Strategy‐based reasoning training modulates cortical thickness and resting‐state functional connectivity in adults with chronic traumatic brain injury

Abstract Introduction Prior studies have demonstrated training‐induced changes in the healthy adult brain. Yet, it remains unclear how the injured brain responds to cognitive training months‐to‐years after injury. Methods Sixty individuals with chronic traumatic brain injury (TBI) were randomized into either strategy‐based (N = 31) or knowledge‐based (N = 29) training for 8 weeks. We measured cortical thickness and resting‐state functional connectivity (rsFC) before training, immediately posttraining, and 3 months posttraining. Results Relative to the knowledge‐based training group, the cortical thickness of the strategy‐based training group showed diverse temporal patterns of changes over multiple brain regions (p vertex < .05, p cluster < .05): (1) increases followed by decreases, (2) monotonic increases, and (3) monotonic decreases. However, network‐based statistics (NBS) analysis of rsFC among these regions revealed that the strategy‐based training group induced only monotonic increases in connectivity, relative to the knowledge‐based training group (|Z| > 1.96, pNBS < 0.05). Complementing the rsFC results, the strategy‐based training group yielded monotonic improvement in scores for the trail‐making test (p < .05). Analyses of brain–behavior relationships revealed that improvement in trail‐making scores were associated with training‐induced changes in cortical thickness (p vertex < .05, p cluster < .05) and rsFC (p vertex < .05, p cluster < .005) within the strategy‐based training group. Conclusions These findings suggest that training‐induced brain plasticity continues through chronic phases of TBI and that brain connectivity and cortical thickness may serve as markers of plasticity.


| INTRODUCTION
A traumatic brain injury (TBI) is caused by external force (e.g., blast or fall) to the head, leading to disruptions in brain structure and function. TBI is a substantial threat to public health in the United States, contributing to 30% of all injury deaths (Faul, Xu, Wald, & Coronado, 2010). In 2010, 2.5 million emergency room (ER) visits, hospitalizations, or deaths were associated with TBI in the United States (http:// www.cdc.gov/traumaticbraininjury/get_the_facts.html). Individuals at the chronic stage of TBI (>6 months postinjury) often have cognitive difficulties such as problems with attention, memory, or executive functions, which are critical to carrying out daily life tasks (Arciniegas, Held, & Wagner, 2002;Rabinowitz & Levin, 2014). As the devastating effects of TBI-related cognitive challenges persist in the lives of millions of individuals, a recent Center for Disease Control and Prevention (CDC) report to Congress emphasized the need for more TBI rehabilitation research that will improve the lives of individuals with TBI and their families (Frieden, Houry, & Baldwin, 2015).
Neuropsychological tests have also been criticized for the lack of ecological validity, meaning the test scores do not adequately reflect daily function levels (Burgess et al., 2006).
Diffuse axonal injury (DAI) (Smith, Meaney, & Shull, 2003) is a primary injury mechanism of TBI, and rsFC has been a promising technique to identify DAI-induced higher-order cognitive impairments (e.g., reasoning, decision making, and selective attention) among individuals with TBI (Sharp, Scott, & Leech, 2014). Notably, large-scale network approaches using rsFC in TBI have effectively revealed the complex patterns of the injured brain (Han, Chapman, & Krawczyk, 2016;Mayer, Mannell, Ling, Gasparovic, & Yeo, 2011;Sharp et al., 2011). The large established literature in large-scale networks in healthy individuals facilitates the interpretation of study findings as applied to individuals after a TBI.
In our previous study, we reported the efficacy of strategy-based reasoning training for chronic TBI (Vas et al., 2016). Although that study contributed to the literature in rehabilitation for TBI, it primarily focused on assessing the neuropsychological performance of individuals with chronic TBI following the strategy-based reasoning training.
Given the limited sensitivity and heterogeneity of neuropsychological test scores of individuals with chronic TBI, assessing training-induced changes in the brain utilizing advanced neuroimaging techniques such as morphometry and rsFC may provide better sensitivity and higher precision. Furthermore, in light of previous morphometry and rsFC studies in TBI, the use of these methods could enable us to better understand the underlying mechanisms of training-induced changes in individuals with chronic TBI.
There is currently a limited literature addressing changes in the brain following rehabilitation for chronic TBI. Thus, multimodal approaches combining both morphometry and rsFC would strengthen the efforts to elucidate potential mechanisms of training-induced neuroplasticity of individuals with chronic TBI. While changes in morphometry following training can confirm that the adjustment of behavior modulates brain structure, the temporal patterns of training-induced changes in morphometry are often complex. As such, the directionality of these changes has varied across previous studies (Draganski et al., 2006;Maguire et al., 2000;Metzler-Baddeley, Caeyenberghs, Foley, & Jones, 2016;Taubert et al., 2010). RsFC can complement morphometry to better understand the complex patterns of changes occurring after training. RsFC can also address a question whether spatially distributed morphometric changes are also accompanied by a reorganization of the architecture of functional networks. For example, Taubert et al. (2011) have reported changes in gray matter volumes using a centrality measure of rsFC (i.e. strength of connectivity of a voxel with rest of the voxels) after balance training in the left supplementary and presupplementary motor areas, indicating that localized training-related changes in the gray matter accompany changes in rsFC over distributed brain regions in healthy adults. The question whether spatially diffuse changes in morphometry are accompanied by reorganization of networks is also critical in the context of TBI as (1) network dysfunction is one of the key mechanisms that explains impairments in high-order cognitive functions following TBI (Sharp et al., 2014) and (2) a TBI markedly disrupts between-network connectivity of the brain, yielding less efficient brain communication after an injury (Han et al., 2014. These characteristics of TBI suggest that it would be interesting to determine whether structural changes following training accompany network reorganization in individuals with TBI. Taken together, identifying changes in rsFC in conjunction with morphometry of individuals with chronic TBI following training would improve our understanding of training-induced neuroplasticity in chronic TBI. In this study, we utilized structural MRI and resting-state functional MRI (rsfMRI) to identify the effects of a strategy-versus a knowledgebased training on cortical thickness and rsFC within individuals at the chronic stage of TBI. Specifically, we randomized individuals with chronic TBI into two 8-week training groups (strategy based vs. knowledge based), and we acquired their MRI scans over three time points (prior to training, after training and at 3-month follow-up after training ended). We then investigated the spatial and temporal patterns of training-induced changes in cortical thickness and rsFC within these individuals. Based on findings from previous studies in neuroplasticity in healthy adults and reported efficacy of strategy-based reasoning training for chronic TBI, we hypothesized that the strategy-based reasoning training method would induce changes in cortical thickness relative to the knowledge-based training. We further hypothesized that the reasoning training would also induce changes in rsFC among regions where cortical thickness changes in chronic TBI cases.

