation of shape) and aprofiling method (achieving flexibility of arbitrary shaperepresentation) and classify human gyral folding patternsinto three classes according to their numbers of hinges:2-hinge, 3-hinge and 4-hinge gyri. It has been demonstrated in

Fig. 3 The sulcal bank segmentation results on 15 randomly selected central sulci (Li et al. 2010b). The central sulcal basins were segmented intoanterior and posterior sulcal banks via geometric similarity based graph partition (Li et al. 2010b)

Fig. 2 The gyral basin segmen-tation results by our hybridmethod in (Li et al. 2010d) onthe 6 cortical surfaces

Brain Imaging and Behavior (2011) 5:189–202 191

the literature, including our own work (Li et al. 2010c; Zhanget al. 2009), that cortical folding patterns at the curvature,meso-scale and gyral-scale are very variable across individuals.Figure 5 depicts an example at the gyral scale. Theremarkable variation of cortical anatomy raises significantchallenges to identifying corresponding brain ROIs indifferent brains.

There are several folding descriptors of more globalscale such as gyrification index (GI) (Zilles et al. 1988) andovercomplete spherical wavelets (Yeo et al. 2008). The GImetric was used to compute the ratio between the pialcontour and the outer contour in successive coronalsections, and already had a variety of applications.

Recently, Toro et al. proposed surface ratio as a descriptionextended from global scale, e.g., GI, to describe localcortical folding pattern (Toro et al. 2008). Also, theovercomplete spherical wavelets (Yeo et al. 2008), consid-ering the trade-off between local and global scale represen-tation of folding pattern, were used for shape analysis ofcortical surface. Chung et al. employed a weighted linearcombination of spherical harmonics for global representa-tion of folded cortical surfaces (Chung et al. 2008).Duchesnay et al. used over 6,000 features to describe theglobal cortex folding pattern (Duchesnay et al. 2007).Those large numbers of feature parameters come from thevarious attributes of extracted sulci, including the pure

Fig. 4 Flowchart of our methods. a DTI data; b Reconstructed corticalsurface; c tracked fibers; d extracted precentral gyrus; e extracted fibersconnecting precentral gyrus; f fiber shape patterns; g fiber connectivitypatterns; h vertices similarity matrix; i gyrus parcellation result; j fMRIactivation maps. (1): surface reconstruction; (2): fiber tracking; (3):

cortical parcellation and gyrus extraction; (4)–(5): fiber extraction; (6):shape pattern clustering; (7) connectivity pattern clustering; (8):similarity matrix clustering using affinity propagation; (9): finalsegmentation result; (10): validation via fMRI

192 Brain Imaging and Behavior (2011) 5:189–202

shape descriptors (length, area and junction length),absolute position (sulcus extremities, sulcus center of masscoordinates, and midpoint between two neighboring sulci inthe Talairach space), and the spatial relationship betweenneighboring sulci.

In general, sulcal fundi are believed to be one of themost reliable features of cortical folding, and sulcal fundiare believed to be associated with functional and cytoarch-itectural boundaries of cortical regions (Welker 1990).Hence, the composition of a collection of sulcal fundicould be a good representation of global cortical anatomy.Recently, we extracted fourteen major sulcal fundiincluding the central sulcus, post-central sulcus, pre-central sulcus, superior temporal sulcus, superior frontalsulcus, inferior frontal sulcus and cingulate sulcus ineach hemisphere, from 281 brain MRI images using themethod proposed in (Li et al. 2010e). Figure 6a and bdisplay all of the extracted sulcal fundi on a smoothedcortical surface. It is evident that the distributions andshapes of these major sulci fundi are remarkably variableacross different individuals. For instance, the superiorfrontal sulcus and inferior frontal sulcus have significant

variability, which imposes significant challenges to brain ROIidentification and correspondence establishment across indi-vidual subjects.

In addition to anatomic variation, functional localizationof brain ROIs on the cerebral cortex in different brains iseven more variable. As an example, Fig. 7 shows thedistribution of 16 ROIs of the working memory network inthe Montreal Neurologic Institute (MNI) atlas space. TheROIs including bilateral insula, superior frontal gyrus,precentral gyrus, paracingulate gyrus, inferior parietallobule, precuneus, left medial frontal gyrus, right dorsolat-eral prefrontal cortex, left occipital pole, and right lateraloccipital gyrus were obtained by the OSPAN task fMRI(Faraco et al. 2011). The widespread distributions of theseROIs suggest the remarkable variation of functional local-izations in different brains.

