Brain changes following four weeks of unimanual motor training: Evidence from behavior, neural stimulation, cortical thickness, and functional MRI

Abstract

Although different aspects of neuroplasticity can be quantified with behavioral probes, brain stimulation, and brain imaging assessments, no study to date has combined all these approaches into one comprehensive assessment of brain plasticity. Here, 24 healthy right‐handed participants practiced a sequence of finger‐thumb opposition movements for 10 min each day with their left hand. After 4 weeks, performance for the practiced sequence improved significantly (P < 0.05 FWE) relative to a matched control sequence, with both the left (mean increase: 53.0% practiced, 6.5% control) and right (21.0%; 15.8%) hands. Training also induced significant (cluster p‐FWE < 0.001) reductions in functional MRI activation for execution of the trained sequence, relative to the control sequence. These changes were observed as clusters in the premotor and supplementary motor cortices (right hemisphere, 301 voxel cluster; left hemisphere 700 voxel cluster), and sensorimotor cortices and superior parietal lobules (right hemisphere 864 voxel cluster; left hemisphere, 1947 voxel cluster). Transcranial magnetic stimulation over the right (“trained”) primary motor cortex yielded a 58.6% mean increase in a measure of motor evoked potential amplitude, as recorded at the left abductor pollicis brevis muscle. Cortical thickness analyses based on structural MRI suggested changes in the right precentral gyrus, right post central gyrus, right dorsolateral prefrontal cortex, and potentially the right supplementary motor area. Such findings are consistent with LTP‐like neuroplastic changes in areas that were already responsible for finger sequence execution, rather than improved recruitment of previously nonutilized tissue. Hum Brain Mapp 38:4773–4787, 2017. © 2017 Wiley Periodicals, Inc.

INTRODUCTION

The brain is capable of remarkable structural and functional change to allow it to optimize performance. Such changes, collectively referred to as plasticity, are key to understanding a variety of real‐life phenomena, such as the learning of new skills [Sanes and Donoghue, 2000], the storing of memories [Xu et al., 2009], and recovering neurological function after brain injury [Nudo et al., 1996].

By definition, plasticity results in functional and/or structural change [Pascual‐Leone et al., 2005]. At the microscopic level, plasticity in the central nervous system can manifest in several different ways. Sprouting of new connections, unmasking of hidden or inhibited synaptic connections, and withdrawal of inhibition are some examples of these changes [Barron et al., 2016; Hoy et al., 1985; Kong et al., 2016]. Another form of plasticity is referred to as long‐term potentiation (LTP), which reflects an increase in synaptic efficacy [Bliss and Collingridge, 1993; Matsuzaki et al., 2004]. Behaviorally, neuroplastic changes typically manifest as altered task performance, such as improved speed of movement. Measuring performance of tasks before and after training provides a means of assessing the functional impact of training. Although such an approach can allow for an investigation of the effects of various factors on plasticity (e.g., aging, attention, and hormones), a purely behavioral approach is unable to shed light on the location and type of brain changes that occur after training. To probe the biological changes subserving plasticity, brain stimulation and brain imaging techniques are required, each with their own relative advantages and disadvantages [Reid et al., 2016].

In the area of motor‐skill acquisition, almost all studies investigating neuroplasticity have, so far, relied on animal models, cross‐sectional data, or functional measures of brain change [Chang, 2014]. In rodents, motor training appears to upregulate synaptic plasticity in the primary of motor cortex [Rioult‐Pedotti et al., 2000], and may also be associated with altered neural recruitment in this area [Costa et al., 2004]. In humans, early work using structural magnetic resonance imaging (MRI) suggested that differences in morphology of the primary motor cortex may exist between expert musicians and the general population [Amunts et al., 1997]. Since then, functional MRI (fMRI) studies have reported altered activity in motor planning in professional sports people [Milton et al., 2007], and transcranial magnetic stimulation (TMS) studies have similarly reported that skilled racquet players have a larger cortical representation of the hand than the general population [Pearce et al., 2000]. Functional MRI work has also suggested that the supplementary motor area (SMA), dorsolateral prefrontal cortex (dlPFC), and caudate nuclei may play integrated roles in error correction and learning of motor sequences [Chevrier et al., 2007; Kübler et al., 2006].

Longitudinal studies are particularly useful for studying neuroplasticity as they eliminate the possibility that neural differences are the cause, rather than result, of participants deciding to learn a skill. Several reports exist of changes in brain function in response to motor training [e.g., Chang, 2014; Orban et al., 2010], including in the pre‐ and postcentral gyri, dlPFC, and basal ganglia [Floyer‐Lea and Matthews, 2005]. Many extensive reviews of such work are available, including Chang et al. [2014], Dayan and Cohen [2011], Doyon et al. [2009], and others. In brief, trained tasks in such studies have been somewhat heterogeneous, and both increases [Karni et al., 1995] and relative decreases in fMRI activation have been reported, often within the same study [Floyer‐Lea and Matthews, 2005; Ma et al., 2010; Xiong et al., 2009]. The direction of change appears to depend on the duration of training [Dayan and Cohen, 2011; Floyer‐Lea and Matthews, 2005; Xiong et al., 2009], potentially due to increases in baseline cerebral blood flow [Xiong et al., 2009], though what constitutes “short‐term” or “long‐term” training appears to be task dependent [Dayan and Cohen, 2011]. A small number of reports have also detailed minute structural changes in grey matter in response to visuo‐motor tasks, particularly in the sensorimotor cortex [Bezzola et al., 2011], visual cortex, and superior parietal lobule [Draganski et al., 2004; Scholz et al., 2009], all areas associated with visuo‐motor skills.