| Participants
The data included in these analyses are a part of an ongoing study (Krawczyk et al., 2013). We analyzed 60 individuals at the chronic stage of TBI who ranged from lower moderate disability to lower good recovery (age 20-65; >6 months postinjury; 5-7 on the Extended Glasgow Outcome Scale (GOS-E; Wilson, Pettigrew, & Teasdale, 1998), who completed MRI scans that passed the quality assurance (QA) procedures described below. We recruited these participants from the Dallas-Ft. Worth community and conducted a phone screening interview before inclusion in the study. The primary causes of TBIs in this group were blasts, blunt force trauma, falls, athletic impacts, vehicle accidents, or combinations thereof. Note that, given the number of years postinjury time, it was not feasible to obtain participants' clinical information on initial injury characteristics such as Glasgow coma scale (GCS; Teasdale & Jennett, 1974) from the inpatient, acute-care facilities where they were hospitalized several years ago. Therefore, initial injury severity was retrospectively estimated utilizing the Ohio State University TBI identification (OSU TBI-ID) method (Corrigan & Bogner, 2007). The OSU TBI-ID method has good interrater reliability and test-retest reliability (Bogner & Corrigan, 2009;Corrigan & Bogner, 2007). Furthermore, the OSU TBI-ID method estimates initial injury severity based on participants' recollections of the incidents including the estimated duration of loss of consciousness (LOC), and the CDC guidelines for the conceptual definition and identification of TBI (National Center for Injury Prevention and Control, 2003;Thurman, Sniezek, Johnson, Greenspan, & Smith, 1995). Specifically, the participants whose estimated LOC duration <30 min, <24 hr, or >24 hr were considered to be probable mild, probable moderate, or probable severe TBI, respectively. Both civilian and veteran participants were included (See Table 1 for demographics). No participants had a history of any significant, clinically diagnosed neurological or psychiatric comorbidities. We also confirmed that all participants' brains did not show visible focal lesions, contusions, mass shifting, or extreme degeneration of white matter on structural MRI scans (see Figure S1 for an example scan that was not included in this study due to white matter degeneration). This confirmation should minimize the potential effects of such macrostructural injuries on preprocessing for cortical surface reconstruction and rsFC analyses. All participants provided written informed consent, and this study was conducted in compliance with the declaration of Helsinki. The study was approved To prevent any potential, systematic effects of out-of-scanner variables (such as injury characteristics and other demographics) on our findings, we randomly assigned all participants into either the SMART or BHW groups. Both training programs comprised of 12 sessions (1.5 h per session) for 8 weeks with quizzes, homework assignments, and projects conducted in small group settings, comprising of 4-5 participants. Briefly, the SMART group focused on selective attention, abstract reasoning, and other thinking strategies (Vas, Chapman, Cook, Elliott, & Keebler, 2011), and the BHW group focused on education regarding brain structure and function and the effects of sleep and exercise on the brain performance (Binder, Turner, O'Connor, & Levine, 2008). More specifically, the SMART participants were trained to (1) manage information by blocking distractions and irrelevant information and avoid multitasking, (2) increase the ability to understand main ideas and take-home messages from information, and (3) examine information from different perspectives. This set of strategies is aimed to improve cognitive control and enhance information processing (goal management). The SMART strategies are introduced in slides presented by a trained clinician. Each of the strategies was sequentially introduced and then reinforced throughout the training sessions. Example materials to practice the strategies included newspaper articles and audio-video clips. The BHW participants learned about brain anatomy, brain function, the effects of a TBI on cognitive function, the principles of neuroplasticity, and the impact of diet, physical exercise, sleep, and social activities on brain health through slides taught by a clinician. The participants were also encouraged to discuss applications of learned information to their daily lives. To control for the effects of group-based social activities on training outcomes, we

| Neuropsychological assessments
We administered a battery of neuropsychological tests to measure training-induced cognitive changes in a variety of domains for the TBI subgroups. As the SMART is aimed at improving multiple domains of cognitive functions of individuals with TBI, we did not focus narrowly on assessments. Rather, we administered a battery of neuropsychological tests probing executive functions of the participants to enable a more exploratory analysis. These tests include full-scale intelligent quotient-2 (FSIQ-2) from the Wechsler Abbreviated Scale of Intelligence (WASI) for estimated current IQ (Wechsler, 1999), FSIQ from the Wechsler Test of Adult Reading (WTAR) for estimated premorbid IQ (Wechsler, 2001), and color-word, verbal fluency, card sorting, trail making from the Delis-Kaplan Executive Function System (D-KEFS) for inhibitory control, switching, verbal fluency, processing speed, and problem solving (Delis, Kaplan, & Kramer, 2001). We also acquired tests that measure psychiatric symptoms of the individuals with TBI as there are relatively common concomitant psychiatric symptoms after TBI (Ashman et al., 2004;Hibbard et al., 1998;van Reekum et al., 1996). To quantify subclinical-but-residual depressive and posttraumatic stress disorder (PTSD) symptom severity of the participants, we measured the Beck Depression Inventory-II (BDI-II) (Beck, Steer, & Brown, 1996) and PTSD Check List Stressor-specific (PCL-S) (Weathers, Litz, Herman, Huska, & Keane, 1993). We also administered the satisfaction with life scale (Diener, Emmons, Larsen, & Griffin, 1985) to measure global cognitive judgments of the participants' life satisfaction.

