Abstract
How does our brain transform when we encounter a new task? To fully answer this question, comparing brain states before and after learning may not be enough, but rather an ongoing, continuous monitoring of brain changes during learning is required. While such continuous examinations of functional learning-induced changes are widely available using functional magnetic resonance imaging (MRI), a continuous investigation of diffusion-detected brain modifications during learning is yet to be reported. Here, we continuously acquire diffusion MRI images during task performance. We then compute the mean diffusivity (MD) using a sliding-window approach, resulting in a continuous measure of diffusivity changes throughout learning. We demonstrate the utility of this method on a motor sequence learning (finger tapping) task (n = 58, 30 females). MD decrease was detected in task-related brain regions, including the parahippocampal gyrus (PHG), hippocampus, inferior temporal gyrus, and cerebellum. Analysis of the temporal patterns of decrease revealed a rapid MD reduction in the right temporal gyrus after 11 min of training, with additional decrease in the right PHG and left cerebellum after 22 min. We further computed “neuroplasticity networks” of brain areas showing similar change patterns and detected similarities between these networks and canonical functional connectivity networks. Our findings offer novel insights on the spatiotemporal dynamics of diffusion-detected neuroplasticity by demonstrating continuous modifications during the encoding phase of learning itself rather than comparing pre- and postlearning states.
Significance Statement
Learning is a dynamic process with widespread functional and structural brain changes unfolding over time. To fully understand how our brain transforms during this process, comparing brain states before and after learning may not be enough. While continuous examinations of functional changes during learning are widely reported, an ongoing, continuous investigation of microstructural learning-induced modifications is lacking. Here, we employ a unique diffusion magnetic resonance imaging protocol and a sliding-window analysis approach to continuously characterize tissue diffusivity changes, while participants perform a motor sequence learning task. Our findings offer novel insights on the spatiotemporal dynamics of neuroplasticity by portraying the real-time unfolding of diffusivity changes in task-related areas across the brain.
Introduction
Learning takes time. While the development of new skills occurs over extended periods of practice and experience (Green and Bavelier, 2008; Knight et al., 2017), learning-related neural modifications are evident already at the very first minutes of encoding (Sagi et al., 2012; Brodt et al., 2018; Tavor et al., 2020). To fully understand the neural underpinnings of learning, it is crucial to consider its continuously evolving nature.
Numerous studies have explored the structural and functional alterations that accompany learning using structural and functional magnetic resonance imaging (fMRI), respectively. Changes in cortical thickness and gray matter volume have been reported (Koch et al., 2016; Legault et al., 2019), as well as changes in white matter structures (Hofstetter et al., 2013; Sampaio-Baptista et al., 2013; for a review, see Sampaio-Baptista and Johansen-Berg, 2017). Functional brain changes have also been detected, e.g., following motor skill acquisition (Ungerleider et al., 2002) or musical training (Luo et al., 2012), mainly in learning-related cortical areas but more recently also in the cerebral white matter (Frizzell et al., 2020, 2022; Ji et al., 2023).
While structural changes have been so far described primarily on a macroscale level, over the last decade diffusion-weighted MRI (dMRI) has been employed to study “microstructural” neuroplasticity in gray matter structures as well (Sagi et al., 2012; Tavor et al., 2013, 2020; Brodt et al., 2018). This recently emerging research differs from previous investigations in that it focuses on structural rather than functional changes and on a “micro-” rather than macroscale level. Additionally, while dMRI has been traditionally used to study the white matter, here it is employed to investigate gray matter plasticity.
dMRI measures the translational displacement of water molecules (Le Bihan and Warach, 1995). This allows to construct a mathematical representation, the diffusion tensor (Basser et al., 1994), from which it is possible to calculate quantitative information on tissue microstructure, such as the mean diffusivity (MD). Learning-induced decrease in MD has been reported (Taubert et al., 2010; Sagi et al., 2012; Sampaio-Baptista et al., 2013; Tavor et al., 2013, 2020; Hofstetter et al., 2017), possibly reflecting changes in the extracellular matrix (Van Der Toorn et al., 1996), formation of synapses and dendrites (Toni et al., 1999), dendritic remodeling, axonal sprouting and pruning (Kays et al., 2012), neurogenesis (Gage, 2002; Knoth et al., 2010; Seib and Martin-Villalba, 2015), astrocyte remodeling (Blumenfeld-Katzir et al., 2011; Sagi et al., 2012; Assaf, 2018), or other glial cell modifications (Fields et al., 2014; Weston et al., 2015).