So far, research in this area has predominantly relied on single data collection modalities, such as TMS, or fMRI. Unfortunately, the differences in study design, contrasts, time‐period, and imaging modalities of these studies can make it difficult to achieve a deeper understanding of the critical experimental parameters that yield reliable brain changes, whether such changes are typically concurrent or independent, and so on. In addition, each modality has its own limitations regarding reliability or interpretability, which can hamper insight into biological processes when only a single modality is acquired. For example, due to its indirect method of measuring brain activity, it is difficult to unambiguously interpret changes in fMRI in terms of biology without independent concurrent information, particularly given that controlling for task equivalency across scans can be difficult to achieve [Reid et al., 2016]. Collecting multiple modalities can also alleviate concerns that results in one modality simply reflect statistical anomalies, rather than subtle brain changes. For these reasons, the longitudinal study of motor training presented here concurrently acquired data on four different measures of neuroplasticity—behavior, TMS, fMRI, and cortical thickness. Findings from a fifth measure, diffusion MRI, appear in a second, companion study [Reid et al., 2017]. Each of these modalities helps to build a global picture of the changes that occur with motor training, by reporting on a slightly different aspect of the neuroplastic response.

TMS can evoke a clearly discernible and quantifiable motor response when motor cortical neurons are stimulated at sufficient intensity [Hallett, 2000]. This TMS‐evoked response, referred to as a motor evoked potential (MEP), reflects the excitability of underlying cortical neurons. Measuring the amplitude of the MEP before and after a training paradigm provides objective measurement of changes attributed to cortical motor plasticity [Stefan et al., 2000]. This is different from other, nonmotor regions, where brain stimulation‐evoked responses are more difficult to quantify (e.g., prefrontal brain regions), yet which also undergo changes via the same, ubiquitous, mechanisms. Although TMS is ideally suited to quantify training‐related change arising in the motor cortex, there are several limitations to the technique. In particular, TMS can only penetrate superficial grey matter structures [Roth et al., 2007], and can only probe the activity of one area of cortex at a time. Therefore, TMS is unable to provide information on whole‐brain changes arising from training.

Whole‐brain assessment of plastic change arising with training can be investigated with both functional and structural MRI. Functional MRI of local blood oxygenation level‐dependent (BOLD) signals provides information about changes in brain activity during tasks with high sensitivity and excellent spatial resolution. However, fMRI cannot readily distinguish whether any measured changes following training reflect excitatory or inhibitory activity [Logothetis, 2008], something that TMS can do. Changes in BOLD signals are also more straightforward to interpret, in terms of biological change, when in the context of information from other modalities, such as TMS [Reid et al., 2016]. By contrast, if conducted carefully, structural MRI can measure changes in cortical thickness that are more directly interpretable in terms of biology but, without accompanying functional measures, are difficult to link to any changes in brain activity.

With these points in mind, the aim of this study was to use a multimodal approach to investigate plasticity in the human cortex. We measured behavior, TMS‐evoked MEPs, BOLD changes using fMRI, and cortical thickness, before and after 4 weeks of motor training. To make interpretation of the results more straightforward, we opted for training that did not have any visual component. Our TMS analyses were conducted on the trained hemisphere and tested for increased MEP amplitude which could reflect altered connectivity induced by an LTP‐like process underlying motor training. For our fMRI analysis, we tested for both task‐induced increases and decreases in activation within sensorimotor cortices, SMAs, and superior parietal lobes. These phenomena may reflect altered processing in the “trained” hemisphere, and potentially an altered balance of interhemispheric inhibition. Although fMRI changes in higher areas, such as the dlPFC, were possible, we did not test for these explicitly. Finally, we tested for increased cortical thickness in motor areas of the trained hemisphere, given reports of subtle changes in visuomotor areas associated with visuomotor learning [Bezzola et al., 2011; Draganski et al., 2004; Scholz et al., 2009]. In combination with our parallel analysis of white matter and network changes [Reid et al., 2017], this work provides a comprehensive assessment of the functional and structural brain changes associated with motor training in humans.

MATERIALS AND METHODS

Overview

Twenty‐four participants were recruited (14 female; aged 28.8 ± 1.5 years; range 18–40 years). Participants were all right handed (laterality quotient 0.92 ± 0.03; range 0.58–1.0) as assessed by the Edinburgh Handedness questionnaire. Participants trained daily on a finger‐thumb opposition task [Karni et al., 1995] for 4 weeks. Behavioral, MRI, and TMS measures were obtained before and immediately after the training period to quantify training‐related changes. Participants were instructed to refrain from the consumption of known neuroactive substances (including caffeine and alcohol) before and during the training and quantification sessions. No participants reported any adverse effects. The study was approved by the University of Queensland Human Research Ethics Committee and all participants gave written informed consent.

Motor Training Task

The training task has been used in a previous independent study, in which it was reported to induce robust behavioral and functional changes [Karni et al., 1995]. The task involved participants performing a sequence of finger‐thumb opposition movements with their non‐dominant (left) hand (Fig. 1). Participants were pseudorandomly assigned one of two sequences that were mirror‐reverse copies of each other. Participants were instructed to perform their assigned sequence as quickly and as accurately as possible for 10 min each day for 4 weeks, and not to look at their hand during training. To help remember the correct sequence, participants were given a printed copy of their allotted sequence (red or blue; see Fig. 1). To minimize any circadian effects on motor training [Sale et al., 2008], each participant was randomized to practicing either during the morning or evening, and conducted training at the same time each day. Participants kept a log‐book to record their training, and were instructed to be honest when training sessions were missed, or were performed outside of the specified training time.

Figure 1.

