Resting state-fMRI approach towards understanding impairments in mTLE
RResting state-fMRI approach towards understandingimpairments in mTLE
Hritik Bansal *
Electrical Engineering DepartmentIndian Institute of Technology Delhi
Nishad Singhi *
Electrical Engineering DepartmentIndian Institute of Technology
Abstract —Mesial temporal lobe epilepsy (mTLE) is the mostcommon form of epilepsy. While it is characeterised by anepileptogenic focus in the mesial temporal lobe, it is increasinglyunderstood as a network disorder. Hence, understanding thenature of impairments on a network level is essential for itsdiagnosis and treatment. In this work, we review recent worksthat apply resting state functional MRI to provide key insightsinto the impairments to the functional architecture in mTLE.We discuss changes on both regional and global scales. Finally,we describe how Machine Learning can be applied to rs-fMRIdata to extract resting state networks specific to mTLE and forautomated diagnosis of this disease.
Index Terms —Independent components, fMRI. resting state,Default mode network, perceptual networks, Machine learning(ML), graph theory, feature selection
I. I
NTRODUCTION
Epilepsy in general is associated with heterogenity in clin-ical history, neuropathological condition, neurophysiologicalfeatures, electroencephalographic and neuroimaging findings.Mesial temporal lobe epilepsy (mTLE) is epilepsy of temporalorigin which is defined by characteristic hippocampal sclerosis(HS) as its pathological substrate. mTLE is considered to besurgically treatable condition, and efforts are being taken tounderstand its characteristics precisely [28].Non-invasive imaging technique such as functional mag-netic resonance imaging (fMRI) use stimulus driven paradigmsto delve into the intracacies of brain function. Recently, therehas been an enormous increase in interest in the applicationof this technique at rest, which termed as resting state fMRI(rs-fMRI).rs-fMRI may provide greater insight into the functionalconnectivity in the neural systems, rather than what is beingactivated by a particular task. It also helps in eliminating thenoise associated with atypical strategies related to task designand the sensitivity and specificity of various metrics. [29], [30].A number of methods are available to analyze the rs-fMRI data like seed-based analysis, Independent componentanalysis (ICA) and graph based approach. Out of these, seed-based analysis requires apriori selection of Regions of Inter-ests (RoIs) based on predetermined templates. This methodrequires thresholding to identify voxels that are significantly * Both authors contributed equally to this manuscript correlated with the RoI. In this study, we will be lookingat ICA specifically because of its various advantages, ashighlighted in Section II. Graph based methods have revealedthat the brain exhibits small-world topology, which allowsnodes (individual RoIs) to have low number of connectionswhile still being connected to all other nodes with a shortdistance [29]. We touch upon this in our study in section VII.With the advent of machine learning (ML) and compu-tational resources, it has been possible to learn complexfunctions that fit the data. Studies have used hand-engineeredfeatures in some cases, and in the other let the model derivethose features automatically from the raw data using moresophisticated techniques like deep learning. In this studyVI, VII-B, we highlight how machine learning techniqueshave been exploited recently for better understanding anddiscovery of epileptic networks which can be missed by sheercorrelation-based statistical methods. studiesII. I
NDEPENDENT C OMPONENTS FROM F
MRI
DATA