Simultaneously, diffusivity changes have been associated with functional (e.g., visual) stimulation, suggested to reflect transient microstructural changes, such as cell swelling, closely tied to neuronal activity (Le Bihan, 2003; Le Bihan and Johansen-Berg, 2012; Tsurugizawa et al., 2013; De Luca et al., 2019; Nunes et al., 2019, 2021). Diffusivity indices such as MD may therefore point to either momentary performance-related or more stable learning-related changes or both, offering a promising marker for neuroplasticity.
Neuroplasticity studies using dMRI have targeted various timeframes of brain changes (Thomas and Baker, 2013), ranging from months to just a few hours of training across different cognitive domains, including spatial navigation (Sagi et al., 2012; Tavor et al., 2013; Keller and Just, 2016; Villemonteix et al., 2023), motor sequence training (Jacobacci et al., 2020; Tavor et al., 2020), language (Hofstetter et al., 2017), and associative memory (Brodt et al., 2018). Notably, these studies have investigated the differences in brain microstructure between two discrete timepoints: before and after learning. However, the temporal dynamics of the learning process and what transpires within the brain during that time have not been fully explored yet.
This study aims to monitor continuous diffusion-detected changes during the learning process itself. To achieve this, we computed a continuous MD measurement and used it to identify temporal patterns of neuroplasticity across the brain, to detect the exact changepoint in MD during learning, and to define “neuroplasticity networks” of brain regions that share common diffusion-detected modification patterns.
Materials and Methods
Participants
Sixty-two right–handed healthy volunteers with no history of neurological disease, psychiatric disorders, drug or alcohol abuse, or use of neuropsychiatric medication were recruited. Four participants were excluded due to technical issues during the scan or excessive movement, resulting in a final sample of 58 participants (30 females; mean age, 27.66 years; SD, 4.4). Participants were randomly assigned to either a learning group (29 participants; 14 females; mean age, 28.47 years; SD, 4.26) or a control group (29 participants; 16 females; mean age, 26.5 years; SD, 4.37; with no significant age or gender differences between groups). The experimental protocol was approved by the Institutional Review Board of the Sheba Medical Center and all participants signed an informed consent form.
Experimental design
Participants underwent an MRI session which included a dMRI scan and several structural scans. During the dMRI scan, participants in the learning group performed a motor sequence learning task (see details below), while participants in the control group were instructed to focus on a fixation cross and not perform any explicit task. MRI acquisition was carried out in the Alfredo Federico Strauss Center for Computational Neuroimaging at Tel Aviv University.
MRI acquisition
Scans were acquired on a 3 T whole-body MRI system (Siemens Magnetom Prisma) equipped with a 64-channel head coil.
Diffusion-weighted images (DWI) were acquired with a spin-echo diffusion–weighted, echoplanar imaging sequences with up to 68 axial slices (whole-brain coverage) and a resolution of 2 × 2 × 2 mm3, with a repetition time (TR)/TE of 3,500/59.4 ms and a multiband acceleration factor of 2. Diffusion parameters were Δ/δ = 28/10 ms; b values of 1,000 s/mm2 were taken at 589 gradient directions with additional 33 nondiffusion weighted (b0) images (spread evenly across the scan duration, such that in each MD calculation, the nearest B0 image was used as a reference to assess diffusion-related decay of the MR signal for each tensor calculation). The total dMRI scan duration was 36:28 min. In addition, five b0 images and one b1000 image were acquired with a reversed phase-encoding direction to correct for susceptibility-induced distortions as described below, dMRI data analysis.
T1-weighted images were acquired with a magnetization-prepared rapid gradient echo sequence with up to 176 axial slices (whole-brain coverage), TR/TE of 2,400/2.98 ms, resolution of 0.9 × 0.9 × 0.9 mm3, and scan time of 4:30 min. T2-weighted images were acquired with up to 176 slices (whole-brain coverage), TR/TE of 3,200/554, resolution of 0.9 × 0.9 × 0.9 mm3, and scan time of 5 min. In addition, fluid-attenuated inversion recovery images (TR/TE/TI, 8,000/81/2,370) were acquired for radiological screening.