Participants were randomized to practicing one of two finger‐to‐thumb opposition sequences. For the “blue” sequence (left), the order of fingers required to make contact with the thumb were little, index, ring, middle. For the “red” sequence (right), the order was little, middle, ring, index. These sequences were mirror images of one another. [Color figure can be viewed at http://wileyonlinelibrary.com]

Behavioral Measure

Participants' performance on the finger‐thumb opposition tasks was assessed as the number of correct sequences completed in 30 s. Performance was documented online with a hand‐held video camera, and quantified offline. Participants performed both red and blue (Fig. 1) sequences with both their left and their right hands. This was to investigate whether training induced any spill‐over of effects to a nontrained sequence and/or the contralateral hemisphere. The order of the sequences performed was randomized for each participant. To minimize errors, and to assist in performance, a print‐out of the sequence to be performed was placed in front of the participant for the duration of that task. Furthermore, prior to the quantification of baseline performance of the sequences (before training), participants were given a brief period of time (∼5–10 sequences) to practice the two sequences.

Performance on the finger‐thumb opposition sequences was analyzed using a three‐way repeated measures ANOVA with within‐participant factors of training (baseline, post), hand (left, right), and sequence (trained, control). Where appropriate, post hoc analyses were performed using Holm–Bonferroni corrected paired t tests.

Functional MRI

Functional MRI images were acquired in the same session as structural MRI images. We acquired 41 axial slices (slice thickness, 3.3 mm) using a gradient EPI sequence (TR 2.67 s; TE, 28 ms; flip angle, 90°; field of view, 210 × 210 mm; voxel size, 3.3 × 3.3 × 3.3 mm). A liquid crystal display projector back‐projected the stimuli onto a screen positioned at the head‐end of the scanner bore. Participants lay supine within the bore of the magnet and viewed the stimuli via a mirror that reflected the images displayed on the screen. Participant head movement was minimized by packing foam padding around the head.

Prior to entering the scanner, participants familiarized themselves with the two movement sequences—“trained” and “control”—they were going to perform within the scanner. The sequences are shown in Figure 1. Within the scanner, participants performed these finger‐thumb opposition movements with their left hand in blocks of 16 s, each followed by 16 s of rest. During rest blocks, a visual display showed a “Rest” command. At the start of each movement block a visual cue—displaying either the red or blue hand and corresponding movement sequence—notified participants whether they would be performing the trained or the control sequence. This was displayed for 2 s, and then removed. Participants then performed the required sequence at a rate of two movements (of the sequence) per second. As a cue to aid in this timing, a fixation cross flashed at 2 Hz on the screen. A tone also sounded at 2 Hz intervals throughout the acquisition. The last two tones in each movement block were at progressively lower pitches to notify participants that a rest block was imminent. Immediately following completion of each movement block (i.e., following the last fixation cross), a “Stop” command was presented for 1 s, which was then replaced with the “Rest” command, Four consecutive “runs” were performed. Each run consisted of four trained‐sequence blocks, four control‐sequence blocks, and seven rest blocks. The order of the trained/control sequence blocks was randomized but kept consistent between participants. Correct performance of the sequence was verified by recording the movements with a video camera, which were later reviewed for accuracy.

Image processing and statistical analyses were performed using SPM12 (Wellcome Department of Imaging Neuroscience, UCL, London, UK). Functional data volumes were slice‐time corrected and realigned to the first volume. The mean image of the resultant time series was co‐registered with the participant's temporally‐unbiased T1 template (see Methods: Cortical Thickness). The time series was then normalized into MNI space, an MNI brain mask was applied, and the result smoothed with an 8 mm FWHM isotropic Gaussian kernel. For first‐level statistical analyses, contrasts were conducted for (a) movement > rest (b) trained > rest, and (c) control > trained sequence blocks. All six motion parameters for each run were included as nuisance regressors. The second level analysis looked at the interaction between sequence and time point (i.e., baseline [control > trained] vs post [control > trained]). This was designed to test whether motor training altered the brain's response to performance of the trained sequence, taking into account any changes that were unrelated to training, or pre‐existing differences between responses to the control and trained sequences (although no baseline differences were expected). It was possible that the trained or control sequence would show greater BOLD responses after training, and so we treated this as a two‐tailed test. To account for this, we set a more conservative cluster significance level of P < 0.025 FWE with minimum cluster size of 200 voxels, after voxel‐wise thresholding at P < 0.0005 uncorrected.

TMS Mapping

Mapping of motor cortical excitability with TMS was performed in a separate laboratory, after MRI acquisition. Surface electromyographic recordings were obtained from the left abductor pollicis brevis (APB) muscle. Recordings were made using silver/silver‐chloride surface electrodes with the active electrode placed over the muscle belly, and the reference electrode placed over the adjacent metacarpophalangeal joint. Signals were amplified (1000×) and band‐pass filtered (20–1000 Hz) using a NeuroLog system (Digitimer), then digitized (2000 Hz) with a data acquisition interface (BNC‐2110; National Instruments) and custom MATLAB software (MathWorks). Signals were monitored online for movement‐related activity using high‐gain electromyography and a digital oscilloscope.

Mapping of motor cortical excitability before and after training involved applying TMS to the right hemisphere in a grid like pattern (described below). The TMS was delivered using a Magstim 200² stimulator (Magstim, UK) and a figure‐of‐eight coil (70 mm diameter). The coil was positioned with the handle pointing backwards at a 45° angle to the sagittal plane to preferentially induce current in a posterior‐to‐anterior direction in the cortex. The optimal scalp position for evoking electromyographic responses in the APB was established as the position that consistently evoked the largest MEP amplitude in this muscle with a slightly suprathreshold intensity. Resting motor threshold (RMT) was determined and defined as the minimum stimulus intensity which evoked an MEP of at least 50 µV in at least 5 out of 10 successive stimuli. The TMS intensity used for the mapping procedure was set at 120% of RMT. This stimulus intensity was established for both pre‐ and post‐training sessions.