| MRI data acquisition
We acquired MRI scans of the participants at three time points: prior to training (TP 1 ), after training (TP 2 ), and 3 months later (TP 3 ). Participants

| Cortical surface reconstruction and cortical thickness measurement
We reconstructed the cortical surface from each of the MRI scans with FreeSurfer v.5.3.0 (RRID:SCR_001847; http://surfer.nmr.mgh. harvard.edu/). The technical details of cortical surface reconstruction procedures have been described elsewhere Fischl, 2012;. After we obtained the gray/white boundary and pial surface estimations (Dale & Sereno, 1993;Dale et al., 1999), cortical thickness was calculated as the closest distance between these surfaces at each vertex across the cerebral cortex (Fischl & Dale, 2000). The cortical surfaces were reconstructed using spatial intensity gradients across tissue classes, thus the measurements were not simply dependent on absolute signal intensity. Note that the surface-based cortical thickness maps were not restricted by the voxel size, which enabled us to detect changes in cortical thickness at submillimeter level. Furthermore, procedures to measure cortical thickness using Freesurfer have been validated against histological analysis (Rosas et al., 2002) and manual measurements (Kuperberg et al., 2003;Salat et al., 2004). Lastly, Freesurfer morphometric procedures have been demonstrated to show good test-retest reliability across scanner manufacturers and field strengths (Han et al., 2006;Reuter, Schmansky, Rosas, & Fischl, 2012).

| Longitudinal analysis of cortical thickness
We utilized the longitudinal processing stream (Reuter et al., 2012) in FreeSurfer to obtain a reliable longitudinal analysis of cortical thickness. Specifically, an unbiased within-subject template was created using robust, inverse-consistent registration (Reuter, Rosas, & Fischl, 2010). Several processing steps, such as skull stripping, Talairach transforms, atlas registration, and spherical surface maps and parcellations, were initialized with common information from the within-subject template. This procedure yielded significant increases in reliability and statistical power (Reuter et al., 2012). This longitudinal processing pipeline also prevented potential bias with respect to any specific point, which is an important issue in longitudinal studies evaluating training-related structural changes within the brain (Thomas & Baker, 2013). For group analysis, we resampled cortical thickness for each of the scans on a standard template, followed by surface smoothing with a 10-mm full-width-at-half-maximum Gaussian kernel.
With the preprocessed longitudinal data, we performed the linear mixed effects model (LME; Bernal-Rusiel, Greve, Reuter, Fischl, & Sabuncu, 2013) analysis using a piece-wise linear model with a breakpoint at TP 2 and a randomly varying intercept. Specifically, the cortical thickness of subject i at time point j, y ij , can be written as: where t ij is the time of measurement for subject i at time point j, t is an average time of measurement at TP 2 , S i is an indicator function for the SMART group for subject i, b i is a subject-specific intercept (cortical thickness of subject i at TP 1 ), A i is the age for the subject i, B is the average BDI score over time for the subject i, B ij is the BDI score for the subject i at time point j, e ij is the measurement error for subject i at time point j, and H(·) represents the Heaviside step function. Previous studies reported the effects of comorbid psychiatric conditions on cortical thickness of individuals with TBI (Hudak et al., 2011;Lindemer et al., 2013). Thus, we controlled for potential effects of depressive symptoms on cortical thickness by including betweenand within-subject BDI covariates in the model. We did not include PCL-S covariates in the model, as BDI and PCL-S scores were highly correlated.
We utilized the LME as opposed to the repeated measures analysis of variance (ANOVA) because the repeated measures ANOVA is less optimal for analyzing longitudinal data. A repeated measures ANOVA does not properly model the covariance structure of serial measurements when the covariance among repeated measurements and variance are not constant across time (i.e., the violation of the compound symmetry assumption). Furthermore, the repeated measures ANOVA cannot handle other common data characteristics present in a longitudinal study. These include nonuniform data acquisition timing from the baseline across datasets and subject attrition. The LME is an extension of the general linear model to handle longitudinal data by incorporating subject-specific random factors into the model. The LME provides both flexible and parsimonious models for the covariance, and it is more appropriate for longitudinal data with increased specificity, sensitivity, and reliability than other alternatives (Bernal-Rusiel et al., 2013;Chen, Saad, Britton, Pine, & Cox, 2013).
We performed subsequent statistical inferences for the within-and between-group contrasts of cortical thickness at (1) TP 2 relative to TP 1 (i.e., H 0 : y(TP 2 )-y(TP 1 ) = 0), (2) TP 3 relative to TP 2 (i.e., H 0 : y(TP 3 )−y(TP 2 ) = 0), (3) monotonic changes over three time points (i.e., H 0 : (y(TP 2 )−y(TP 1 ))+(y(TP 3 )−y(TP 2 )) = y(TP 3 )−y(TP 1 ) = 0)), and (4) nonmonotonic changes over three time points (i.e., H 0 : (y(TP 2 )− y(TP 1 )) + (y(TP 2 )−y(TP 3 )) = 0), respectively. See Figure 1 for illustrations of monotonic and nonmonotonic changes. We identified statistically significant training-induced temporal changes in cortical thickness from the between-group contrast over all time points (i.e., monotonic or nonmonotonic changes) at p vertex < .05 and p cluster < .05. To determine which group(s) drove statistically significant between-group differences, we then used the results of the within-group contrasts for changes over all three time points. If the corrected version of the within-group contrast results did not clearly reveal which group(s) accounted for the observed betweengroup differences, we further assessed an uncorrected version (p vertex < .05) of the within-group contrast results. We assessed contrasts for changes in two successive time points to confirm the patterns of training-related change over all time points.