Motor sequence learning task
Participants in the learning group performed a finger tapping task (first introduced by Karni et al., 1995). The task consisted of 24 trials of 60 s, each followed by a 25 s rest period, for a total duration of 33:35 min, with additional 2:53 min of rest at the end of the task. During task trials, participants were instructed to repeat a five-digit sequence (2-4-1-3-2) using their left (nondominant) hand as correctly and quickly as possible, while the sequence is constantly presented on the screen to eliminate executive memory differences. Rest periods were added to allow for off-line gain learning (Bönstrup et al., 2019; Jacobacci et al., 2020). Finally, participants performed an additional task trial while in the scanner, but with no scan being performed, serving as a behavioral test.
Behavioral assessment
Two (nonindependent) measurements of task performance were calculated: (1) accuracy, the total amount of correct sequences completed, calculated as the number of correct sequences a participant performed within a specific trial, and (2) speed, the mean duration of a successful key press sequence, calculated as the average duration of correct sequences in a specific trial (sequence duration being the time passed between pressing the first and last keys of a sequence). Both measurements were computed for each trial and for the test trial. Paired t tests were performed for each of these measurements, between the first and last trials, and the test trial.
dMRI data analysis
dMRI images were corrected for head motion, susceptibility-induced distortions, and eddy current-induced distortions, as well as registered to the T1-weighted image using the minimal preprocessing pipeline proposed by the Human Connectome Project (HCP; Glasser et al., 2013; Sotiropoulos et al., 2013) including the FMRIB Software Library (FSL) TOPUP and EDDY tools (Andersson et al., 2003; Andersson and Sotiropoulos, 2016). For optimized registration to the Montreal Neurological Institute (MNI) standard space, we applied a tissue-probability–based registration routine, as recently described by Malovani et al. (2021). Data were spatially smoothed with a Gaussian kernel with a full-width at half-maximum (FWHM) of 5 mm using FSL.
MD was calculated using the diffusion tensor model (Basser et al., 1994). To obtain a continuous measurement of diffusivity changes during learning, we calculated the diffusion tensor by applying a sliding-window approach: We used a window length of 13 TRs [12 DWI plus one nondiffusion weighted (b0) image] and a window offset of a single TR as described in Figure 1A. To determine the optimal window length, we used data from three different sources, using three different diffusion protocols: (1) 4 participants who were excluded from the present study due to technical issues or excessive movement (these participants were scanned with the exact same diffusion protocol as the remaining participants in this study; see above, MRI acquisition); (2) 10 participants who were scanned in the same MRI scanner but using a different dMRI protocol, containing 64 different gradient directions at b = 1,000 s/mm2; and (3) 5 HCP participants (Van Essen et al., 2012) who underwent a dMRI scan containing 90 different gradient directions at b = 1,000 s/mm2. We calculated the MD from an increasing number of directions, ranging from 1 to 61 (to allow for an adequate number of possible permutations for the shortest dataset (Number 2) or
Continuous MD using a sliding-window approach. A, An illustration of a sliding window over multiple DWI volumes. The window length was set to 13 TRs (12 DWI in different gradient directions + the nearest b0 volume) and the window offset was one TR. B, For each window, an MD value in each voxel was calculated to create a continuous. To determine the optimal window length, we used data from three different sources; see Extended Data Figure 1-1 for a detailed analysis. All subsets were constant for all participants and were spread evenly of a sphere (see Extended Data Fig. 2-1 for examples).
In addition to this continuous characterization of MD changes, we also calculated MD “pre” and “post” learning, as the mean MD in the first and last 10 MD timepoints in the continuous MD curve, respectively.