Nonlinear properties of ROIs

Cortical ROIs exhibit highly nonlinear properties in termsof structural and functional connectivity profiles. For

Fig. 5 Illustration of differentcortical folding patterns in tworandomly selected humanbrains. Each color represents aroughly corresponding corticalregion

Fig. 6 a–b All the extracted 14 sulcal fundi for the 281 subjects. a Sulcal fundi extracted from left hemispheres: 7 major sulcal fundi are shown indifferent colors (1, 2), and overlaid on a smoothed cortical surface (3, 4); b Sulcal fundi extracted from right hemispheres

Brain Imaging and Behavior (2011) 5:189–202 193

instance, a slight change to the location of an ROI mightdramatically alter its structural and functional connectivitypattern. As shown in Fig. 8, when the ROI is moved fromthe green position to the red position (Fig. 8a), its structuralfiber connection pattern derived from DTI tractographychanges dramatically from that in Fig. 8b to the one inFig. 8c. Here, the fibers derived from streamline tractog-raphy were projected onto the cortex using methods in(Zhang et al. 2010), since it is difficult for DTI tractog-raphy to enter the gray matter due to the low diffusionanisotropy. At the same time, the fMRI BOLD signal forthis ROI also changes significantly, as shown by the greenand red time-series signals in Fig. 8b and c. The remarkablenonlinearity demonstrated in Fig. 8 imposes significantchallenge to ROI identification within and across subjects.For example, the altered functional BOLD signal of brain

ROI will have significant influence on the measurement offunctional connectivity between the ROI in considerationand other ROIs. Similarly, a slight change to the size of anROI might also dramatically alter its structural andfunctional connectivity pattern, as the fibers emanatingfrom the ROI and fMRI signals contained in the ROI willbe changed as well.

A few thoughts on ROI identification

In response to the above discussed challenges in brain ROIidentification, we made some initial effort to optimize,model and predict functionally meaningful ROIs within andacross individual brains. The general principle underlyingour work is that the optimized or predicted ROIs should

Fig. 7 a ROI distribution in onesubject. b ROI distributions of15 subjects in the atlas space.Each ROI is represented by asphere and different colors referto different functional ROIs. 15subjects were randomly selectedand displayed. It is noted thatthe ROIs’ colors in (a) do notcorrespond to those in (b)

Fig. 8 Illustration of structuraland functional connectivitychanges when the location of anROI is changed slightly. a ROIlocation moves from the greento the red bubble. b Structuraland functional (fMRI BOLDsignal) profiles before themovement. c Structural andfunctional profiles after themovement

194 Brain Imaging and Behavior (2011) 5:189–202

have similar structural and functional connection patternswithin a group of subjects.

Optimization of brain ROIs via maximizationof group-wise consistency

Despite the high anatomic variability across subjects, there isdeep-rooted regularity of brain architecture on which we basethe proposed approaches. First, across subjects, the functionalROIs should have similar anatomical locations, e.g., those inthe MNI atlas space (please refer to Fig. 7b). Second, theseROIs should have similar structural connectivity profilesacross subjects. In other words, fibers penetrating the same

functional ROIs should have at least similar target regionsacross subjects (Fig. 9a). Last, individual networks identifiedby task-based paradigms, such as the working memorynetwork we adapted as a test bed in our pilot study, shouldhave similar functional connectivity patterns across subjects.The neuroscience bases of the above premises include: 1)structural and functional brain connectivity are closely related(Passingham et al. 2002; Honey et al. 2009); Hence, it isreasonable to put these three types of information in a jointanalysis framework (Li et al. 2010a). 2) Extensive studieshave already demonstrated the existence of a commonstructural and functional architecture of the human brain(Van Dijk et al. 2010), and it is reasonable to assume that thecommon brain networks have similar structural and func-

Fig. 9 a Comparison of structural connection profiles before and afteroptimization. Each column shows the corresponding before-optimization (top) and after-optimization (bottom) fibers of onesubject. The landmark (right precuneus) is presented by the redsphere. b The movement of right precuneus before (in red sphere) and

after (in green sphere) optimization for eight subjects. The “C”-shapedred dash curve for each subject depicts a similar anatomical landmarkacross these subjects. The yellow arrows in subjects 1, 4 and 7visualize the movement direction after optimization

Brain Imaging and Behavior (2011) 5:189–202 195

tional connectivity patterns across individuals. Based onthese premises, we optimize the locations and sizes ofindividual ROIs by jointly modeling anatomic profiles, andstructural and functional connectivity patterns (Li et al.2010a). The goal is to minimize the group-wise variance (ormaximize group-wise consistency) of the jointly modeledprofiles. Mathematically, we modeled the problem as anenergy function, which was minimized via the simulatedannealing (Li et al. 2010a).

