Remodeling of brain morphology in temporal lobe epilepsy

Abstract Background Mesial temporal lobe epilepsy (TLE) is one of the most widespread neurological network disorders. Computational anatomy MRI studies demonstrate a robust pattern of cortical volume loss. Most statistical analyses provide information about localization of significant focal differences in a segregationist way. Multivariate Bayesian modeling provides a framework allowing inferences about inter‐regional dependencies. We adopt this approach to answer following questions: Which structures within a pattern of dynamic epilepsy‐associated brain anatomy reorganization best predict TLE pathology. Do these structures differ between TLE subtypes? Methods We acquire clinical and MRI data from TLE patients with and without hippocampus sclerosis (n = 128) additional to healthy volunteers (n = 120). MRI data were analyzed in the computational anatomy framework of SPM12 using classical mass‐univariate analysis followed by multivariate Bayesian modeling. Results After obtaining TLE‐associated brain anatomy pattern, we estimate predictive power for disease and TLE subtypes using Bayesian model selection and comparison. We show that ipsilateral para‐/hippocampal regions contribute most to disease‐related differences between TLE and healthy controls independent of TLE laterality and subtype. Prefrontal cortical changes are more discriminative for left‐sided TLE, whereas thalamus and temporal pole for right‐sided TLE. The presence of hippocampus sclerosis was linked to stronger involvement of thalamus and temporal lobe regions; frontoparietal involvement was predominant in absence of sclerosis. Conclusions Our topology inferences on brain anatomy demonstrate a differential contribution of structures within limbic and extralimbic circuits linked to main effects of TLE and hippocampal sclerosis. We interpret our results as evidence for TLE‐related spatial modulation of anatomical networks.


| INTRODUC TI ON
Temporal lobe epilepsy (TLE), one of the most common forms of focal epilepsy, is associated with progressive cognitive dysfunction and resistance to antiepileptic drug therapy (Wiebe & Jette, 2012).
Although considered for many years as disorder related to focal temporal lobe pathology, there is strong evidence for disruptions in a widespread cortico-subcortical network (Bernhardt, Hong, Bernasconi, & Bernasconi, 2013;Bonilha et al., 2013;Concha, Kim, Bernasconi, Bernhardt, & Bernasconi, 2012). Given the fact that the TLE clinical phenotype is modified as function of disease progression, there is clear need to shed light on spatial and temporal dynamics of changes within affected brain circuits.
Computational anatomy studies in TLE, using in vivo brain magnetic resonance imaging (MRI), demonstrate a specific pattern of cortical volume loss and changes in corresponding white matter pathways that extend beyond mesial temporal lobe structures (Bernhardt, Hong, et al., 2013;Keller & Roberts, 2008). Theoretical work and studies on animal models suggest that TLE-associated network remodeling follows a specific temporal trajectory (Leite et al., 2005;Sutula, 2004). Recent report provided empirical evidence for the assumption of bidirectional brain anatomy changes during disease progression with initial seizure-dependent boost in neurogenesis followed by gliosis due to depletion of hippocampal stem cells and shift toward astrocytes production (Sierra, Grohn, & Pitkanen, 2015;Sierra, Martin-Suarez, et al., 2015). In humans, cross-sectional (Bonilha et al., 2004) and longitudinal studies (Bernhardt, Kim, & Bernasconi, 2013) corroborated continuous changes affecting hippocampal volume loss in chronic TLE stages. Unpublished findings from our own group confirmed the notion of bidirectional hippocampus alterations and demonstrated hippocampus volume increase in early stages of TLE, followed by progressive atrophy of the hippocampus ipsilateral to seizure onset (Roggenhofer et al., 2019). The assumption of differential temporal dynamics of brain anatomy changes within the TLE network remains to be tested, particularly in relationship with individual clinical phenotype.
Compared to investigation of temporal trajectories of TLEinduced brain circuit changes, our knowledge in the spatial domain of network modulation remains very limited. Previous studies demonstrated strong links between clinical phenotype and functional brain network organization in the case of left-or right-lateralized TLE (Doucet, Osipowicz, Sharan, Sperling, & Tracy, 2013), however, comparable work in the field of brain anatomy lacks specificity. Current statistical analyses in the framework of computational anatomy describe the spatial pattern of TLE pathology without formally testing the question about interdependencies between spatially segregated structural findings. At present, it remains unclear to which degree a particular structure within the TLE specific network is involved in underlying pathological mechanisms and how these regional interdependencies evolve over time.
Main goal of our study is to investigate differential topology of brain anatomy changes within TLE circuits. We hypothesize that the ipsilateral hippocampus is the structure with strongest contribution to TLE-induced brain anatomy remodeling. Additionally, effects of TLE and hippocampal sclerosis. We interpret our results as evidence for TLE-related spatial modulation of anatomical networks.  (Roggenhofer et al., 2019). The protocol was approved by the local Ethics Committee. Informed consent was obtained from each participant. All procedures were performed in accordance with national and international guidelines.