Prior to participants' first TMS mapping session, their individual T1 MRI scan was processed with the neuro‐navigation software ASA‐Lab (ANT, The Netherlands). Markers were placed on the scalp surface of the structural scan in a grid‐like pattern spanning the entire right hemisphere. The grid commenced at the vertex, and markers were placed every 1 cm anteriorly and posteriorly from the vertex, and then extended out laterally (to the right) in 1 cm increments. Marker sites were targeted with TMS using a Polaris‐based infrared frameless stereotaxic system and Visor software (ANT, The Netherlands). Five TMS pulses were applied to these markers every 5 s. The markers were stimulated systematically, moving in a medial‐to‐lateral direction. Stimulation of the marker sites continued until the average MEP amplitude from a marker site fell below 50 µV. Those sites where average MEPs met or exceeded this amplitude were referred to as active. Once a row of markers had been assessed, stimulation was moved either 1 cm anteriorly or posteriorly (chosen randomly), until all active marker sites had been stimulated and identified. The MEP volume [Schabrun et al., 2009] was also calculated before and after training. This was calculated by summing the mean MEP amplitude of all the active sites.

Training‐related increases in the number of active sites and MEP volume were investigated using separate one‐tailed, paired t tests to compare pretraining measurements with those acquired after training. Significance was set as P < 0.05 after Holm–Bonferroni multiple comparisons correction.

Cortical Thickness

We examined cortical thickness changes arising in response to motor training. We acquired MPRAGE T1 images (0.9 mm isotropic) immediately before (“baseline”) and after (“post”) the 4 week training period using a 3 T MR system (Magnetom Trio, Siemens) and a standard 32‐channel head coil. Images were processed with Advanced Normalization Tools (ANTs; v2.1.0, source pulled 9th Feb 2016). An overview of the processing pipeline is shown in Figure 2.

Figure 2.

Cortical thickness pipeline. Steps are indicated with numbers. (1) Structural volumes from the baseline (left column) and post (right column) were preprocessed, skull‐stripped, and registered using a symmetrical registration. The resulting half transforms were applied to non‐skull‐stripped images to achieve an intermediate image (top, middle column), which was processed into a sharp single subject template. (2) The single‐subject template was then skull‐stripped and segmented. The result was carefully visually inspected. (3) If the segmentation was considered inaccurate, the brain mask was manually edited and Step 2 was rerun. (4) If the segmentation was successful, each time point was then re‐skull‐stripped and segmented using the single subject template (middle row of left and right columns), and cortical thickness calculated (bottom row of left and right columns). (5) Cortical thickness at each time point was moved into single‐subject template space, subtracted, and this difference transformed into MNI space for statistical analysis using the known transform between the single‐subject template and MNI space. Statistical analyses were then performed across all subjects, resulting in statistical maps (bottom row, middle column). Images here are purely illustrative; orientation differences have been exaggerated to convey concepts clearly. [Color figure can be viewed at http://wileyonlinelibrary.com]

Single participant analyses

In longitudinal analyses, it is important to utilize templates that are invariant to either time point, to boost statistical power and avoid false‐positives induced by registration bias [Thomas et al., 2009]. Although ANTs provides a longitudinal cortical thickness script, we found that the templates produced by this script on our system were commonly biased toward one time point. Therefore, we constructed templates for each participants by rigid‐registering the baseline and poststructural images together using ANTS SyN (symmetric) transform [Avants et al., 2008] after skull‐stripping, N4 bias‐field correction and intensity normalization. Time points were transformed by half of the resulting matrices then averaged, producing an approximate template that was unbiased with respect to time point. This initial template was converted into a sharper, segmented, template using the antsMultivariateTemplateConstruction script, followed by the antsCorticalThickness script using the ANTs NKI template. Each axial, coronal, and sagittal slice was then carefully visually inspected. In instances where dura or skull were classified as grey matter, or the brain undersegmented, edits were made to the brain mask using ITK Snap v3.4.0 [Yushkevich et al., 2006], and the script was rerun with the updated brain mask. For this study, particular emphasis was placed on achieving highly precise segmentations of the parietal and frontal lobes (Supporting Information, Fig. 1) due to their known association with, or proximity to, areas involved in motor training and/or learning. The process was repeated until we were satisfied with the final segmentations. Examples of acceptable and unacceptable segmentations are provided in Supporting Information, Figure 2. Posterior tissue probabilities were then converted into priors in a manner consistent with ANTs antsCookTemplatePriorsCommand script.

The antsCorticalThickness script was applied to structural images from both time points, utilizing the single subject template, producing a cortical thickness map in single‐subject‐template space. Subtraction of the pretraining from post‐training cortical thickness images produced a cortical thickness difference image.

Statistical analysis

We hypothesized that any region of change would likely be substantially smaller than atlas ROIs, and so opted for voxel‐based morphometry. Single‐subject template T1s were registered to FSL's 1 mm isotropic MNI152 atlas (Fig. 2) using ANTs SyN, and the resultant transforms applied to cortical thickness difference images. Voxels where the white matter was the most probable tissue, as defined by the mean of all single‐subject tissue priors, were excluded. Images were then smoothed with a 5 mm FWHM kernel and placed into a factorial model in SPM 12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/). This model regressed change in cortical thickness against time point. This model included sex as a factor to account for its previously reported effects on cortical thickness [Luders et al., 2006], and included ANCOVA normalization for nuisance effects to account for any remaining global differences. In the interests of accuracy and statistical power, we restricted our analysis to the parietal and frontal lobes—the areas which received particular focus during segmentation correction (Supporting Information, Fig. 1). We set our significance criteria as P < 0.05 FWE corrected, or a cluster comprising more than 20 voxels expressing values P < 0.001 uncorrected.