| Post hoc regions-of-interest analysis of cortical thickness
To confirm that the statistically significant changes in cortical thickness following training were not occurring due to outliers, we identified training-induced cortical thickness trajectories at the single-subject level. More specifically, we obtained an average cortical thickness of each participant within each of the cortical regions where statistically significant between-group differences in overall temporal changes in cortical thickness occurred at the vertex level. Within each of these cortical regions, we also counted how many participants showed significant changes in cortical thickness compared to the other training group. We defined a significant training-induced change in cortical thickness in a given region for a participant if the amount of cortical thickness change fell outside of the 2 SD band obtained from the distribution of cortical thickness changes for the other training group.
Note that, in this analysis, we did not perform any additional statistical tests (thus no subsequent p-values) for the regions identified by the group analysis as those statistical tests could yield a selection bias.

| Volumetric rsfMRI preprocessing
Volumetric rsfMRI data were preprocessed in a subject-native space using a modified version of a shell script generated by afni_proc.py (http://afni.nimh.nih.gov/pub/dist/doc/program_help/afni_proc. py.html) from AFNI [RRID:SCR_005927; Cox, 1996). Each subject's whole-brain structural images were first skull stripped and registered (affine transform with 12 parameters) to the initial frame of the first rsfMRI run. For each rsfMRI run, the initial four frames were discarded to allow T 1 magnetization saturation. Standard preprocessing methods were then applied, including despiking, slice timing correction, motion correction, spatial resampling (4 mm isotropic), normalization to whole-brain mode of 1000, band-pass filtering (0.009 < f < 0.08 Hz), and linear regression. At the motion correction stage, the six rigid body motion profiles were obtained for the linear regression. In the linear regression, the rsfMRI time series were third-order detrended, and several sources of signal fluctuation unlikely to be of neuronal origin were regressed out as nuisance variables: (1) six parameters for rigid body head motion acquired from the motion correction (Johnstone et al., 2006), (2) the signal averaged over the lateral ventricles (Fox et al., 2005), (3) the signal averaged over a region centered in the deep cerebral white matter (Fox et al., 2005), and (4)

| Surface mapping
We used SUMA (Saad, Reynolds, Argall, Japee, & Cox, 2004) for surface mapping and surface standardization. The preprocessed volumetric functional time series were projected onto mesh surfaces of each subject by averaging the time series across the voxels belonging to five equally spaced coordinates along the line between two matching nodes of 1 mm inside of the white surface and the outside of the F I G U R E 1 Illustrations of monotonic (a) and nonmonotonic (b) changes over three time points. H 0 represents the null hypothesis of statistical inferences for contrasts at monotonic and nonmonotonic changes over three time points in the linear mixed model, respectively. TP 1 , prior to training; TP 2 , after training; TP 3 , 3 months after training t y(t) H 0 : (y(TP 2 )-y(TP 1 )) + (y(TP 3 )-y(TP 2 ))=0

Monotonic changes
H 0 : (y(TP 2 )-y(TP 1 )) + (y(TP 2 )-y(TP 3 ))=0 t y(t) pial surface. In the same way, the voxel-based subject masks were converted into surface-based subject masks. For group analysis, we standardized surface meshes (36,002 nodes per hemisphere with a 2-mm spatial resolution) of each individual. We also obtained an intersection mask for the mesh nodes in which fMRI signals existed across all subjects and time points.

| Network-based statistics of longitudinal rsfMRI data
We obtained a connectivity matrix for each of the scans by calculating the Pearson correlation coefficients for average time series from each of the ROIs where statistically significant between-group differences in overall temporal changes in cortical thickness occurred. After performing a Fisher's Z-transform to ensure the normality of correlations and scaling to Z-scores (i.e., zero mean and unit variance), we performed the LME analysis with the same model used for the cortical thickness analysis, additionally including between-and within-subject FD covariates. The FD covariates were included because the LME analysis on (1) the percentage of motion censored volumes and (2) FD after censoring and trimming revealed statistically significant changes in FD from TP 2 to TP 3 and from TP 1 to TP 3 (nonmonotonic) within SMART (Table   S1). Statistically significant changes in connectivity were identified at

| Analysis of connectivity patterns according to large-scale networks
To further understand training-induced functional connectivity among the specified ROIs from the perspective of large-scale networks, we first determined which were the most affiliated resting-state networks to the ROIs. We used the seven network version of the resting-state cortical parcellation , which is a large-scale networkbased atlas of the cerebral cortex. The seven large-scale networks from the Yeo atlas include the visual, somatomotor, dorsal attention, salience (or ventral attention), limbic, frontoparietal, and default networks. All of these networks are frequently reported in the rsFC literature. We overlaid the ROIs obtained from group comparisons of training-induced changes in cortical thickness over time onto the Yeo atlas to identify which ROI affiliates with which large-scale network.
We then identified the number of connections with statistically significant group contrasts according to connectivity within and between the resting-state networks that each of the ROIs was affiliated with.
We also assessed Euclidean distance of these connections according to connectivity within and between the resting-state networks that each of the ROIs was affiliated with.

| Analysis of effect sizes
As there are no standard effect size statistics for the LME model, we ates that were included in the LME analyses (e.g., the BDI covariates).

| Quality assurance
We visually inspected all structural MRI scans to ensure that subjects had no significant brain atrophy. In rsfMRI preprocessing, the quality of the preprocessed data was visually inspected at each step.
After motion 'scrubbing', we confirmed that the total time of remaining frames after the 'scrubbing' exceeded 4 min, the minimum length required to reliably estimate rsFC (van Dijk et al., 2010). We also ensured that there were no MRI scans or neuropsychological measures that were acquired too late (i.e., outside the 2 SD band from the mean) for all time points. Lastly, we excluded MRI scans from the LME analysis when corresponding BDI scores were not available. See Table 2 for the number of MRI scans after the QA procedure.