K E Y W O R D S
The diagnosis of TLE followed the well-established criteria of the International League Against Epilepsy (Berg et al., 2010;Engel, 2006) including (a) clinical aspects of seizures like semiology, onset, and history, (b) standard and/or sleep electroencephalography with or without hyperventilation and intermittent photic stimulation additional to long-term video-electroencephalography (King et al., 1984) monitoring, and (c) neuro-radiological assessment.

| Magnetic resonance imaging and data processing
The MRI protocol consisted of T1-weighted images acquired on a  (Ashburner & Friston, 2005) using a novel set of brain tissue priors showing increased accuracy for subcortical structures (Lorio et al., 2016). Following this step, gray and white matter probability maps were spatially registered to a standardized Montreal Neurological Institute space using the diffeomorphic algorithm based on exponentiated lie algebra-DARTEL (Ashburner, 2007). The resulting gray matter probability maps were scaled with the corresponding Jacobian determinants to preserve the initial total amount of signal intensity followed by spatial smoothing using an isotropic Gaussian kernel of 8 mm full-width-at-half-maximum.

| Multivariate Bayesian analysis
The MVB statistics followed a mass-univariate analysis investigating the temporal dynamics of structural remodeling in TLE (Roggenhofer et al., 2019). We used whole-brain two-sample t tests on gray matter volume maps to demonstrate a pattern of volume differences and their modulation by side of seizure onset additional to presence or absence of hippocampal sclerosis. Here, we combined the significant clusters of the classical VBM approach separately for right and left laterality followed by binarization of the resulting pattern.
We used atlas information (probabilistic and maximum probability tissue labels-derived from the "MICCAI 2012 Grand Challenge and Workshop on Multi-Atlas Labeling" www.masi.vuse.vande rbilt. edu/works hop2012) to label anatomically distinct areas within the binarized VBM pattern. The anatomical labels as region-of-interest (ROI) provided the spatial constraints for multivariate decoding and Bayesian model selection.
Multivariate Bayesian was implemented to decode diseaserelated patterns from structural brain images (Friston et al., 2008).
This decoding framework relies on a model inversion using a variational Bayesian implementation of expectation maximization, to furnish the model evidence and the conditional density of the model's hyperparameters (Friston & Kiebel, 2009). Free energy is an information theory quantity derived from physics that bounds the log evidence of a model of data (Friston, 2009). The multivariate model allows drawing inferences about where and how TLE pathology is represented in the brain by evaluating competing anatomical coding hypotheses on a group level. Therefore, MVB provides the statistical evidence for each regional model that the structural data predict the pathology (Kherif & Muller, 2020). We emphasize that the objective of this MVB approach is not to predict the pre-or absence of epilepsy itself as the diagnosis is a known parameter but to enable inferences on distinct regional models and differentiate the individual degree of contribution to a pattern of dynamic epilepsy-associated brain remodeling.
We used 6 different designs as MVB models including two groups-healthy volunteers and a subgroup of TLE patients. The subgroups were defined based on clinical phenotype-laterality of seizure onset (left and right TLE) and presence (MTS) or absence (MRI−) of temporal lobe pathology. Bayesian model comparison implies the prior generation of multiple regional hypotheses (i.e. ROI or models) to be compared. As input for each Bayesian model, we use the voxel information based on gray matter probability maps within anatomically distinct atlas-defined ROI (see above "Magnetic resonance imaging and data processing"). The calculation of the model evidence permits Bayesian Model Comparison and selection (Penny, Flandin, & Trujillo-Barreto, 2007). Bayesian Model Selection (Penny et al., 2007;Stephan, Penny, Daunizeau, Moran, & Friston, 2009) is applied on the created models to compare different spatial hypotheses (Hulme, Skov, Chadwick, Siebner, & Ramsoy, 2014) using both, the model log evidence (Free Energy) and the parameter densities.
Regional models maximizing free energy are more likely to confirm the model-specific prediction of TLE disease and subtype pathology. Statically significant clusters at the group level were identified using the classical SPM mass-univariate approach. Each cluster was labeled according to their ROI.
We quantify the contribution of regions to the presence of diagnosis within the pattern of structural brain reorganization and