As an example, we used the brain ROIs detected in aworking memory task (Faraco et al. 2011) as a test bed, andFig. 9 shows the white matter fiber connections derivedfrom DTI (Fig. 9a) and ROI locations before and after theoptimization (Fig. 9b). Figure 9a shows the fibers penetrat-ing the right precuneus for eight subjects before (top panel)and after optimization (bottom panel). The ROI is high-lighted in a red sphere for each subject. As we can see fromthe figure (please refer to the highlighted yellow arrows),after optimization, the third and sixth subjects havesignificantly improved consistency with the rest of thegroup. By visual inspection, most of the ROIs moved tomore reasonable and consistent locations after the optimi-zation. As an example, Fig. 9b depicts the locationmovements of the ROI in Fig. 9a for eight subjects. Aswe can see, the ROIs for these subjects share a similaranatomical landmark, which appears to be the tip of theupper bank of the parieto-occipital sulcus. If the initial ROIwas not at this landmark, it moved to the landmark after theoptimization, which was the case for subjects 1, 4 and 7.The results in Fig. 9 indicate the significant improvement ofROI locations achieved by the optimization procedure (Liet al. 2010a). Hence, this ROI optimization method can beused to remedy the fMRI activation peak shift due to spatialsmoothing. In general, this ROI optimization offers apromising solution to the challenges of unclear boundariesbetween cortical regions, the individual variability ofcortical anatomy and connection, and nonlinear ROIproperties.

Interestingly, a major observation that can be made from ourresults in Fig. 9 and in (Li et al. 2010a) is that white matterfiber connection pattern could be a good predictor offunctional ROI. That is, the white matter fiber connectionpatterns for the same functional landmark in different brainsafter optimization are quite consistent (Fig. 9a). This is inagreement with the “connectional fingerprint” concept pre-sented in (Passingham et al. 2002) and the widely acceptedconcept that brain connection and function are closely related.

Identification of ROIs by visual analytics

Effective visual representations and interactions provide themechanism for allowing experts to see and understand large

volumes of information at once. Through a graphic userinterface (GUI) (Fig. 10), raw multimodal imaging data canbe transformed into compact and descriptive representationthat is appropriate for the analytical task and effectivelyconveys the most important content of the large, complex,and dynamic neuroimaging data. Specifically, we havedesigned and implemented an in-house software for theGUI, as shown in Fig. 10. Figure 10a visualizes a corticalsurface, on which the working memory ROIs are overlaid.The red arrow highlights a blue ROI under interactiveediting, e.g., drag-and-drop to change its location. The DTI-derived fibers connecting to the moving blue ROI and thechanging network connectivity are being updated in realtime in response to the ROI change, as shown in Fig. 10b.The representative fMRI signal within this ROI, e.g., thefirst component after a PCA transform on the signals, isbeing updated in Fig. 10c, and the ROI’s locations insagittal, axial and coronal volumetric images are updated inFig. 10d.

In a pilot study, we predicted a functional ROI (leftoccipital pole, highlighted by the yellow arrows in Fig. 11(c)) with the aim of finding similar structural connectivityprofiles for the ROI across subjects. In essence, this is aROI prediction process based on visual structural connec-tivity pattern matching. In Fig. 11, panel (a) shows the fibertracts that penetrate the ROI for 15 subjects (obtained in (Liet al. 2010a)). As can be seen, these fiber tracts have similarshapes, revealing this ROI has strong consistency instructural profiles across subjects. That is the reason whywe consider these shapes as prior knowledge of structuralconnectivity pattern for this ROI, which serves as thematching target in ROI editing and prediction. Panels (b)and (c) depict the ROI’s fiber tracts before and afterinteractive visual editing. The first and third images showthe ROI location from different views. The second imageshows the corresponding fiber tracts, and the fourth imageshows the functional connectivity network. It is apparentthat after interactive editing, the ROI of this subject has amuch better match in shape with other subjects, suggestingthat this visual analytics approach offers an effectivesolution to the challenges of the individual variability ofcortical anatomy and nonlinear ROI properties.