To further explore the evolving pattern of MD changes during learning before inspecting the full continuous change in time, we investigated discrete timepoints along the continuous MD curve shown in Figure 1B. Specifically, we computed the percentage change in MD between the “prelearning” timepoint (as described above) and three additional timepoints: ∼11 min into the learning process, ∼22 min into the learning process, and at the end of the scan, after training is completed, which is identical to the “pre”/”post” comparison described above. The timepoints at 11 and 22 min were chosen as two middle points during the learning process. Each of these additional points was calculated as the mean of 10 MD timepoints, starting at 11 or 22 min into learning. Since an MD timepoint lasted 42 s and was separated by 3.5 s from the next one, 10 MD timepoints lasted
Statistical analysis
To detect clusters of MD decrease, which were later used for investigation of continuous modifications during the training process itself, we first performed a voxelwise two-way ANOVA of MD timepoint (before/after learning) × group (learning/control) using FSL's PALM tool with 1,000 permutations (Winkler et al., 2014). Only gray matter voxels were included, and a threshold-free cluster enhancement (Smith and Nichols, 2009) was utilized. P values were FDR corrected for multiple comparisons. Clusters with over 50 voxels that showed an interaction effect (i.e., MD decrease in the learning but not the control group) were included in the following stages of analysis. To rule out the possibility that MD changes across time merely reflected fluctuations in b0 image intensity, we performed the same statistical analysis on the first and last nondiffusion weighted (b0) images. No pre–post differences were found.
For each of these clusters, we computed the “pre” and “post” learning MD as described above, as well as the differences between the “prelearning” MD and the additional timepoints throughout the learning process, averaged within the learning or control groups. Paired t tests were performed between the “prelearning” MD and the three subsequent timepoints (11 min, 22 min, and postlearning).
Next, to obtain a continuous characterization of MD change, we averaged MD values per MD timepoint within each cluster across voxels and across participants for each group independently, resulting in two (for the learning and control groups) distinct MD timeseries per cluster.
Changepoint detection analysis
To further compare between the patterns of MD change across clusters, we performed a changepoint analysis (Killick et al., 2012) per timeseries to detect a single specific timepoint in which the greatest change in the slope has occurred. The changepoint for each cluster was calculated per participant and averaged per group. This analysis was performed first on different cutoff frequencies for the low-pass filter (and on the unfiltered data) and second at different window lengths to examine whether the found changepoint represents the overall trend of the data and not a specific fluctuation. To achieve a more detailed understanding of the spatial pattern of temporal changes, changepoint analysis was additionally used to examine the change in the slope separately for each voxel within a cluster.
Networks analysis
We next examined the similarities between diffusion-detected plasticity patterns across brain areas. Gray matter was parceled into 273 areas [Brainnetome Atlas (BNA); Fan et al., 2016]. We calculated the individual continuous MD change averaged across voxels per parcel and computed Pearson's correlations between all region pairs. Individual correlation matrices were then averaged across participants. We applied a hierarchical clustering on the averaged correlation matrices using Python library SciPy (Virtanen et al., 2020) to detect “neuroplasticity networks,” i.e., networks of brain areas that share similar temporal patterns of MD change during learning. This analysis resulted in a division of the brain into five “networks” (the optimal number of clusters was determined by the silhouette value; Rousseeuw, 1987). To examine the correspondence between these “neuroplasticity networks” and known functional networks, we computed the Dice index between each of the five “neuroplasticity” and the seven canonical resting-state functional connectivity networks (Yeo et al., 2011). Since the Yeo networks do not include subcortical regions or the cerebellum, those areas were excluded for the purpose of this comparison. We have also preformed this network analysis with an alternative parcellation, using the Glasser Atlas (Glasser et al., 2016) and using partial correlation instead of full correlation, and found similar division to networks.
Classification analysis
To explore the relationship between the continuous change in MD and the behavioral manifestation of training, two support vector machine (SVM; Cortes and Vapnik, 1995) classifiers were trained. First, we attempted to classify between learning and control participants using a whole-brain approach. Second, we examined classification between high performers and low performers within the learning group, based on diffusion-detected changes following the learning task.
To classify between groups, we used the individual correlation matrices described in the previous section and tested whether learning and control participants can be distinguished based on the correlation profile of the whole-brain continuous MD change. We used a leave-two-out cross–validation routine, such that in each iteration, the model was trained on 28 participants from each group (56 participants in total) and tested on two participants—one from the learning and one from the control groups.
To classify between high and low performers within the learning group, we used the 10 best and 10 worst performers based on their speed and accuracy scores, separately. Features for this classification were the continuous MD curves of the parahippocampal gyrus (PHG), the hippocampus, the temporal gyrus, and the cerebellum (notably, these could not have been employed as features for the learning vs control classification as they differ between groups by definition). A separate SVM model was built for each ROI and each behavioral score, resulting in 10 models in total.
For both analyses, the chance success rate was calculated based on a permutation test with 1,000 iterations. For the high versus low performers classification, p values were FDR corrected for 10 comparisons (10 models).