Analytical techniques like time-frequency analysis, Analy-sis of variance (ANOVA) and principal component analysis(PCA) have several drawbacks pertaining to extraction ofuseful information from changes in fMRI data (includes bloodoxygen level-dependent (BOLD) signal). Generally, cogni-tive/perceptual task related fMRI signals are typically small( < ), suggesting other time-varying phenomena must becontributing to the bulk of the measured signal [1].In time-frequency analysis, the time signals associated witheach voxel are converted into fourier (frequency) domain.Such techniques assume that the task performed by the subjectcauses distinguishable changes in the frequency spectrum ofthe data. Some techniques also assume the source signals ofthe data to be periodic, which is infact a strong assumption tomake. ANOVA [4], [5] makes certain assumptions on the priordistribution of the data (e.g. Gaussian), time series associatedwith each voxel being independent and variances betweenrepeated measurements are equal. However, such techniquesdo not extract the intrinsic structure of the data, and do notcapture transient task related changes.PCA [3] has long been considered as a way to identify thetendency of signals at each pair of voxels to covary. It findsthe spatial patterns (eigenvectors) which capture the greatest a r X i v : . [ q - b i o . N C ] S e p ariance of the data. However, when task related signal is verysmall in comparison to the bulk, getting eigenvectors (images)associated with greatest variance in the data might not beuseful. There is a tendency to miss overall associations in thebrain networks using this technique when multiple voxels getactivated simultaneously [1].Independent Component Analysis (ICA) [1], [6] techniqueposits that separate processes associated with multifocal brainareas may be represented by multiple spatially-independentcomponents. Each independent component has a single timesignal and spatial map associated with it (Fig. 1). We presumethat high-order correlations between the voxels values in thepairs of components is zero. ICA is posed as solving: X = M C (1) C = W X (2)where C ∈ R n × K is a matrix of independent components,rows of which consists of n independent components withK being the number of voxels. M ∈ R T × n is the mixingmatrix, which controls the contribution of each independentcomponent to the observed signal X ∈ R T × K . T is thenumber of time datapoints in the observed signal. W is theunmixing signal, which helps in retrieving the underliningcomponents C. We estimate W Using an iterative unsupervisedlearning algorithm [6] based on mutual information principlesand without making an prior assumptions on the activation’sspatiotemporal extent.Rather than considering the observed signal as a linear combi-nation of independent components, non-linear ICA [8] posesthe problem as: x ( t ) = f ( s ( t )) (3)where x(t) is an n-dimensional data point at time t, and fis smooth and invertible mixing function, and s(t) is an n-dimensional vector of independent components s i ( t ) . Even thetime series s i are assumed to be mutually independent. [2],[7] show that non-linear ICA of the fMRI data can expressfunctional traits associated with resting state, cognitive andperceptual task in the human brain.III. R ESTING S TATE N ETWORKS
Positron emission tomography (PET) and functional Mag-netic Resonance Imaging (fMRI) studies [10] have providedevidence of transient changes ( activations ) induced by cog-nitive and perceptual tasks. Several studies [11], [14] havesuggested that fMRI images obtained using blood oxygen leveldependent (BOLD) contrast show signal fluctuations at rest(awake or anesthetized brain).The regions displaying coherent low frequency (0.01-0.1 Hz)fluctuations constitute a resting state network (RSN). Theunderstanding of the neurophysiological basis of these RSNshelps in identifying the functional role of spontaneous activity.[15] identify resting state patterns from the BOLD signals byusing Independent Component Analysis (ICA) (Fig. 2).
Fig. 1. fMRI data decomposed into spatial independent components. [1]. Firstcolumn of the mixing matrix (M) gives the time course modulation of the firstIC, second column gives the time course modulation of the second IC and soon.Fig. 2. Example of 1 IC seperated from a subject using ICA [15]. The figureshows the sagittal, coronal and axial spatial map of the IC. Orange coloredbrain areas show positive correlation with the IC waveform, and the blue onesshow negative correlation.
Across subjects, [15] identified six RSNs, as exemplified inthe Table I.Several studies [15], [21] have shown the correlation be-tween the resting state fMRI and electroencephalography(EEG) power data collected simultaneously from the subjects.EEG-fMRI and rs-fMRI have frequently shown networks re-lated to epileptic transients, characterization of which becomesimportant in understanding the neurological differences in thecontrols vs epilepsy patients.