ROI prediction

In the above sections, we used multimodal fMRI and DTIdata for ROI identification and optimization. However, inmany application scenarios, there is no task-based fMRIdata available. For instance, it is challenging to conducttask-based fMRI studies for young children or elderlysubjects such as Alzheimer’s disease patients. Instead, aDTI scan typically needs around 10 min, is much less

196 Brain Imaging and Behavior (2011) 5:189–202

demanding, and is widely available. Therefore, we arestrongly motivated to accurately estimate the locations offunctional ROIs by learning and applying statistical priormodels of ROIs. Specifically, based on the above optimizedROIs (Fig. 9) and DTI data, we constructed consistentwhite matter fiber models as the predictors of functionalROIs in the training stage. In the prediction stage, DTI dataof an individual subject was aligned to the standard spacefirst. Then, starting from the average location of warpedROIs from the training dataset, an energy function wasdesigned to iteratively optimize the ROIs’ locations. Thisenergy function consists of internal and external terms. Theinternal term was measured by ROI coordinate reconstruc-tion error, which corresponds to the ROI coordinate PCAmodel component; while the external term was defined bythe Hausdorff metric of ROIs’ fiber bundles, whichcorresponds to the fiber bundle template component.

As a pilot study, we applied the ROI predictionapproach on the ADNI-2 DTI datasets (Jack et al.2010). To visualize the consistency of fiber bundles ofpredicted ROIs, we showed the fibers emanating fromthree predicted ROIs and fibers emanating from benchmarkROIs for two randomly selected testing ADNI-2 subjects inFig. 12. In addition, corresponding fiber bundles of threesubjects randomly selected from the training dataset areshown in Fig. 12 for comparison. It is evident that the fiber

bundles of the predicted ROIs are quite similar to those ofthe corresponding ROIs in the training dataset (the first tothird columns in Fig. 12). We also performed ROI predictionon healthy subjects with coincident fMRI and DTI datasets.The activation peaks provided by task-based fMRI was usedas benchmark data. On average, the prediction error for ROIsover 5 separate subjects is 3.8 mm.

Scales of brain ROIs

In many previous studies, e.g., in (Bullmore and Sporns2009), structural brain networks were typically constructedin a single resolution, that is, the sizes and numbers ofgraph nodes were fixed for the same dataset. However,structural brain networks can be viewed and defined atmultiple scales, e.g., at micro-, meso-, or macro- scales. Forinstance, in our recent studies in (Yuan et al. 2011), weapplied our recently developed approach based on fiberdensity guided cortical parcellation (Zhang et al. 2010) onhigh resolution DTI dataset to construct multi-resolutionstructural brain networks, shown in Fig. 13 as an example.Then, we use state-of-the-art graph analysis algorithms inGraphCrunch (Milenkovic et al. 2008) to measure theglobal and local graph properties of the structural networks.It turns out that graph properties in different resolutions

Fig. 10 The GUI. a the view of cortical surface and ROIs; b the view of fibers and network; c fMRI signal view; d volume view

Brain Imaging and Behavior (2011) 5:189–202 197

could be significantly different (Yuan et al. 2011). It is alsoconceivable that structural connectivity patterns in networksof different scales are very different. At the current stage, itis unclear which resolution is the best one to representstructural connectivity patterns of the brain. Interestingly,in a recent study of axonal wiring of Drosophila brain(Chiang et al. 2010), it was reported that the localprocessing unit (LPU) has remarkable consistency andreliability across many different Drosophila brains. Thus,we postulate that group-wise consistency is one of thefundamental principles that should guide the definition ofbrain ROIs for measurement of brain connectivity, nomatter if it is at the micro-, meso-, or macro- scale. Actually,this principle has already been used in our ROI optimizationmethod in Fig. 9. In addition, future exploration of otherfundamental principles is warranted.