| Statistical analysis
Finally, we compared rankings based on multivariate pattern modeling to the established VBM approaches. For detecting group differences between left-or right-lateralized TLE and healthy volunteers, we applied a statistical threshold at p < .001, uncorrected. VBM rankings depended on statistical parametric mapping and T-map values, averaged across distinct anatomical structures, that is, ROIs or regional models. For visualization of T-statistics, we adopted the same ranking sequences like for MVB results (Figure 2). The multivariate statistics (model log evidences) and the mass-univariate statistics (averaged T-statistics) cannot be compared for the same ROI. Instead, we propose to compare the overall ranking of the ROIs.
In a first analysis, we computed the MVB models using four types of structural abnormalities representations: compact, sparse, smooth, and support representations for each parcellated ROI (Friston et al., 2008

| Main effects of disease
Using VBM, we defined the TLE-associated pattern and labeled the involved distinct anatomical structures. Based on MVB, we calculated free-energy parameters for each of these structures.
The topological MVB ranking determines that volume estimates in ipsilateral mesial temporal lobe regions most contribute to TLE (Figure 1a, b). The two most contributing regions cover the ipsilateral hippocampus and para-hippocampal gyrus independent of laterality. The

| Comparison between mass uni-and multivariate methods
To contrast the method with established volumetric morphology approaches, we paralleled the MVB-based ranking with a ranking based

| Structural remodeling in TLE subtypes
To evaluate regional specificity between TLE subtypes, MVB mod-

| D ISCUSS I ON
Our study provides unique empirical evidence for a differential contribution of brain regions to the process of brain anatomy remodeling in TLE. Using multivariate statistical methods allowing for topology inferences, we quantify the individual contribution of structures within limbic and extralimbic circuits to the TLE-associated brain anatomy pattern. We identify the ipsilateral hippocampal complex as main driver of spatial dynamics in TLE, whereas thalamus and a number of cortical areas show differential contribution depending on the laterality of the epileptogenic focus and mesial temporal lobe pathology. We interpret our findings as correlates of differential spatial trajectories within the TLE network that are also subjected to changes in due course of disease. The translation of our approach to clinical usage could provide a novel in vivo diagnostic tool for noninvasive localization of an epileptogenic focus, not detectable by conventional radiological assessment.

| Main effects of disease
Our main finding is that the hippocampal complex is the most discriminative region for TLE-related remodeling of brain anatomy.
Distinct morphology inferences dependent on hippocampal sclerosis corroborate that the hippocampal topology plays a decisive role in TLE-related spatial modulation of anatomical networks, supplementary to the previously described temporal dynamics. We observed a common pattern for TLE patients with mesial temporal lobe pathology that includes the ipsilateral hippocampus, parahippocampal gyrus, amygdala, and bilateral thalamus-all structures representative of the brain anatomy network implied in TLE  (Kemmotsu et al., 2011). Further, we demonstrate a decentralized, but less severely affected pattern in left TLE. These findings are supported by previous reports where patients with left TLE showed more widespread and diffuse abnormalities including cortical volume loss (Kemmotsu et al., 2011;Riederer et al., 2008) and changes in white matter fiber tracts (Ahmadi et al., 2009) extending to contralateral regions.
The asymmetry in topology patterns could be a consequence of hemisphere-specific rates of brain fiber tract maturation.
Quantitative and diffusion imaging corroborated that maturation occurs earlier and evolves quicker in the left than in the right hemisphere (O'Muircheartaigh et al., 2013) whereas global and local efficiencies are significantly decreased until early adulthood (Zhong, He, Shu, & Gong, 2016). In the context of quicker left-hemispheric development and less integrated connection with less efficient communication, the left hemisphere tends to be more susceptible to initial events like febrile convulsions and early onset seizures (Kemmotsu et al., 2011). Analogously, hippocampal sclerosis is observed more often in left rather than right hemisphere after febrile convulsions (Janszky et al., 2003). Correspondingly, age-related maturation of white matter is delayed in children with new-onset epilepsy (Chiron et al., 1997;Hutchinson et al., 2010). In our dataset, the age at the first seizure was not different between left-and right-lateralized TLE favoring the argumentation that rates of maturation differ across hemispheres. Furthermore, left hemispheric white matter connectivity in individuals with left-sided language dominance exhibit a more widespread pattern and comprises more complex hippocampal connections (Powell et al., 2007). This can provide a physiological network basis for a more diffuse and extensive left-lateralized seizure propagation underlying network deterioration.