SN
ESTING STATE NETWORKS AND THEIR CORRESPONDING REGIONS ANDTASK THAT THEY ARE ASSOCIATED WITH . IV. S
TRUCTURAL AND F UNCTIONAL ABNORMALITIES IN
DMNThe focal point in mTLE is usually in the mesial temporalregion (hence the name), usually hippocampal sclerosis (HS).But, mTLE is increasingly being understood as a networkdisorder, meaning that its effects can extend to regions farbeyond the mesial temporal lobe. Previous studies have shownthat epileptic activity can extend from the temporal lobe toother regions of the ‘Default mode network’ [9], which is anetwork of regions that are functionally connected to eachother and are usually active during wakeful rest. It has beenshown that functional and structural connectivity overlap inthe case of DMN and that structural organization is directlyrelated to the functional synchronization of DMN [12] [13].Hence, [16] studied how structural and functional connectivitychanges in the DMN.Functional connectivity was measured using Pearson cor-relation (in fMRI) and structural connectivity was measuredusing connection density (number of connections per unitsurface) and path length (average length of all connectingfibres). As compared to controls, patients showed a decreasein both connection density and temporal correlation betweenposterior cingulate cortex (PCC)/precuneus (PCUN) and bothmesial temporal sclerosis (mTLs). These results suggest thatthe sturctural connecticity between these two areas is de-graded and along with this, decreased functional connectivitymay cause DMN abnormalities in patients suffering frommTLE. Patients showed a decrease in the maximum fractionalanisotropy value between PCC/PCUN and medial prefrontalcortex (mPFC). Apart from this, no significant differencewas found between these two regions. Interestingly, meantemporal correlation was found to be significantly correlatedwith connection density between PCC/PCUN and mTLs inboth groups, further suggesting that functional connectivityand structural connectivity between these areas are related.Broadly, these results suggest that while HS is the epilepto-genic focus, it may cause structural degradation of connec-tions in DMN (like the connections between PCC/PCUN andmTLs), which contributes to functional decline in patients.V.
RSF
MRI
AND P ERCEPTUAL NETWORKS IN M
TLEmTLE has shown to adversely affect the cognitive function-ing such as memory, largely associated with DMN IV, andlanguage [22]. Studies have also shown impairments in theattention as well as perceptual networks [23], [24]. Perceptualnetworks are linked to the functionality of the visual, auditory and sensorimotor systems (Table I). ICA based analysis ofBOLD-fMRI data in resting state has provided significantinsights into neurophysiological mechanisms underlying func-tional impairments in mTLE. The results [23] obtained fromthis analysis of RSN agreed with earlier findings [25], [26].By performing correlation based statistical analysis of theRSNs corresponding to perceptual networks of the mTLEpatients and controls, we can assess the functional connectivitydifferences between the two Fig. 3. It is important to note asper Fig. 3 the increased functional connectivity within primarycortex might be linked to the plasticity behavourial changesoccuring in the mTLE patients in response to impairments inother parts like bilateral MT+ areas.In [23], the authors show that mean z-values within the regionof interests (RoIs) from Fig. 3 are negatively correlated witha clinical parameter: epilepsy duration, but uncorrelated toseizure frequency.
Fig. 3. Functional connectivity differences between the controls and mTLEpatients [23]. Blue indicates decreased connectivity within a region for patientswhen compared with the controls; Red indicates the increased connectivity.