Dynamics of functional connectivities between brainROIs

Functional connectivities between brain ROIs have beenextensively studied via functional MRI (fMRI) in recentyears. Currently, many approaches for analysis of functionalconnectivities on brain networks assume temporal stationarity,that is, functional connectivities are computed over theentire scan and used to characterize the strengths ofconnections across regions. However, accumulating liter-ature, including our own recent studies (Lim et al. 2011;Hu et al. 2011), have shown that functional brainconnectivity is under dynamic changes in different timescales. As an example shown in Fig. 14a and b, wemeasured the temporal correlation of fMRI time series oftwo ends of a DTI-derived fiber to define the functional

Fig. 11 Interactive ROI predic-tion by using fiber profile asprior knowledge. a left occipitalpole ROIs (highlighted in darkred) from 15 subjects with fibertracts overlaid. b and c show thecomparison before (b) and after(c) user interaction. The ROI ishighlighted by yellow arrows. Ineach panel, the first image andthird image show the ROI loca-tion from different views. Thesecond image shows thecorresponding fiber tracts, andthe fourth image shows thefunctional connectivity network

198 Brain Imaging and Behavior (2011) 5:189–202

connectivity (FC) between the voxels it connects. Then,the FC patterns of all the white matter fibers within thewhole brain are concatenated into a feature vector(Fig. 14c) to represent the brain’s state (Lim et al. 2011).It is evident that the FC patterns of individual fibers, as

well as the whole-brain, changes dramatically along thetime axis.

Neuroscience research suggests that the function of anycortical area is subject to top-down influences of attention,expectation, and perceptual task (Gilbert and Sigman 2007).

Fig. 12 Visualization of fiber bundles of three predicted ROIs in MCIpatients (Jack et al. 2010). Left gray frames: the template fiber bundlesemanating from corresponding ROIs in the training dataset (yellow

bubbles); Right green frames: fiber bundles emanating from predictedROIs (green bubbles) in MCI patients

Fig. 13 Examples of multi-resolution structural brain networks. 10 resolutions were illustrated for one subject. The number of nodes in eachnetwork is 2063, 2051, 1802, 1667, 1495, 1221, 447, 82, 38 and 21, respectively

Brain Imaging and Behavior (2011) 5:189–202 199

For instance, dynamic interactions between connections fromhigher- to lower-order cortical areas and intrinsic corticalcircuits mediate the moment-by-moment functional switchingin brain. Even in resting state, the functional connectivity isstill under dynamic changes within time scales of seconds tominutes (Lindquist et al. 2007). In the literature, there hasbeen a growing number of studies that tackled the problemof modeling the dynamics of brain states and functionalconnectivity [e.g., (Lim et al. 2011; Hu et al. 2011; Gilbertand Sigman 2007; Lindquist et al. 2007; Robinson et al.2010; Gao and Yee 2003; Morgan et al. 2004; Chang and

Glover 2010)]. Despite significant progress in the literature,it is still an open and challenging problem, in our view, dueto at least two critical reasons. 1) FMRI blood-oxygenationlevel dependent (BOLD) signal is characterized by its non-linearity, non-stationarity and composition of signal compo-nents at multiple time scales (Heeger and Ress 2002;Logothetis 2008; Buzski and Draguhn 2004), which imposessignificant challenges to inferring meaningful informationfrom it. 2) Functional brain dynamics is a system behavior,and its quantitative modeling entails mathematically soundand biologically meaningful approaches.

Fig. 14 a DTI-derived fibers.Five fibers are highlighted andnumbered. b Dynamic functionalconnectivities (DFCs) of fivepairs of fiber end points alongthe time axis (seconds). Eachrow represents the DFC curve ofa fiber in (a). The FC range is−1.0 to 1.0. Color bar is on thetop right. c DFC curves of allfibers in a brain are representedby the CSVs (columns) along thehorizontal time axis

200 Brain Imaging and Behavior (2011) 5:189–202

Discussions and concluding remarks

In this article, the working memory network was used as atest bed system. Multimodal fMRI and DTI data was usedfor the ROI optimization, modeling and prediction. In thefuture, we plan to further evaluate and validate our ROIoptimization and prediction framework (Li et al. 2010a)based on multimodal data for other brain networks such asattention, language, motor, emotion and vision systemsusing independent datasets. We envision that our frame-work would be a general framework that is applicable toreveal the common architecture of the human brain,represented by consistent structural connectivity patternsacross individuals. In a broader sense, the optimized andmodeled dense map of brain ROIs derived from multiplemultimodal DTI and fMRI datasets can be considered andused as the next-generation brain atlas, which will havemuch finer granularity and better functional homogeneitythan the Brodmann brain atlas that has been used in thebrain science field for over 100 years. Furthermore,advanced diffusion imaging techniques such as HARDI(Tuch et al. 1999) have the potential to significantlyimprove current spatial and angular resolutions of DTIdata, and more extensive evaluation studies are warrantedto elucidate the relationships between structural connectivityand brain function.