Results
Behavioral results
The success of behavioral training was assessed as improvement in both accuracy and speed (Fig. 2). Comparing the first and last trials of learning, the mean improvement in speed was 42.9% (SE, 2.8%), and the mean improvement in accuracy was 67.2% (SE, 11.2%). In the test trial, performed after a 12 min break (inside the scanner), the improvement in speed was 52.3% (SE, 2.2%) and, in accuracy, 108.55% (SE, 14.7%) compared with the first trial. Differences between the first, last, and test trials were significant for both measurements (paired t test,
Behavioral effects of motor sequence learning. A, Averaged behavioral scores for 29 participants who performed a finger tapping task for 33:35 min (24 trials). The figure shows, on the right y axis, the mean sequence duration per trial and, on the left y axis, the mean number of correct sequences per trial. Error bars indicate the standard error per trial. B, Differences in performance between the first, last, and test trials (paired t test;
Diffusion-detected neuroplasticity at discrete timepoints throughout training
In a voxelwise two-way ANOVA of MD timepoint (before/after learning) × group (learning/control), we found a 1–4% MD decrease in the right PHG, the right hippocampus, the right inferior temporal gyrus (ITG), and the cerebellum (Fig. 3A,B). Similar clusters were detected (1) when pre- and postimages were calculated based on 1, 5, or 15 MD timepoints and (2) when applying smoothing of higher or lower Gaussian kernels (with a FWHM of 3 and 7 mm; Extended Data Fig. 3-1).
Microstructural neuroplasticity at discrete timepoints throughout motor sequence learning. A, We first examined the pre-/postlearning effect following training: A 2 (learning vs control groups) by 2 (pre/post) ANOVA revealed a significant interaction effect. Clusters of MD decrease following learning were found in the right PHG, the right hippocampus, the right ITG, and the cerebellum. The same statistical analysis was preformed using a different number of MD timepoints to define the pre- and postlearning data and using different smoothing kernels (Extended Data Fig. 3-1). B, MD change in the clusters displayed in A, for the learning (dark blue) and control groups (light blue). C, A three-point temporal course of MD change in the same clusters. Paired t tests were performed between each timepoint and the prelearning timepoint; asterisks denote a significant MD decrease in the learning group. Note that the rightmost bars in each plot in panel C are identical to those presented in panel B. Error bars indicate the standard error.
To further explore the evolving pattern of MD decrease during learning, we examined additional timepoints throughout the training process, beyond the traditional comparison of before versus after learning. Each panel in Figure 3C includes, therefore, three bars, depicting MD change between the “prelearning” baseline and the following timepoints: ∼11 min into the learning process; ∼22 min into the learning process; and after learning is completed, which is identical to the bars shown in Figure 3B. A rapid reduction in MD within the right temporal gyrus was observed after 11 min of learning, extending consistently until task's completion. A decrease in MD was also observed after 22 min of learning in the right PHG and the left cerebellum.
Continuous training-induced diffusion-detected neuroplasticity
While we have so far focused on MD at discrete timepoints, Figure 4 presents the continuous MD change during training (in Extended Data Fig. 4-1, we present the unfiltered continuous curve). Notably, within all clusters, the reduction in MD manifested gradually during training; however, distinct variations were detected across regions.
Continuous MD changes during learning. MD timeseries per cluster are shown after applying a low-pass filter of 0.0015 Hz (for unfiltered curves, see Extended Data Fig. 4-1). Curves were calculated individually and averaged across participants per group (standard errors were added around the continuous curves in lighter colors). While a decrease in MD was found in all areas for the learning group compared with the control group, the patterns of decrease differed across areas.
In Figure 5, we present the results of a changepoint analysis aimed to examine variations in the slope of MD decrease across clusters. Notably, in the PHG, we observed a relatively consistent slope throughout the training process, while in the hippocampus, we detected a significant changepoint occurring ∼17 min into the learning task. In both areas of the cerebellum, a change in the slope was found ∼21–22 min into the learning task. In the temporal gyrus, however, there is a slight increase in MD ∼20 min into the learning task, resulting in a steep decrease afterward. During subsequent phases of training, the MD decrease was considerably steeper compared with initial stages (Fig. 5A).