TMS and MRI Overlay

To interpret TMS, fMRI, and cortical thickness results together, it was important to show reasonable anatomical correspondence between the methods. To achieve this, TMS MEP responses were projected into MNI 152 space. This was achieved by converting each active site from Talairach to MNI coordinates, using the Lancaster transform [Lancaster et al., 2007], then connecting neighboring measurement nodes to form a mesh. An edge of zero‐value nodes was added to the outside of this mesh, based on the position of neighboring nodes. A duplicate of this mesh was projected 20 mm toward the midline of the base of the brain (x = 0, y = 0), reflecting the approximate penetration of the TMS pulse. Values on the grid were normalized between 0 and 1 on a per‐participant basis. For each participant, a volume was generated by linearly interpolating all voxels in MNI space between the inner and outer surfaces of the grid by the nearest surrounding 8 nodes. All participants' volumes were averaged to produce a mean image. This overlay was used for visual comparison of the modalities only; it was not used for quantification of TMS MEPs.

RESULTS

Behavioral Performance

There was no difference in behavioral performance for the morning and evening training groups, and therefore the data for the two groups were pooled for subsequent analyses. Prior to training, participants were equally proficient at performing the trained and control sequences with their left and right hands. Following training, there was a significant improvement in the number of sequences participants could complete in the 30 s period (Fig. 3; effect of time P < 0.001, partial η ² = 0.16; side × sequence × time interaction P < 0.01, partial η ² = 0.026). Using the left (trained) hand, the number of correct sequences completed on the trained sequence increased by 53%, from 24.0 ± 1.2 (mean ± SEM) to 36.7 ± 2.1 (Holm–Bonferroni adjusted P = 2.73 × 10⁻⁹). There were also smaller, but significant improvements in performance of the control sequence with the left hand (24.48 ± 0.94 to 25.65 ± 0.79 [mean ± SEM]; 6.5% increase; P = 0.025 [Holm–Bonferroni adjusted]), the trained sequence with the right hand (22.22 ± 1.11 to 26.48 ± 1.22; 21.0% increase; P = 7.53 × 10⁻⁶), and the control sequence with the right hand (21.82 ± 1.4 to 24.74 ± 1.15; 15.8% increase; P = 9.85 × 10⁻⁶). Although participants improved to a greater degree on the trained sequence than the control sequence (both hands P < 0.05 FWE), this difference between sequences was substantially greater for the left than the right hand (P < 0.05 FWE)

Figure 3.

Increase in performance of motor training tasks following 4 weeks of training. Group data (n = 24) showing number of correct sequences performed prior to (blue bars) and following (red bars) 4 weeks of training of a finger‐thumb opposition movement sequence (trained sequence). Participants performed the trained and control sequence with their left hand (trained) and right hand (not trained). Training improved execution speed for all four hand‐sequence combinations assessed (each P < 0.05 FWE). Data represent mean ± SEM. [Color figure can be viewed at http://wileyonlinelibrary.com]

fMRI Analysis

Of the 23 participants included in the fMRI analysis, none exhibited excessive (>2 mm or 2°) head movement during any session. Prior to training (Fig. 4), the trained sequence versus rest contrast revealed four significant (P < 0.05 FWE) clusters that were located across a variety of motor areas, predominantly in a bilateral manner, including the precentral gyri, postcentral gyri, SMAs, and superior parietal lobule. At this time, the activation of the trained and control sequences were, as expected, equivalent (i.e., there were no significant clusters when contrasted). Following training, however, the control sequence evoked greater activation than the trained sequence in a number of sensorimotor areas (Fig. 4), despite slight reductions in control‐sequence activation in these areas after training (Supporting Information, Fig. 3). To formally compare these differences, we calculated the interaction between training type and time point (baseline [control > trained] versus post [control > trained]). For this contrast, positive t values indicated locations in which the trained sequence showed reduced signal after training, relative to the control sequence. In the left hemisphere, a significant cluster overlapped the postcentral gyrus and superior parietal lobule (1947 voxels, cluster p‐FWE < 0.001; Table 1). In the right hemisphere, a similar cluster overlapped the precentral gyrus, postcentral gyrus, and superior parietal lobule (864 voxels, cluster p‐FWE < 0.001; Fig. 5). Both hemispheres also each demonstrated a significant cluster crossing the premotor cortex and SMA (left hemisphere, 700 voxels; right hemisphere, 301 voxels; each cluster p‐FWE < 0.001). Reversal of this contrast, highlighting areas of increased relative activation of the trained sequence, revealed no significant clusters.

Figure 4.

Mean of participants' normalized TMS responses (Top; normalized mV), positive BOLD activation during the trained task (middle; t values for task vs rest), and positive t value map contrasting BOLD for control sequence > trained sequence (bottom), overlaid on an MNI 152 T1 image. The left and right columns show measures before and after the training period, respectively. Crosshairs show equivalent anatomical positions (MNI 152 coordinate: 37, −16, 57) to aid viewing. TMS and BOLD signals show a large degree of overlap in the primary motor cortex. The TMS responses were centered more anteriorly (motor + premotor), particularly after training, than the BOLD responses (predominantly motor+ sensorimotor). Details of TMS projection into MNI space is contained in the text. [Color figure can be viewed at http://wileyonlinelibrary.com]

Table 1.

Locations of significant cluster fMRI peaks for the time point–sequence interaction (baseline [control > trained] > post [control > trained])

Cluster size	Peak coordinates	Anatomical locations
1947	−26, −54, 54	L SPL
	−44, −34, 36	L PoG
	−32, −44, 52	L SPL
864	18, −56, 60	R SPL
	40, −30, 42	R PrG/PoG
	20, −48, 62	R SPL/PoG
700	−12, −6, 64	L SMA
	−24, 6, 40	L premotor/SMA
	−24, 10, 42	L premotor
301	24, −8, 54	R premotor/SMA
	24, −2, 64	R premotor/SMA
	28, −10, 62	R premotor

Cluster size is in voxels. Peaks are in MNI 152 standard space.

Abbreviations: L, left; PrG, precentral gyrus; R, right; SMA, supplementary motor area; SPL, superior parietal lobule.