In addition, the optimized ROI map can be used as ageneral reference coordinate system for functional brainmapping via fMRI. Over the past two decades, fMRI hashad an enormous impact on how we study the human brain.However, the current common practice is to reportstereotaxic coordinates for brain activations, usually inrelation to the Talairach or the Montreal NeurologicalInstitute coordinate system (74% of over 9,400 fMRIstudies (Derrfuss and Mar 2009)). The accuracy with whichwe can compare coordinates will depend greatly on thesimilarity of the images, atlases, and normalization techni-ques used. In particular, comparison of functional localiza-tion and brain function from different fMRI imaging studieshas been complicated by reporting the stereotaxic coordi-nates. With the availability of a dense ROI map optimizedand modeled in the future, it will become possible to reportfunctional activities in fMRI in relation to the identifiedROIs. Since the ROIs are consistent across different brainswith intrinsically established correspondences, the compari-son and pooling of results from different fMRI studies becomemuch more effective and efficient.

Finally, the ROI map will have significant and wideapplications in clinical neuroscience. In general, clinicaldiagnosis, therapy and followup of numerous brain diseasesentail accurate localization of brain ROIs, which has beenan open problem for years due to the practical difficulty ofacquiring task-based fMRI data for patients. Essentially, the

ROI prediction framework presented in this paper providesa promising solution to this challenging problem byoffering an effective and efficient approach to estimatingthe localizations of functionally meaningful ROIs based onwidely available DTI data. It should be noted, however, thatstructural and functional connectivities in diseased brainsmight have already been altered during disease progression.In this case, application of our approaches on a specificgroup of patients using an adaptive strategy will bewarranted.

Acknowledgements T Liu was supported by the NIH Career Award(NIH EB 006878), NIH R01 HL087923-03S2 and The University ofGeorgia start-up research funding. The author would like to thankGang Li, Kaiming Li, Dajiang Zhu, Tuo Zhang, Chulwoo Lim, andJianfei Yang for providing some of the figures used in this paper.

References

Ashburner, J., Friston, K., & Penny, W. (2004). Human brain function.Academic Press.

Beckmann, C. F., DeLuca, M., Devlin, J. T., & Smith, S. M. (2005).Investigations into resting-state connectivity using independentcomponent analysis. Philosophical Transactions of the RoyalSociety of London. Series B: Biological Sciences, 360, 1001–1013.

Behrens, T. E. J., et al. (2003). Non-invasive mapping of connectionsbetween human thalamus and cortex using diffusion imaging.Nature Neuroscience, 6, 750–757.

Biswal, B. B. (2010). Toward discovery science of human brainfunction. PNAS, 107(10), 4734–4739.

Brett, M., Johnsrude, I. S., & Owen, A. M. (2002). The problem offunctional localization in the human brain. Nature Reviews.Neuroscience, 3(3), 243–249.

Bullmore, E., & Sporns, O. (2009). Complex brain networks: graphtheoretical analysis of structural and functional systems. NatureNeuroscience Reviews, 186(10).

Buzski, G., & Draguhn, A. (2004). Neuronal oscillations in corticalnetworks. Science, 304(5679), 1926–1929.

Cachia, A., Mangin, J.-F., Rivière, D., Kherif, F., Boddaert, N.,Andrade, A., et al. (2003). A primal sketch of the cortex meancurvature: a morphogenesis based approach to study thevariability of the folding patterns. IEEE Transactions on MedicalImaging, 22, 754–765.

Calhoun, V. D., Pekar, J. J., & Pearlson, G. D. (2004). Alcoholintoxication effects on simulated driving: exploring alcohol-doseeffects on brain activation using functional MRI. Neuropsycho-pharmacology, 29, 2097–3017.

Chang, C., & Glover, G. H. (2010). Time—frequency dynamics ofresting-state brain connectivity measured with fMRI. Neuro-image, 50(1), 81–98.

Chiang, A., et al. (2010). Three-dimensional reconstruction of brain-wide wiring networks in drosophila at single-cell resolution.Current Biology, 21(1), 1–11.

Chung, M., Hartley, R., Dalton, K., & Davidson, R. (2008). Encodingcortical surface by spherical harmonics. Statistica Sinica, 18,1269–1291.

Cohen, A. L., et al. (2008). Defining functional areas in individualhuman brains using resting functional connectivity MRI. Neuro-image, 41(1), 45–57.