Changepoint detection analysis. A, For each of the five clusters, a change in the slope was determined as the maximum change between the slope of the two parts of the divided curve. The changepoint was calculated individually and averaged across participants and is presented on the average MD timeseries. Extended Data Figure 5-1 for the same analysis using different low-pass filters and different window sizes. B, Voxelwise changepoint detection within the left cerebellum. Color coding represents the time (minutes into the learning process) of changepoint for each voxel in the cerebellum. See Extended Data Figure 5-1 for the same analysis on all clusters.
It is worth highlighting that variations in the progression of MD decrease were not only evident between different brain regions but also within each area. For instance, in the cerebellum, we identified voxels displaying a change in slope around the 15 min mark, while others exhibit this change at later timepoints (Fig. 5B; see Extended Data Fig. 5-1A for the voxelwise changepoint analysis results in other clusters). To test the stability of the changepoint, we performed this analysis on first different cutoff frequencies for the low-pass filter (and on the unfiltered data) and second at different window lengths to examine whether the found changepoint represents the overall trend of the data and not a specific fluctuation (Extended Data Fig. 5-1B,C).
Neuroplasticity networks
We next searched for “neuroplasticity networks,” reflecting a network-level organization of brain regions that undergo similar diffusion-detected changes during training. To examine the similarities in the patterns of MD decrease across brain regions, we first parceled the brain into 273 areas based on the BNA. We then calculated the continuous MD for each parcel across all learning participants and computed the Pearson's correlations between each pair of parcels, resulting in a 273 × 273 correlation matrix. We then used hierarchical clustering to characterize five “neuroplasticity networks” (Fig. 6).
Neuroplasticity networks. A, A matrix depicting the pairwise Pearson's correlations between the patterns of MD decrease across 273 brain regions, ordered by a hierarchical clustering algorithm and parceled to “neuroplasticity networks” (black squares). B, The five networks displayed on a surface (top row) and a medial view (bottom row) of the brain. For the same analysis using the Glasser Atlas (Glasser et al., 2016), see Extended Data Figures 6-1 and 7-1. In addition, we tested our clustering on a similarity matrix composed using partial correlation (Extended Data Fig. 8-1).
Finally, we examined the correspondence between these networks and canonical functional connectivity networks (Yeo et al., 2011) by computing the Dice index between each pair of networks (Fig. 7). Similarities were detected between each of our “neuroplasticity networks” and either the limbic, the control, or the visual networks.
Similarities between neuroplasticity networks and functional connectivity networks. Similarity matrix showing the Dice index between each of our five main neuroplasticity networks (left) and Yeo's seven connectivity networks (bottom). Highest Dice indices were found between the neuroplasticity networks and the limbic network (Dice = 0.31), the control network (Dice = 0.32), and the visual network (Dice = 0.47). For the same analysis using the Glasser Atlas (Glasser et al., 2016), see Extended Data Figure 6-1.
For the same analysis with an alternative parcellation, using the Glasser Atlas (Glasser et al., 2016), see Extended Data Figures 6-1 and 7-1. We have also preformed this network analysis using partial correlation instead of full correlation and found similar division to networks (Extended Data Fig. 8-1).
Associations between diffusion-detected neuroplasticity and behavior
Whole-brain correlation matrices of continuous MD change patterns significantly discriminated between learning and control participants (model accuracy, 0.74; p = 0.002; Fig. 8, left column).
Classification between learners based on MD change patterns. Model accuracy for classification between learning and control participants based on the individual whole-brain correlation matrices (left column) and for classification between high and low performers based on the continuous MD change within the left cerebellum (two right columns). Chance levels (marked by the dashed lines) were calculated based on 1,000 permutations. The P value was FDR corrected for the left cerebellum.
Classifying between high and low performers, defined based on their accuracy scores, was successful when using the continuous MD change patterns in the left cerebellum as features for prediction (model accuracy, 0.75; p = 0.03; FDR corrected). Classification between high and low performers, defined based on their speed scores, was not significantly successful (Fig. 8, right columns).