Figure 5.

Training‐related differences in BOLD activation. Groupwise functional MRI t‐map, showing voxels which, after training, demonstrated higher signal during the control sequence than the trained sequence. A series of axial slices is shown. Significant clusters were apparent in the superior parietal lobule and primary sensorimotor cortices in the right (864 voxels; P < 0.001 FWE) and left (1947 voxels; P < 0.001 FWE) hemispheres. Significant clusters were also seen in the supplementary motor area/superior premotor area of the right (301 voxels; P < 0.001 FWE) and left (700 voxels; P < 0.001 FWE) hemispheres. Only voxels belonging to significant clusters are shown; a structural template is displayed for anatomical reference only. Left of image is the left of brain. Axial slices shown are evenly spread between MNI‐152 z‐slices 67 and 26 inclusive. [Color figure can be viewed at http://wileyonlinelibrary.com]

TMS Mapping

There was no significant change in RMT following training (40.8% vs 41.6% maximum stimulator output; P > 0.05). By contrast, training induced an increase in the volume of MEP amplitudes evoked at the targeted sites by an average of 58.6% (P = 0.015; Fig. 6). That is, with the same relative TMS intensity (120% of RMT), MEPs evoked by TMS at the active sites were significantly larger following training. There was no significant increase in the number of active sites evoked with TMS (P = 0.09). A cortical heat‐map from a representative participant is shown in Figure 6, together with group‐wise changes. A normalized group‐wise heat map (see Methods) is also shown in Figure 4.

Figure 6.

Training‐related changes in motor cortical excitability. Top: A heat‐map representation of the cortical representation of the abductor pollicis brevis before (left) and after (right) training in one representative participant. Coordinates are referenced to the vertex (0, 0). The average MEP amplitude evoked at each site is indicated by the color scale (mV). Bottom: Mean number of active sites (left) and MEP map volume (right) before (blue) and after (red) training across all participants. There was no increase in active sites following training, but MEP volume increased significantly. [Color figure can be viewed at http://wileyonlinelibrary.com]

Cortical Thickness

One participant's dataset was lost during transfer to the server, and so was excluded from both cortical thickness and fMRI analyses. A second participant displayed slight MRI artefacts on the T1 images in the right sensorimotor cortex, and so was excluded from cortical thickness analyses. For a third dataset, we were unable to achieve high‐quality tissue segmentation and so we excluded this dataset from cortical thickness analyses, leaving 21 participants in total for these analyses. Unthresholded t‐value images are provided in Supporting Information, Figures 4 and 5. At an uncorrected threshold of P < 0.001, training resulted in an increase in cortical thickness in the right precentral gyrus (81 voxel cluster), right post central gyrus (34 voxel cluster) and right dlPFC at approximately Brodmann's area 9 (22 voxels; Fig. 7). The right SMA also contained two nearby clusters, 11 and 13 voxels in size, whose individual volumes did not cross our prespecified threshold, but whose sum did. Although these five clusters did not survive multiple comparisons correction (P < 0.05 FWE) all changes were observed in regions that are likely to be involved with motor tasks, and all were in the right (“trained”) hemisphere. Notably, clusters at the pre‐ and postcentral gyri were consistent with the expected location of sensory and motor finger representations [Penfield and Boldrey, 1937; Wahnoun et al., 2015], and close to peaks in the (post time point) TMS and fMRI maps. All clusters were within the group‐wise region of successful TMS excitation, and all but the prefrontal cluster were consistent with regions of group‐wise fMRI activation. To test the robustness of these findings, we removed data from the two participants who displayed the strongest performance improvements. In this reanalysis, the previously found clusters in the precentral gyrus, postcentral gyrus, and prefrontal cortex were still apparent (data not shown).

Figure 7.

Areas of increase in cortical thickness overlaid on the FSL MNI 152 template. Light blue indicates voxels with significant (cluster size > 20 voxels at P < 0.001 unc) increases in cortical thickness estimation. Increases were seen in the right prefrontal cortex (top row), and approximate hand areas of the right precentral (middle row) and postcentral (bottom row) gyri. Red–yellow indicates mean normalized TMS response, where yellow indicates a strong response (mean ≥ 30% normalized peak motor evoked potential) and red indicates a weaker response (mean ≥ 1% of normalized peak motor evoked potential); see text for details. Left of coronal and axial images indicates left of brain. Axial slices (top‐to‐bottom) show MNI y coordinates 30, −3, and −18. Sagittal slices show MNI x coordinates 44, 58, and 61. [Color figure can be viewed at http://wileyonlinelibrary.com]

DISCUSSION

The neural changes that accompany a period of motor training contribute to the increase in performance. Understanding how these neural changes manifest themselves is important in both guiding rehabilitation strategies [Reid et al., 2015], and understanding normal brain function. Such neural changes can be quantified in several ways, each with their respective advantages and limitations. Here, we show how utilizing several different methods to concurrently quantify plasticity—behavioral, brain stimulation, functional brain imaging, and structural brain imaging—can offer broad insight into the plastic changes that arise following training. To our knowledge, these multimodal data from healthy adults provide the most comprehensive assessment of the functional and structural changes that occur following training to date. Multimodal assessments are particularly important in this context as the measurement of minute brain changes is difficult. Positive results from single modalities, then, are often interpreted with caution (e.g., perception of potential publication bias). By contrast, coherent results from multiple, concurrently acquired and independently analyzed modalities—as reported in this study, and its companion paper [Reid et al., 2017]—lowers the probability of an overall false positive.