| Cortical thickness versus neuropsychological performance
To confirm whether improved neuropsychological performance after training was related to changes in cortical thickness, we performed the LME analysis with a modified model by additionally including within-and between-subject factors from the numberletter switching versus motor speed scores of the trail-making test from the D-KEFS. We chose the trail-making test scores because the SMART group showed statistically significant (p < .05) improvement on this test relative to the controls (see the results section).
We adjusted for age, years of education, estimated current IQ, and BDI scores before we included the covariates to minimize potential effects of these measures on the trail-making test scores. As we were interested in how and where training-induced changes in cortical thickness were associated with improved neuropsychological performance within subjects, we focused on assessing the withinsubject covariate for the trail-making test in the revised LME model.
As temporal patterns of cortical thickness change within the SMART group were often nonmonotonic (i.e., returning to the baseline at TP 3 ; see the results section), we also assessed associations between nonmonotonic changes in cortical thickness and improvement in the trail-making test scores. Specifically, we performed the LME analysis on nonmonotonic components of cortical thickness by reflecting cortical thickness at TP 3 over the axis of cortical thickness at TP 2 (see Figure S2 for more details). This made the reflected thickness at TP 3 artificially made overall changes appear monotonic when actual temporal pattern of changes was nonmonotonic. As with the cortical thickness analysis, we identified statistically significant associations between changes in the two measures at p vertex < .05 and p cluster < .05.

| Resting-state functional connectivity versus neuropsychological performance
We also performed the LME analysis on seed-based connectivity additionally including within-and between-subject covariates from the number-letter switching versus motor speed scores of the trailmaking test from the D-KEFS in the model. We defined seeds as the regions that showed statistically significant associations between changes in cortical thickness and improvement in trail-making scores (refer to the result sections). The LME analysis was then performed for each of the seeds. We identified statistically significant associations between changes in the two measures at p vertex < .05 and p cluster < .005 (=.05/10) by additionally correcting for the number of seeds. between the brain and behavior in the context of cognitive training after TBI. As such, we did not correct for multiple comparisons across the neuropsychological measures. Our previous study (Vas et al., 2016) reported the efficacy of the SMART for chronic TBI based on neuropsychological assessment with larger sample sizes, which was the primary focus of that prior study.

| Demographics
All participants were at a long-term chronic phase of TBI (approximately 8 years postinjury time on average). There were no statistically significant group differences in age, estimated current and premorbid IQs, gender, proportions of civilians and veterans, postinjury time, distribution of estimated initial injury severity, or distribution of primary injury types at α = 0.05 (Table 1). However, group differences in years of education and estimated current IQ were marginally significant.

| Neuropsychological measures
The average times of assessments were 9 and 18 weeks from the baseline, respectively (Table 2). Although we attempted to match injury characteristics and other demographics between the two training groups by randomized group assignment, statistically significant (p < .05) group differences were observed in performance at baseline in test scores from the D-KEFS (

| Cortical thickness analysis results
The QA procedure allowed us to include 148 structural MRI scans from 58 participants (N = 30 for SMART and 28 for BHW) in the cortical thickness analysis ( Table 2). The average timing of MRI scans that passed the quality assurance procedures were 9 and 21 weeks from the baseline, respectively (Table 2).

| Whole-brain, group analysis results
The whole-brain, LME analysis demonstrated temporal changes in cortical thickness for the training groups over three time points ( Figure 2). Overall, the patterns of group differences in monotonic and nonmonotonic temporal changes in cortical thickness were distributed over the whole brain. Specifically, statistically significant (p vertex < .05, p cluster < .05) group differences in nonmonotonic (in-

| Post hoc ROI analysis results
The group analysis results yielded 12 ROIs (Figure 2, Table 5). We further identified individual patterns of cortical thickness changes within these ROIs. Overall, patterns of changes in cortical thickness at the single-subject level were consistent with the observed patterns at the group level, confirming that observed patterns at the group level were not due to outliers (Figures 4 and S4). For example, cortical thickness in the R DPFC of the SMART participants was increased then decreased back to the baseline, whereas temporal changes in cortical thickness of the BHW participants in this region were stable over time ( Figure 4a). More specifically, from TP 1 to TP 2 six SMART participants showed changes in cortical thickness in R DPFC that exceeded the 2 SD band from the BHW group in this region. From TP 2 to TP 3 two of these six SMART participants exceeded the 2 SD band from the BHW group in this region. Across the three time points, four of these six SMART participants showed nonmonotonic changes in cortical thickness greater than the 2 SD band from the BHW group. Note that one of these six SMART participants did not undergo an MRI scan at TP 3 .
None of the participants would be expected to be outside of the 2

| The sizes of training-related effects
The SMART-induced changes in scaled scores for the trail-making number-letter switching versus motor speed from 8.2 at TP 1 to 9.8 at TP 3 on average. These scaled scores at TP 1 and TP 3 correspond to 25% and 50% in percentile ranks, respectively. This indicates that the SMART group participants showed average performance on the trail-making test compared to normative samples. SMART induced 0.04-0.13 mm of cortical thickness changes within the regions that showed statistically significant between-group contrasts (Table 5).
These change ranges correspond to 1.8~5.2% change compared to baseline thickness. SMART also yielded training-induced changes in rsFC, ranging from 1.0 to 2.8 in Z-scores and from 29.1 to 172.6% ( Figure 8).