Derrfuss, J., & Mar, R. A. (2009). Lost in localization: the need for auniversal coordinate database. Neuroimage.

Brain Imaging and Behavior (2011) 5:189–202 201

Duchesnay, E., Cachia, A., Roche, A., Rivière, D., Cointepas, Y.,Papadopoulos-Orfanos, D., et al. (2007). Classification based oncortical folding patterns. IEEE Transactions on Medical Imaging,26(4), 553–565.

Faraco, C. C., Unsworth, N., Lagnely, J., Terry, D., Li, K., Zhang, D.,et al. (2011). Complex span tasks offer a unique view of workingand long term memory. NeuroImage.

Fischl, B., et al. (1999). Cortical surface-based analysis II: inflation,flattening, and a surface-based coordinate system. Neuroimage, 9,195–207.

Frey, B. J., & Dueck, D. (2007). Clustering by passing messagesbetween data points. Science, 315, 972–976.

Friston, K. J. (2009). Modalities, modes, and models in functionalneuroimaging. Science, 326(5951), 399–403.

Friston, K. J., et al. (2003). Dynamic causal modeling. Neuroimage,19, 1273–1302.

Gao, J.-H., & Yee, S.-H. (2003). Iterative temporal clustering analysisfor the detection of multiple response peaks in fMRI. MagneticResonance Imaging, 21(1), 51–53.

Geissler, A., Lanzenberger, R., Barth, M., Tahamtan, A. R., Milakara,D., Gartus, A., et al. (2005). Influence of fMRI smoothingprocedures on replicability of fine scale motor localization.Neuroimage, 24, 323–331.

Gilbert, C. D., & Sigman, M. (2007). Brain states: top-downinfluences in sensory processing. Neuron, 54(5), 677–696.

Hagmann, P., et al. (2010). MR connectomics: principles andchallenges. Journal of Neuroscience Methods, Jan 22.

Handbook of Functional Neuroimaging of Cognition, 2nd Edition, byRoberto Cabeza and Alan Kingstone.

Heeger, D. J., & Ress, D. (2002). What does fmri tell us aboutneuronal activity? Nature Reviews. Neuroscience, 3(2), 142–151.

Honey, C. J., et al. (2009). Predicting human resting-state functionalconnectivity from structural connectivity. PNAS, 106(6), 2035–2040.

Hu, X., et al. (2011). Assessing the dynamics on functional brainnetworks using spectral graph theory, ISBI 2011.

Hyvärinen, A., & Oja, E. (2000). Independent component analysis:algorithms and applications. Neural Networks, 13(4–5), 411–430.

Jack, C. R., Jr., Bernstein, M. A., Borowski, B. J., Gunter, J. L., Fox,N. C., Thompson, P. M., et al. (2010). Update on the magneticresonance imaging core of the Alzheimer’s disease neuroimaginginitiative. Alzheimer’s & Dementia, 6(3), 212–220.

Jo, H. J., Lee, J.M., Kim, J. H., Choi, C. H., Gu, B.M., Kang, D. H., et al.(2008). Artificial shifting of fMRI activation localized by volume-and surface-based analyses. Neuroimage, 40(3), 1077–1089.

Li, K., et al. (2009a). Review of methods for functional brainconnectivity detection using fMRI. Computerized MedicalImaging and Graphics, 33(2), 131–139.

Li, G., et al. (2009b). Automatic cortical sulcal parcellation based onsurface principal direction flow field tracking. Neuroimage, 46(4), 923–937.

Li, K., et al. (2010a). Individualized ROI optimization via maximizationof group-wise consistency of structural and functional profiles.Advances in Neural Information Processing Systems (NIPS).

Li, G., et al. (2010b). Cortical Sulcal Bank segmentation via graphpartition: methods and applications. MIAR (Medical Imaging andAugmented Reality).

Li, K., Guo, L., Li, G., Nie, J., Faraco, C., Cui, G., et al. (2010c).Gyral folding pattern analysis via surface profiling. NeuroImage.

Li, G., Guo, L., Zhang, T., Nie, J., & Liu, T. (2010d). Automaticcortical gyral parcellation using cortical probabilistic atlas andgraph cuts. MIAR (Medical Imaging and Augmented Reality).

Li, G., Guo, L., Nie, N., & Liu, T. (2010e). An automated pipleline forsulci fundi extraction. Medical Image Analysis, 14(3), 343–359.