Four weeks of training of a sequence of finger‐thumb opposition movements resulted in a substantial improvement in performance. This was particularly true of the trained sequence performed with the trained hand, for which correct sequence completions improved by 53% (Fig. 3). This was similar (albeit slightly lower) than reported in two similar, yet smaller, studies [Karni et al., 1995; Xiong et al., 2009]. To further investigate the brain changes arising from training, we incorporated several other probes of cortical structure and function. Although transfer of skill did occur, both between sequences and between hands, our fMRI and TMS analyses were not explicitly designed to investigate such effects. Diffusion MRI analyses detailed in a companion paper [Reid et al., 2017] explored transfer effects in the same participant group in more detail.

Functional MRI

Our fMRI analyses revealed that, only after training, the trained sequence evoked lower cortical activation bilaterally than the control sequence (Fig. 5), predominantly in the sensorimotor cortices and superior parietal lobe (Table 1). The apparent relative reduction in functional activation reported here is at odds with an earlier smaller (n = 6) study that used a very similar training and scanning approach [Karni et al., 1995]. A subsequent replication of this work showed that fMRI activation decreased during the third and fourth weeks of training, and PET data suggested that this was, at least in part, due to an increase in baseline blood flow rather than a decrease in actual brain activity [Xiong et al., 2009]. This study provides an interesting insight into these works: our contrast focused on the interaction between time point and task, which means that, although changes in resting blood flow may have occurred, such changes cannot explain our results. Specifically, as both the control and trained tasks evoked the same patterns of activation at baseline (Fig. 4), changes in resting cerebral blood flow would be expected to affect both equally, and so cannot explain the statistical interaction reported here. An alternative, nonmutually exclusive, explanation is that local changes, such as LTP, allowed local grey matter that was responsible for performing the trained task prior to training to perform the trained sequence more efficiently, reducing the need for recruitment of surrounding areas. Further, because we controlled for execution speed during scanning, the parietal changes we report (i.e., increased activation in the control vs trained sequence) might reflect a reduction in attentional resources required to perform the trained sequence compared to the control sequence. This finding is consistent with the idea of more efficient processing. Future research using fMRI could be useful in investigating such attentional network‐related changes following motor training.

Previous research has highlighted that motor training is supported by changes in the allegiance of specialized regions to a given brain system, or module [Bassett et al., 2011]. For example, the analysis of fMRI data acquired during the performance of a motor sequence task over six weeks revealed that increased automaticity in task performance is supported by a reduction in the functional integration between motor and visual areas composing two distinct brain modules [Bassett et al., 2011]. Moreover, the gained automaticity was related to the disengagement of cognitive control hub regions comprising fronto‐parietal and cingulo‐opercular systems [Bassett et al., 2011]. These findings echo recent results showing that a focal reduction of baseline activity in the primary motor cortex causes a significant reorganization of functional connectivity patterns within the sensorimotor module and between this module and the rest of the brain [Cocchi et al., 2015]. Specifically, a reduction in motor cortex excitability was paralleled by enhanced functional connectivity within the sensorimotor module and reduced connectivity with other modules of the brain. In this study, we did not directly assess changes in connectivity as a function of training but, previous findings suggest that the observed changes in local neural activity occur in the context of broader changes in brain network dynamics. Future work will need to clarify the functional interplay between training‐induced local changes in neural activity and global modulation of brain network dynamics.

Although motor training paradigms, similar to the one employed in this study, have been shown to induce LTP‐like changes in cortical activity [Ziemann et al., 2004], it would be fair to consider this hypothesis speculative if based on fMRI alone. A novel aspect of the present work, however, is providing evidence in support of this hypothesis from two additional lines of enquiry in the same participants: changes in TMS‐evoked responses and changes in cortical thickness.

Transcranial Magnetic Stimulation

TMS provides an indirect way to assess LTP‐like changes in cortical excitability. Since TMS activates motor cortical output neurons trans‐synaptically, if synaptic efficacy is increased, this should lead to an increase in the amplitude of the MEP at a given stimulus intensity [Di Lazzaro and Ziemann, 2013]. Here, we showed that MEP amplitudes were larger following training (Fig. 6). Notably, there was no significant change in the area that could evoke an MEP in the APB. This suggests that the changes indexed by TMS were predominantly driven by changes in neural networks that already played a role in contraction of the APB.

To consider TMS and MRI evidence together, it is important that these are viewed with respect to one another, to ascertain that regions of measurement (or signal change) for these modalities have reasonable anatomical overlap. In this study, the areas of cortex that were activated by the trained sequence (in the trained hemisphere) during fMRI scans were very similar, though not identical, to areas eliciting MEPs during TMS mapping (Fig. 4). TMS and fMRI map markedly different aspects of motor control, and so perfect overlap between them should not be expected: fMRI here contrasted BOLD responses to sequential movement of the fingers and thumb versus rest, while TMS targeted neurons functionally relevant to activating the APB muscle.

Keeping in mind that there was no increase in the RMT, the TMS findings suggest that at least one of three processes have taken place: enhanced trans‐synaptic transmission (e.g., through LTP), increased neurite density, and/or improved conduction of the corticospinal tract. Both the first and second of these possibilities support our earlier hypothesis introduced in the context of the fMRI results. The third possibility, regarding changes in white matter, is investigated in detail in our follow‐up publication [Reid et al., 2017].