| Training-induced changes in the brain versus improvement in neuropsychological performance
We observed statistically significant (p vertex < .05, p cluster < .05) associations between training-induced changes in cortical thickness and improvements in neuropsychological performance after the SMART ( Figure 9a, Table 6). These associations occurred with both monotonic and nonmonotonic changes in cortical thickness after SMART.
Statistically significant associations between changes in these two measures did not occur with the controls. Note that the SMART group showed statistically significant (p < .05) monotonic improvement in scores on the trail-making number-letter switching versus motor speed, whereas the control group showed trends in reduced scores over time. Thus, positive association in the given region shown in Figure 9a indicates that the participants with the improved trailmaking scores after the SMART showed monotonic increases or increases then decreases in cortical thickness in that region. Negative association indicates that improvement in the test scores was associated with monotonic decreases or decreases followed by increases in cortical thickness in the given region within the SMART group.
Trajectories of cortical thickness and the trail-making test scores over time (Figure 9b-c) confirmed the observed group analysis results.
Analyses of rsFC with seeds identified from the results of the cortical thickness analysis from SMART exhibited the patterns of connectivity that were associated with the improved trail-making test scores (p vertex < .05, p cluster < .005; Figure 10). Within the SMART group, positive associations between the two measures indicate that participants with training-induced improvement in the trail-making scores showed monotonically increased rsFC between the given seed and region. Note that most of the temporal patterns of associated functional connectivity were monotonic increases. These patterns of associations even occurred in seed regions where changes in cortical thickness were nonmonotonic (e.g., the right lateral occipital lobe seed).

| DISCUSSION
We demonstrated changes in cortical thickness and resting-state connectivity of individuals at chronic stage of TBI following strategybased reasoning training. To our knowledge, this is the first MRIbased study to report brain plasticity as measured by both cortical thickness and rsFC in chronic TBI following cognitive training. Novel methods employed in this study included comparing individuals undergoing strategy-based training with an active control condition matching on all factors except training content. We assessed groupby-time interaction effects, considered to be the gold standard evidence for training-related neuronal and behavioral changes (Thomas & Baker, 2013). Furthermore, we employed an LME analysis to account for nonuniform data acquisition timing across the participants and participant dropouts, which frequently occurs in longitudinal studies.

| Changes in cortical thickness following training
Our cortical thickness findings supported our hypothesis that training could induce changes in brain morphometry of individuals even at the chronic stage of TBI. These findings extend a line of research in structural brain plasticity in clinical populations (Alosco et al., 2015;McEwen et al., 2015;Reiter et al., 2015;Sehm et al., 2014) by demonstrating training-induced changes in cortical thickness in the context of cognitive training after TBI. Another novel component of our study is that we demonstrated structural plasticity following cognitive reasoning training, adding to similar evidence following physical activity (Alosco et al., 2015;McEwen et al., 2015;Reiter et al., 2015) and balance training (Sehm et al., 2014  Thus, even though a TBI alters morphometry of the gray matter, the injured brain at the chronic stage of TBI responds to cognitive training by exhibiting dynamic changes in cortical thickness (Figures 2-5 and 9). The need for effective rehabilitative trainings for individuals with chronic TBI is growing (Frieden et al., 2015). In this context, our findings on changes in cortical thickness following strategy-based reasoning training for individuals with TBI suggest that cortical thickness may be a sensitive measure to identify training-induced neuroplasticity in TBI. However, measuring cortical thickness alone may not be sufficient to evaluate the efficacy of neurorehabilitation for TBI as half of the changes in cortical thickness over time within the SMART group were transient and the controls also showed changes in cortical thickness (Table 5).
The spatial patterns of training-related changes in cortical thickness were specific to the type of training that was carried out (Figures 2-5 and S3-S5, Table 5). As such, regions with statistically significant (p vertex < .05, p cluster < .05) changes in cortical thickness of the SMART and BHW groups did not overlap (Figures 2-3 and   S3). Among the 12 regions identified in the group analysis, statistically significant between-group differences in changes in cortical thickness occurred by within-group changes that occurred in 10 regions in the SMART group compared to 4 regions in the BHW group (Figures 2-3 and S2, Table 5). SMART training is targeted at improving higher-order cognition (i.e., executive functions), and is likely to involve different cognitive processes than the BHW.
SMART involves strategic attention, abstract reasoning, and cognitive flexibility (Vas et al., 2016), whereas the BHW primarily involves learning and memory. A previous study carried out in healthy individuals (Metzler-Baddeley et al., 2016) reported that changes in cortical thickness occurred in three regions after highly taxing working memory training compared to change in one region after less-taxing working memory training. Taken together, a greater number of regions exhibited cortical thickness changes over time after SMART training compared to the BHW group, suggesting that strategybased reasoning training programs may induce more spatially extensive gray matter structural changes than simply engaging in new learning. Temporal patterns of cortical thickness changes were also training specific. However, half of the changes in cortical thickness over time within the SMART group were transient (Table 5)  F I G U R E 3 Colormaps for within-and between-group contrasts for changes in cortical thickness from TP 1 to TP 2 (a) and from TP 2 to TP 3 (b). The colormaps were thresholded at p vertex < .05 and p cluster < .05. See Figure 2 for the abbreviations  '↑ then ↓', '↓ then ↑', '↑', '↓', and '−' symbols indicate increases then decreases, decreases then increases, monotonic increases, monotonic decreases, and no significant changes, respectively.
F I G U R E 4 Trajectories of cortical thickness over time and scatter plots for cortical thickness changes in the right dorsal prefrontal cortex (a), right anterior medial prefrontal cortex (b), right postcentral gyrus (c), and left lingual gyrus (d). Within each panel, the top row demonstrates trajectories of cortical thickness over time for each of the participants from the SMART (left) and BHW (right) groups. The other two rows represent scatter plots for cortical thickness changes from TP 1 to TP 2 (middle left) and from TP 2 to TP 3 (middle right), and monotonic (i.e., from TP 1 to TP 2 plus from TP 2 to TP 3 ; bottom left) and nonmonotonic (i.e., changes from TP 1 to TP 2 minus changes from TP 2 to TP 3 ; bottom right) changes over all time points, respectively. The I bars indicate the means and standard deviations of the BHW (a-c) and SMART (d) groups, the dotted (solid) horizontal bar is the 2 SD from the mean of the BHW (SMART) group, and the solid horizontal bars in the SMART (a-c) and BHW (d) are the means of the SMART (a-c) and BHW (d), respectively. Filled triangles (circle) represent TBI individuals from the SMART (BHW) with "significant" changes in cortical thickness, located outside the dotted (solid) horizontal bars (see the Methods section for more details of the term "significant"). See Figure 2 for the other abbreviations and locations of the regions