Discussion
In this work, we investigate continuous training-induced diffusion–detected remodeling of the brain tissue, leveraging a unique diffusion MRI protocol to continuously assess MD in the gray matter throughout a motor training routine. This protocol allowed us to detect not only changes between discrete timepoints, as was done previously and also shown here in Figure 3, but also the continuous change pattern throughout the entire training process, as shown in Figure 4. Different patterns of MD decrease were found in several brain areas, suggesting their involvement at different stages of training. Specifically, while the PHG demonstrated a relatively consistent slope of MD decrease, the hippocampus and cerebellum showed a significant change in slope 17–20 min into the training process. Furthermore, we were able to capture different change patterns across single voxels within a cluster, suggesting high sensitivity of our method to the spatiotemporal properties of training-induced changes. Based on similarities in the patterns of MD decrease across the brain, we detected “neuroplasticity networks,” consisting of areas that undergo similar diffusion-detected alterations during training. Finally, we were able to classify between learners and control participants, as well as between high and low performers within the learning group, based on patterns of continuous MD change, suggesting that these patterns are associated with behavioral learning outcomes.
Our main goal was to examine where do signatures of neuroplasticity develop in the brain and how they change over ongoing training. While the use of diffusion MRI to study short-term plasticity has gained increasing interest over the last decade (Sagi et al., 2012; Tavor et al., 2013, 2020; Keller and Just, 2016; Hofstetter et al., 2017; Brodt et al., 2018; Jacobacci et al., 2020; Villemonteix et al., 2023), these studies have all investigated the differences in brain structure between two timepoints: before and after training. This work is therefore the first to explore the temporal dynamics of these diffusion-detected changes as they occur during the encoding stage of the learning process itself.
Motor sequence learning requires a complex ensemble of distributed brain regions which are involved at different stages of learning (Penhune and Steele, 2012; Dahms et al., 2020; Tavor et al., 2020). Based on evidence from functional imaging studies (e.g., fMRI or PET), it has been suggested over 20 years ago that early stages of motor learning involve corticocerebellar and corticostriatal mechanisms, whereas later stages of consolidation and retention engage the striatum, motor, and parietal cortices (Toni et al., 1998; Doyon et al., 2003). Our results go beyond these classical models of learning by, first, describing diffusion-detected rather than functional dynamics and, second, by “zooming in” into the very first minutes of motor training, tracking continuous changes and detecting the exact moments in time when different brain areas join the process.
Of particular interest is the involvement of the hippocampus, which has been rarely reported in studies of functional plasticity following motor learning (Coynel et al., 2010; Berlot et al., 2020) but is in line with recent diffusion MRI studies showing an early engagement of the hippocampus during learning (Brodt et al., 2018; Jacobacci et al., 2020). Our findings may support a role of the hippocampus in short-term memory stabilization (Schapiro et al., 2019) or in connecting encoded items of different modalities over space or time (Buzsáki and Tingley, 2018). Another motor-related area that emerges from our study is the cerebellum, which is known to be involved in the fast-learning stage of motor learning (Ungerleider et al., 2002). Specifically, we show that continuous MD changes in the left cerebellum have successfully classified accuracy-based high and low performers of the learning task. This finding corresponds with previous reports of associations between learning-induced cerebellar modifications and accuracy of learning (Sagi et al., 2012).
While it is evident that the unique diffusion MRI protocol developed here is sensitive to dynamic, flexible brain modifications over short timescales, the biological underpinnings of such modifications are yet to be fully understood. A possible explanation is that training-induced alterations in MD may reflect the rapid modification of astrocytes structures during and following learning (Johansen-Berg et al., 2012; Sagi et al., 2012; Tavor et al., 2013; Assaf, 2018).
The division to “neuroplasticity networks” based on similarities of MD change patterns across brain regions offers further insights into the complex, multiregional process that ultimately gives rise to learning and memory. These networks are not to be confused with the widely studied functional or structural networks, which reflect the temporal synchronization in brain activity across regions or the physical connections between them, respectively. Considering the accumulating evidence on a high correspondence between functional connectivity and brain activity (Smith et al., 2009; Tavor et al., 2016; Tik et al., 2021, 2023; Gal et al., 2022) and the recent suggestion that functional network architecture may relate brain activity and behavior (Bijsterbosch et al., 2020; Bernstein-Eliav and Tavor, 2024), a network-organization approach may facilitate the study of neuroplasticity and its behavioral manifestation as well. Specifically, grouping areas by patterns of microstructural changes may shed new light on the intricate mechanisms underlying skill acquisition.