VI. I
DENTIFICATION OF RS - F MRI
NETWORKS IN
TLE
USING
MLResting-state networks were introduced in Section III. WhileICA (see Section II) can extract a large number of RSnetworks, research in fMRI typically focuses on a few well-studied resting state networks (see Table I) because theirspatial structure is consistent with the understanding of brainfunctions developed using task-based fMRI. However, power-ful machine-learning techniques have made it possible to studythe remaining components derived by ICA. [17] analysed rs-fMRI data of 42 patients and 90 controls with the aim ofidentifying ICs that could be indicative of TLE.After extracting 88 ICs using ICA, top 10 ICs were iden-tified with the help of an elastic-net based feature selectionmethod. These networks included frontal, temporal, perisyl-vian, cingulate, posterior-quadrant, thalamic, and cerebellarregions. The intensity of these ICs were correlated with clinicalvariabled and hippocampal volumes and it was found thatany of these components were significantly correlated withduration of epilepsy, number of anti-epileptic drugs, durationof epilepsy, etc. Two ICs showed significant correlation withthe affected hippocampal volume and not with the unaffectedone.The significance of this study is clear from the fact thatthe networks identified using ML correlated strongly withseveral clinical variables, indicating that they are specific tothis disease. Moreover, the strengths of these ICs was onlyrelated to the affected hippocampus and not the unaffected one,which again supports the epilepsy-specific nature of these ICs.Also, IC strength is negatively correlated with hippocampalvolume which implies that hippocampal volume is lost as thedisease condition worsens, which is consistent with previousfindings [19]. Finally, the method achieves accuracy that iscomparable to previous studies using different techniques.VII. G
RAPH THEORY BASED APPROACHES
A. Altered functional connectivity and small world
Several studies have shown that functional connectivitywithin DMN is altered in patients suffering from mTLE.However, most of these studies focus on local changes inconnectivity, analysing changes in individual connections. [20]analysed rs-fMRI data of 18 patients and 27 controls usinggraph theoretical measures that provide a more global pictureof differences between patients and healthy subjects.90 RoIs were defined in the brain and their correlation ma-trix was computed. Each region served as a node in a graph andan edge existed between the two nodes if their correlation wasabove a certain threshold. Several graph theory measures werecomputed and compared between the two groups. Significantdifferences were found between the two groups for severaltopological variables, suggesting a macroscopic reorganizationin mTLE. It was found that connectivity decreased in thefrontal, parietal, and occipital lobes, which are some of thekey areas in DMN. It is known that both functional andstructural connectivity in the DMN are altered in patients(See Section IV), hence these findings are consistent with theliterature. Significant negative correlation between rIFGoperand lIFGtri with the epilepsy duration was found, providingfurther evidence that changes in functional connectivity affectobservable clinical variables. Along with decreased functionalconnectivity in the DMN, an increase in functional connectiv-ity was observed in other areas of the brain, which suggeststhat epilepsy may arise from an imbalance of connectivity inregions of the brain.Several important regions showed decreased values of de-gree in patients, which suggests that their connectivity to otherregions of the brain was affected, which may inhibit the flowof information. Particularly, PCC/PCUN, which is a crucialarea in the DMN (see Section IV), showed a decrease in thevalue of degree. There were significant differences in the n-to-1 connectivity values of several regions across the groups,suggesting altered functional connectivity.The clustering coefficients were smaller in patients, suggest-ing that the graphs were sparser. Shorter path lengths were found in patients which indicates faster and more efficient in-teractions between regions. Finally, the small-world propertieswere altered in patients in a way that their graphs were closerto random graphs. These findings suggest that graph theorymeasures can be used as markers for this disease.
B. Using ML and graph-based features to detect lateralizationin mTLE
One of the most important aspects of preseurgical testing inmTLE is successful lateralization of the affected hemisphere.Usually, this is done by manual assessment of structural MRIimages, which leads to accurate classification in 70-85% cases.As we have previously seen, functional MRI provides rich in-formation about changes in mTLE. [27] employed graph-basedfeatures as used in the previous study and machine learningto develop an automated lateralization pipeline (henceforthreferred to as CADFIG).CADFIG could accurately lateralize 95.8% of the patients,which was better than manual assessment (66.7% accuracy).Moreover, CADFIG showed a 42.9% increase in sensitivityfor right TLE and 44.4% increase for left TLE (see Figure 4).Clearly, CADFIG is superior than manual assessment, whichis treated as the ‘gold standard’ in presurgical evaluation oflateralization. When the two methods were used together, allpatients were classified correctly, which means that a multi-modal approach may work best. Since functional connectivitycould correctly lateralize most of the patients, it can be saidthat right and left mTLE differ as far as functional connectomeis concerned.
Fig. 4. Classification results of CADFIG, Manual assessment (denoted byMRI) and combined approaches [27].
VIII. C
ONCLUSIONS AND FUTURE DIRECTIONS