Trajectories (SMART) Trajectories (BHW) Trajectories (SMART) Trajectories (BHW)
Interestingly, the majority of the training-induced changes in cortical thickness occurred within the sulcal span at sulcal banks ( Figures 5 and S5). More specifically, increases in cortical thickness were accompanied by decreases in sulcal span, and decreases in cortical thickness were accompanied by increases in sulcal span. Although this phenomenon has not previously been documented in studies of training-related brain plasticity, our finding is in line with previous reports on inverse relationships between cortical thickness and sulcal span in adolescence (Aleman- Gomez et al., 2013) and aging populations (Kochunov et al., 2008).
However, it remains unclear whether the phenomenon observed in a subset of participants ( Figure 5) occurred in the other participants. Thus, further studies will be required to quantitatively assess sulcal span and potentially other morphometric measures at the group level.
Within the SMART group, temporal patterns of changes in cortical thickness varied among brain regions (Figures 2-5 and S3-S5, Table 5).
We observed (1) Table 5 for the other abbreviations and details diverse temporal trajectories of cortical thickness in both present and in previous studies demonstrate the complexity of brain plasticity.
Further studies are required, but these different cortical thickness dynamics across brain regions may be attributable to different biological mechanisms and time scales of brain plasticity across the different brain regions. This may also reflect different patterns of brain response to subcomponents of the SMART occurring over the 8-week timeframe of the training.

| Changes in resting-state functional connectivity following training
TBI-related deficits in rsFC can be caused by DAI (Smith et al., 2003).
This is one of the most common mechanisms of a closed head injury and is well documented in the TBI literature (Sharp et al., 2014), yet training-related changes in rsFC of individuals with TBI are still poorly characterized. In this regard, our rsFC findings  extend the TBI literature with results that demonstrate the utility of neuroimaging not only for diagnosing TBI but also for identifying brain changes associated with cognitive training for individuals with TBI. Furthermore, results for rsFC portion of our study ( Figure 6) are consistent with two theories of rsFC. First, rsFC is based upon the underlying structural connectivity, and therefore reflects the functional anatomy of brain systems (Damoiseaux et al., 2006). On the other hand, rsFC is also sculpted by the repeated history of coordinated activation between brain regions during experience-driven activities (Lewis et al., 2009). Our rsFC findings (Figure 6) further support the latter aspect of rsFC by demonstrating training-related changes in rsFC of injured brain. Note that our findings do not contradict the former aspect of rsFC (i.e., the patterns of rsFC follow functional  (Achard, Salvador, Whitcher, Suckling, & Bullmore, 2006). The strength of brain activity in regions with high between-network connectivity is proportional to the number of cognitive functions engaged in a task (Bertolero, Yeo, & D'Esposito, 2015).
Furthermore, damage to these regions disrupts the brain's modular organization (Gratton, Nomura, Pérez, & D'Esposito, 2012) and yields widespread deficits in neuropsychological measures (Warren et al., 2014). Critically, long-range and between-network connections are also vulnerable to damage from TBI (Han et al., 2014. Although further graph-theoretic studies comprehensively assessing brain networks of our participants are needed (e.g., studies evaluating greater numbers of nodes), the present training-related changes which showed a preferential impact to long-range and between-network connectivity suggest that the SMART program might contribute to more efficient neural communications thus perhaps healthier brain systems.

| Changes in cortical thickness and resting-state connectivity
Changes in rsFC among the regions which also exhibited temporal These previous multimodal imaging studies (Takeuchi et al., 2014;Taubert et al., 2011) separately investigated changes in rsFC and gray matter structure at the whole-brain level. However, we directly examined whether training-related changes in cortical thickness co-occur with connectivity changes or these cortical thickness changes remain spatially isolated. This was achieved by constraining regions of assess- F I G U R E 1 0 Colormaps for statistically significant associations between the improved trail-making test scores and changes in seed-based resting-state functional connectivity within the SMART group. See Figure 9 and Table 6 for the seed regions. The maps were corrected for multiple comparisons across vertices and seeds at p vertex < .05 and p cluster < .005 (=.05/10) of energy conservation in the brain (Laughlin & Sejnowsk, 2003). As such, diverse and complex changes in cortical thickness following training may reflect adaptive and dynamic processes of reallocating neural resources to support more 'successful' information processing relevant to the training program, while maintaining energy efficiency of the brain. In this context, training-induced increases in rsFC (particularly, long-range and between-network connections) might reflect more 'successful' information processing compared to the pretraining stage. Results of brain and behavior relationships indirectly support this hypothesis. As such, improvement in neuropsychological performance after the SMART was associated with monotonic increases in rsFC in most of the seeds and regions while the patterns of associations with cortical thickness included a mixture of monotonic and nonmonotonic changes (Figures 9 and 10). Future studies will be needed to directly measure the energy efficiency of the brain.