An examination of the hierarchical clustering results reveals that the largest “neuroplasticity networks” may roughly correspond with the functionally defined limbic and control networks (Yeo et al., 2011), which is consistent with the known involvement of frontal, temporal, and parietal cortices in motor sequence learning (Witt et al., 2008; Turesky et al., 2018). This correspondence is particularly intriguing as Yeo's networks are defined based on the functional MRI signal during resting-state, while our “neuroplasticity networks” are driven from diffusion properties during task performance. It may therefore provide converging evidence for neuroplasticity mechanisms supported by brain areas within the control network, including the prefrontal, premotor, and parietal cortices which play a critical role in motor learning and execution. Interestingly, disrupted functional connections between areas in the control and limbic networks have also been associated with motor deficits in aging (Michely et al., 2018), neurodegenerative disorders (Poston and Eidelberg, 2012), or following a stroke (Siegel et al., 2016).
This work introduces a novel method to continuously track diffusion properties throughout learning. Our results provide a promising proof-of-concept, yet further research is required to examine methodological aspects and optimize parameters. It should also be considered that due to the nature of MD, which is calculated based on a series of DWI acquired from various directions at intervals of a few seconds (depending on TR), the temporal resolution of our method is lower than that of functional imaging methods such as fMRI or even functional diffusion MRI (Le Bihan, 2003; Le Bihan and Johansen-Berg, 2012; Tsurugizawa et al., 2013; De Luca et al., 2019; Nunes et al., 2019, 2021). While in this paper we focused on the low frequencies compiling the continuous MD change to assess the trend of MD decline, the raw data also show some similarities to the functional dMRI suggested by Le Bihan (2003). The raw continuous MD signal most likely reflects both the functional aspect of diffusivity change and their microstructural origins during training. Taken together, the rapid onset of the dMRI signal (Nunes et al., 2021) and the continuous progression of the MD signal during task performance, as described here, may offer a more comprehensive picture on the temporal dynamics of learning-induced changes. Further research is required to address the correspondence between these different protocols and integrate between them.
Additionally, in this study, we employed a relatively simple motor task; future work may apply our method to study higher-level cognitive tasks and complex learning routines. Insights on the timing and the network-level organization of learning-induced changes may extend previous findings on neuroplasticity following spatial navigation (Sagi et al., 2012; Tavor et al., 2013; Keller and Just, 2016; Villemonteix et al., 2023), language (Hofstetter et al., 2017), and associative memory tasks (Brodt et al., 2018). Notably, the current training procedure included a test block performed 12 min following learning completion, but no contralateral tests (Karni et al., 1995) or retention after longer time periods (Lohse et al., 2014). The lack of additional tests limits our ability to fully differentiate between learning-dependent and use-dependent effects. Although the successful classification of high and low performers based on continuous MD change patterns may support a learning-depended account, future studies may benefit from combining our continuous diffusion approach with more elaborated, longitudinal training paradigms to fully disentangle learning-dependent from use-dependent effects.
Future studies are also required to examine the signal decomposition of the continuous MD timeseries, which may show spontaneous fluctuations alongside training-related effects. For example, the control group in the present study showed higher MD dynamics (mainly in the temporal gyrus and the left cerebellum) which are unexplained by task manipulation. This can be originated in unrelated biological issues or physical issues relating to the MRI protocol (Jones and Cercignani, 2010) but may also reflect “resting-state”–like physiological changes in tissue diffusivity.
In conclusion, we provide evidence for the utility of a continuous acquisition of diffusion MRI during task performance to explore the dynamic aspects of diffusion-detected neuroplasticity. Rather than a single, postlearning “snapshot” that indicates that changes have occurred at some, undetectable, point and pace, we show that different task-related brain areas demonstrate distinctive temporal patterns of training-related modifications. The continuous characterization of neuroplasticity is therefore a promising approach, offering a temporally sensitive, network-level view on the orchestrated processes underlying learning and memory in the human brain.
Footnotes
Funding for this research was provided by the Minducate Science of Learning Research and Innovation Center of the Sagol School of Neuroscience, Tel Aviv University, and the Israeli Science Foundation (ISF Grant Number 1790/24). Data used for method optimization were provided by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657), funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research and by the McDonnell Center for Systems Neuroscience at Washington University.
The authors declare no competing financial interests.
This paper contains supplemental material available at: https://doi.org/10.1523/JNEUROSCI.1152-24.2025
- Correspondence should be addressed to Ido Tavor at idotavor{at}tauex.tau.ac.il.