Lim, C., Li, X., Li, K., Guo, L., & Liu, T. (2011). Brain state changedetection via fiber-centered functional connectivity analysis, ISBI2011.

Lindquist, M. A., Waugh, C., & Wager, T. D. (2007). Modeling state-related fMRI activity using change-point theory. Neuroimage, 35(3), 1125–1141.

Logothetis, N. K. (2008). What we can do and what we cannot dowith fmri. Nature, 453(7197), 869–878.

Milenkovic, T., et al. (2008). GraphCrunch: a tool for large networkanalyses. BMC Bioinformatics, 9, 70.

Morgan, V. L., Price, R. R., Arain, A., et al. (2004). Resting functionalMRI with temporal clustering analysis for localization ofepileptic activity without EEG. Neuroimage, 21(1), 473–481.

Passingham, R. E., et al. (2002). The anatomical basis of functionallocalization in the cortex. Nature Reviews. Neuroscience, 3(8),606–616.

Robinson, L. F., Wager, T. D., & Lindquist, M. A. (2010). Changepoint estimation in multi-subject fMRI studies. Neuroimage, 49(2), 1581–1592.

Shen, D., & Davatzikos, C. (2002). HAMMER: hierarchical attributematching mechanism for elastic registration. IEEE Transactionson Medical Imaging, 21(11), 1421–1439.

Shen, D., et al. (2002). HAMMER: Hierarchical Attribute MatchingMechanism for Elastic Registration. IEEE Transactions onMedical Imaging, 21(11), 1421–1439.

Sobel, D. F., Gallen, C. C., Schwartz, B. J., Waltz, T. A., Copeland,B., Yamada, S., et al. (1993). Locating the central sulcus:comparison of MR anatomic and magnetoencephalographicfunctional methods. AJNR. American Journal of Neuroradiology,14(4), 915–925.

Sporns, O., Tononi, G., & Kötter, R. (2005). The human connectome:a structural description of the human brain. PLoS ComputationalBiology, 1(4), e42.

Toro, R., Perron, M., Pike, B., Richer, L., Veillette, S., Pausova, Z., etal. (2008). Brain size and folding of the human cerebral cortex.Cerebral Cortex, 18(10), 2352–2357.

Tuch, D. S., et al. (1999). High angular resolution diffusion imagingof the human brain. In ISMRM’1999.

Van Dijk, K. R., Hedden, T., Venkataraman, A., Evans, K. C., Lazar,S. W., & Buckner, R. L. (2010). Intrinsic functional connectivityas a tool for human connectomics: theory, properties, andoptimization. Journal of Neurophysiology, 103(1), 297–321.

Van Essen, D. C., & Dierker, D. L. (2007). Surface-based andprobabilistic atlases of primate cerebral cortex. Neuron, 56.

Welker, W. (1990). Why does cerebral cortex fissure and fold? Areview of determinants of gyri and sulci. Cerebral Cortex 8b.

White, T. O., Leary, D., Magnotta, V., Arndt, S., Flaum, M., &Andreasen, N. C. (2001). Anatomic and functional variability:the effects of filter size in group fMRI data analysis. Neuroimage,13, 577–588.

Yeo, B. T., Yu, P., Grant, P. E., Fischl, B., & Golland, P. (2008). Shapeanalysis with overcomplete spherical wavelets. In Proc. MICCAI,LNCS 5241, 468–476.

Yuan, Y., et al. (2011). Assessing graph models for description ofbrain networks, ISBI 2011.

Zang, Y., et al. (2004). Regional homogeneity approach to fMRI dataanalysis. Neuroimage, 22(1), 394–400.

Zhang, T., Guo, L., Li, G., Nie, J., & Liu, T. (2009). Parametricrepresentation of cortical surface folding via polynomials.Medical Image Computing and Computer Assisted Intervention(MICCAI).

Zhang, D., et al. (2010). Automatic cortical surface parcellation basedon fiber density information. International Symposium ofBiomedical Imaging (ISBI) 2010.

Zhu, D., et al. (2011). Fine granularity parcellation of Gyrus via fibershape and connectivity based features, ISBI 2011.

Zilles, K., Armstrong, E., Schleicher, A., & Kretschmann, H. J.(1988). The human pattern of gyrification in the cerebral cortex.Anatomy and Embryology (Berl), 179, 173–179.

202 Brain Imaging and Behavior (2011) 5:189–202