| Potential underlying mechanisms
The underlying microscopic-level mechanisms of the training-related changes in cortical thickness and rsFC are unknown, which is common in most of the MRI studies. However, we can speculate potential underlying mechanisms based on the literature. As in the case of learning and memory (Chklovskii, Mel, & Svoboda, 2004), functional plasticity (i.e., changes in synaptic strength without changing anatomical connectivity between neurons) alone may be insufficient to explain the underlying biological mechanisms of our findings because structural plasticity (i.e., changes in anatomical connectivity between neurons) is likely to be involved in the observed, training-induced changes in cortical thickness. Neurogenesis may not be one of the underlying mechanisms, as neurogenesis outside the hippocampus in human appears unlikely (Zatorre, Fields, & Johansen-Berg, 2012).
Rather, the primary underlying mechanisms may be synaptic plasticity (i.e., structural changes in synapses), which has been reviewed elsewhere (Butz, Wörgötter, & van Ooyen, 2009;May, 2011;Zatorre et al., 2012). More precisely, (1) synaptogensis by growing dendritic spines, increases in the number of synapses per neuron, dendritic arborization, and axonal sprouting, (2) selective synaptic pruning, and (3) synaptic remodeling by dendritic rebranching and axonal rerouting may have led to the training-induced changes in both cortical thickness and rsFC as we report herein. At present, it remains unclear which precise mechanism of synaptic plasticity plays a role at specific time phase. Several or all of these complex mechanisms may contribute to rewiring neural circuits for more efficient information processing, as indicated by increased rsFC following training. Second, changes in glial cells such as oligodendrocytes, astrocytes, and microglia (reviewed in Markham &Greenough, 2005 andZatorre et al., 2012) may also reflect training-induced changes in cortical thickness and rsFC. In particular, changes in oligodendrocytes, and subsequent myelination and myelin remodeling in the white matter are plausible candidate processes, as axonal connections are pivotal in the establishment of rsFC (Zhou et al., 2014). Further neuroimaging studies assessing white matter microstructure such as diffusion tensor imaging and myelin water imaging (Alonso-Ortiz, Levesque, & Pike, 2015) may help to further clarify the validity of this hypothesis. Third, changes in the cerebrovasculature may be another mechanism. Experience-dependent changes in cerebrovasculature in rats have been reported (Black, Sirevaag, & Greenough, 1987;Sirevaag, Black, Shafron, & Greenough, 1988).
Furthermore, Zatorre et al. (2012) suggested that changes in cerebrovasculature are a potential candidate mechanisms for gray matter changes following training, and Tak, Wang, Polimeni, Yan, and Chen (2014) reported a close association between the cerebrovasculature and measures of rsFC. Importantly, changes in cerebral blood flow and rsFC of the adult human brain following cognitive training have been previously reported Takeuchi et al., 2013).

| Limitations and future directions
Our study sample was heterogeneous, which may not be scientifically ideal, particularly in the context of diagnosis and characterization of pathology. However, in practice, TBI is marked by heterogeneity. Training-related changes within our heterogeneous TBI sample demonstrate the robustness and applicability of the SMART program for individuals with chronic TBI. Thus, the heterogeneity of our TBI sample is advantageous in terms of relating to the clinical population of interest. Although our TBI sample is inherently heterogeneous, we carefully randomized the two training groups making them demographically similar and we maintained training protocols consistently across individuals within each of the training groups. Second, even though we matched demographics, injury characteristics (e.g., initial injury severity and postinjury time), and severity of subclinical psychiatric symptoms of the two groups by carefully randomizing the participants, there were group differences in scores on multiple neuropsychological test at the baseline. These baseline group differences yielded cross-sectional group comparisons at TP 2 and TP 3 less meaningful. Thus, we focused on reporting group contrasts in temporal changes in neuropsychological measures (conceptually same as group-by-time interactions in ANOVA). The heterogeneity of neuropsychological test performance in TBI (Tellier et al., 2009;Thaler et al., 2013) and other potential factors that we did not measure such as genetics (Diaz-Arrastia & Baxter, 2006) might have led such group differences at the baseline. Future research in identifying factors that are affecting neuropsychological test performance in chronic TBI may provide us with better strategy to randomize training group assignment. Third, we did not systematically assess the effects of training on rsFC at the whole-brain level. As such, spatial and temporal patterns of rsFC following training outside the selected twelve regions, and training-related changes in network topology at the whole-brain level remain unknown. Thus, in the future, we will utilize a comprehensive graph-theoretic approach to address this concern. Fourth, it remains unclear whether the SMART program in individuals with chronic TBI led to recovery or compensation of altered neural circuitry that resulted from TBIs. Although improvement in neuropsychological tests performance is observed, "true" recovery of neural circuitry of injured brain following training may be unlikely (Kolb & Muhammad, 2014).
Future comparisons with healthy individuals may address this question. Lastly, our study would provide better understanding of brain and behavior relationships following rehabilitation for chronic TBI if we assessed more appropriate behavioral measures for improvement associated with the SMART program or if we adopted more refined and specialized training program. The SMART is an integrative training that aims to improve multiple domains of cognitive functions such as abstract reasoning, goal management, and selective attention (Vas et al., 2011). The number-letter switching versus motor speed of the trail-making test measures cognitive processing speed, working memory, and the ability to switch tasks while maintaining a goal (Sánchez-Cubillo et al., 2009). Although improvements in cognitive functions following the SMART for TBI can reflect improvement in the trailmaking test scores, the trail-making test does not measure the same level of cognitive functions that the SMART improves. This may be the reason why spatial patterns of between-group contrasts for changes in cortical thickness ( Figure 2) and statistically significant associations between changes in cortical thickness and the trail-making test within the SMART group ( Figure 9) did not overlap. Future studies utilizing carefully designed fMRI tasks that can tease apart SMART-induced cognitive domains may allow us to better understand the brain-behavior relationships associated with the SMART for chronic TBI.
Our future directions include carrying out assessments of other morphometric measures and graph-theoretic measures in this TBI cohort to better understand the current findings, and further address the concerns discussed above. We also plan to investigate and identify pretraining conditions that are predictive of training-related changes in cortical thickness and rsFC of the individuals with TBI, and are linked to training outcomes (Arnemann et al., 2015;Ventura-Campos et al., 2013).

| CONCLUSION
In conclusion, we provided neural evidence of the effects of cognitive rehabilitation in chronic TBI. Specifically, we demonstrated that strategy-based reasoning training led to dynamic changes in cortical thickness and rsFC in individuals with chronic TBI relative to an information-based training comparison group, even 3 months after training was completed. Our findings suggest that brain plasticity continues through the chronic phases of TBI, and a combination of cortical thickness and rsFC may be sensitive biomarkers for evaluating the efficacy of cognitive rehabilitation in the chronic TBI populations.

ACKNOWLEDGMENTS
First of all, we would like to thank all study participants and their families that supported this study. Our thanks also go to Weikei Yu