| Differential regional contribution in TLE subtypes
Another important finding is the evidence for differential contribution of specific brain structures to the TLE pattern under the modulatory impact of presence or absence of mesial temporal lobe pathology. The ipsilateral hippocampus volume is highly contributory to group differences between healthy controls and all TLE subtypes except for right MRI− patients where thalamus, temporal polar cortex, and insula achieved highest ranking. Our results, showing a principal difference of regional contribution to brain remodeling depending on the presence or absence of mesial temporal lobe sclerosis, confirm the supposition that these represent the same nosological entity or lie along a biologic continuum (Labate, Cerasa, Gambardella, Aguglia, & Quattrone, 2008;Mumoli et al., 2013). We go further to claim that the evidence for differential regional contribution adds another layer of complexity linked to time-dependent modulation of brain structure in TLE. This is in line with our previous findings of a switch from hippocampus volume increases to decreases that could be interpreted as turning point with impact on remote regions within the TLE-affected network (Roggenhofer et al., 2019).
Preceding studies in MRI− TLE support the involvement of frontal and parietal cortices, especially the sensorimotor cortex mainly due to excitotoxicity induced neuronal loss (McDonald et al., 2008;Mueller et al., 2009).

| Methodological considerations
The descriptive comparison demonstrates the significant advantage in discriminative power when using multivariate instead of univari- ate methods to answer the question about differential regional contribution to brain remodeling in TLE. The MVB findings, confirming the hippocampus as main target of spatial remodeling, are in line with previous computational anatomy studies using univariate statistics (Bernhardt et al., 2008;Moran, Lemieux, Kitchen, Fish, & Shorvon, 2001).
The VBM analysis downgraded the role of the mesial temporal lobe structures in both right and left TLE and increased rankings in favor of cortical and subcortical structures known to be secondarily implicated in seizure. Indeed, one can infer generic disease-associated structural differences by established VBM approaches. But the method does not allow drawing causal inferences about inter-regional dependencies in brain anatomy. Disseminated patterns of structural reorganization are less likely to be identified in VBM approaches and inter-regional dependencies are not considered.
In particular, mesial temporal lobe regions vary highly anatomically across healthy individuals in gyral and sulcal shape and patterning with a convoluted cortical ribbon. Characteristic signs of TLE are an increased folding complexity in temporo-limbic cortices, hippocampal malrotations, and developmental anomalies Voets, Bernhardt, Kim, Yoon, & Bernasconi, 2011) in addition to disease progression-related atrophy (Roggenhofer et al., 2019).
Individual anatomical particularities and disease-dependent remodeling of medial temporal lobes can influence the spatial distribution of an affected structural pattern on the group level. In analogy, it was not possible to determine mesial temporal lobe structures as the most affected regions concerning structural differences between TLE and healthy volunteers in the VBM approach.
A multivariate approach can reveal information jointly encoded by several voxels as the multivariate distance between the two categories accounts for correlations among these. Extending classical inferences of mass-univariate analysis, the multivariate technique is particularly suited to quantify local changes in brain morphology and does not depend on a statistical threshold. Using MVB, we were able to answer the question which brain regions are jointly informative for the disease pathology and which anatomical structures allow to separate TLE patients and healthy controls. In this respect, the established multivariate method provides a reliable and robust ranking.

| Limitations and conclusion
We We present topology inferences of disease-related remodeling that highlights a key dissociation in the brain anatomy contributing to epilepsy pathology and healthy control conditions. Structural Bayesian modeling furnishes an appropriate framework for population-based analysis to identify regions being most important for disease-related structural remodeling and quantitatively estimate the involvement of distinct structures in spatially restricted morphological remodeling in focal epilepsy. An individualized focus identification can provide preoperative clinical benefits for targeting electrodes used for neurostimulation therapy and is relevant to guide and monitor surgical intervention, especially in the context of increased use of minimally invasive approaches, such as MRI-guided thermal ablation.

ACK N OWLED G M ENTS
We would like to thank the participants for their participation in the study. BD and ER are supported by The Swiss National Science for their generous financial support.

CO N FLI C T O F I NTE R E S T
Nothing to report.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/brb3.1825.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data available on request due to privacy/ethical restrictions.