Cortical Thickness

Increased cortical thickness was observed after motor training within the right SMA, right middle‐frontal gyrus, and right precentral gyrus and right postcentral gyrus. Although cortical thickness changes did not reach our stricter FWE‐corrected threshold, the clusters of significant voxels seen were sizeable (22–81 voxels @ P < 0.001) and striking in their location. As previously mentioned, one motivation for the present study was to index motor training related brain changes across a variety of modalities to better understand to what degree, if any, these are concurrent, and to build a more global picture of the processes that allow behavioral improvement. Toward this aim, it is then interesting that these changes in cortical thickness were substantially less strong than TMS, fMRI, behavior, and the diffusion measures reported in the accompanying paper [Reid et al., 2017]. Although these findings are not definitive in themselves, they are interesting in the context of our other findings. Specifically, the location of clusters in the pre‐ and post‐central gyri were relatively consistent with our fMRI maps, our TMS maps (Fig. 7), and with areas responsible for sensorimotor representation of the fingers elucidated through electrocorticographic and brain stimulation techniques [Penfield and Boldrey, 1937; Wahnoun et al., 2015]. The neighboring clusters seen in the right SMA were also within the TMS and fMRI maps from our study, and consistent with previous findings that this area plays a role in motor training [Taubert et al., 2010]. That cortical thickness changes were very subtle, while task performance was strongly improved, and other brain‐change measures relatively clear cut, may be indicative that the processes driving relatively short‐term skill acquisition are not those that drive strong increases in cortical thickness. That these changes appeared near the peaks of fMRI and TMS maps adds further credibility to the suggestion that changes in fMRI and TMS patterns were a reflection of altered neurite organization, density, and/or connection strength in regions that, at baseline, were already predominantly responsible for execution of the (to‐be) trained sequence, rather than altered responsibility of surrounding areas.

The final location in which we saw changes in cortical thickness was the dlPFC, which is known to play an important role in error‐correction of motor output (along with the caudate nucleus) [Chevrier et al., 2007; Kübler et al., 2006]. Structural changes in the dlPFC were toward the edge of the TMS response map, and not in an area of fMRI change. The role of the dlPFC is to modulate motor responses, not generate them directly, and so it should not be surprising that only small TMS responses were recorded. Furthermore, we do not consider cortical thickness change here to be at odds with fMRI findings because the fMRI analysis was optimized for detection of motor output and conducted at around half the speed participants were capable of at baseline; it did not contrast very‐challenging‐versus‐relaxed motor performance, as would be optimal to highlight an area that plays a role in error detection. In fact, we believe this apparent discrepancy highlights the usefulness of multimodality imaging in measuring neuroplasticity to provide a more accurate overview of changes that take place during training [Reid et al., 2016]. The dlPFC and related areas were further investigated with diffusion imaging; these analyses are described in a separate publication [Reid et al., 2017].

Biological Interpretation

MRI is not an optimal tool for the interpretation of biological changes in terms of cellular physiology. Nevertheless, it provides an excellent means of localizing brain changes and, particularly when applied in a multimodal manner, can suggest which family of processes may underlie changes. The behavioral, fMRI, TMS, and cortical thickness changes presented here are consistent with an LTP‐like change in the grey matter. Functional MRI showed a task‐specific reduction in BOLD signal, consistent with more efficient processing in the grey matter. TMS findings suggested that enhanced trans‐synaptic transmission and/or improved corticospinal tract conduction took place. Cortical thickness findings suggested subtle structural changes in the same regions. The fact that cortical thickness findings were not strong suggests that gross morphological changes to the grey matter were not a primary driving force toward behavioral change. This is more consistent with more subtle (dendritic or synaptic) changes rather than larger scale changes, such as glial cell proliferation or angiogenesis. The precise combination of processes is not currently something that can be determined with MRI alone. However, the locus of these changes—regions already involved in performing the nontrained task at baseline—indicates that improved function is at least partially driven by more efficient grey matter processing in regions that, at baseline, were already predominantly involved with the execution of the trained sequence, rather than altered activation of surrounding areas.

Generalization to the Untrained Hand

We report improvements in task performance for both sequences in both hands, despite only one sequence with the nondominant hand being trained. Such cross‐hemisphere generalization is at odds with one smaller study (n = 6) that employed a similar training task [Karni et al., 1995], but is consistent with findings from several other motor studies that utilized other tasks [Hicks et al., 1983; Teixeira, 2000]. Our fMRI analyses revealed that, after training, execution of the trained sequence with the left‐hand elicited reduced activation relative to the control sequence, bilaterally. However, the present results and those of our companion paper [Reid et al., 2017] revealed no changes in the left hemisphere as measured with structural or diffusion MRI. This discrepancy suggests that altered fMRI and behavioral measures of the “untrained” side may reflect changes in grey matter, such as LTP or altered neurite density, that were too subtle to be detected by structural MRI. Techniques such as neurite orientation dispersion and density imaging [Zhang et al., 2012], or bilateral TMS, EEG, or fMRI tasks targeting the opposite hand, might allow future studies to shed further light on this issue.

Limitations

This study was designed to maximize the amount of concurrently acquired information. Practical considerations, however, meant that the depth of investigations with each modality had to be limited, leaving some open questions for future studies. For example, more challenging rates of task execution were not assessed using fMRI. Including such a manipulation might have helped us interpret the changes in dlPFC and related structures, as reported both here and in our diffusion MRI work [Reid et al., 2017]. Similarly, TMS was not acquired for the “untrained” hemisphere. Including such a condition might help to understand processes responsible for skill transfer between hemispheres. Future work could also further illuminate these processes by including an additional time point, several days or weeks following the cessation of training. Another possibility would be to include daily behavioral assessments, to assess whether training‐related brain changes reflect continuous skill acquisition, or instead track ongoing learning processes themselves and subside once skill level reaches a plateau. In a multiple‐time point study of this nature, collecting behavioral error rates, in addition to successful‐sequence completion rates, would also be advantageous.

CONCLUSION

We have shown that 4 weeks of motor training can invoke robust changes in behavior, cortical thickness, fMRI activation, and TMS‐evoked motor maps in motor regions of the brain. Taken together with our diffusion MRI findings [Reid et al., 2017], the work presented here provides the most cohesive and comprehensive longitudinal study of motor plasticity in healthy adult humans to date. All three brain measures suggested that motor training was driven by LTP‐like plasticity that was relatively widespread across the sensorimotor system—even when a participant is trained on a simple task solely on the basis of proprioceptive feedback.

CONFLICT OF INTEREST

The authors declare that they have no conflicts of interest.

Supporting information

Supporting Information