Abstract
People with cerebellar degeneration show characteristic ataxic motor impairments. Despite cerebellar dysfunction, they can improve motor performance through sensorimotor training. Yet, how such training changes functional brain networks affected by cerebellar degeneration is unknown. We here investigated neuroplastic changes in the cortico-cerebellar network after a 5-d forearm training in 40 patients with mild to severe cerebellar degeneration and 40 age- and sex-matched healthy controls. Human participants (41 female) were assigned to one of four motor training conditions, varying online visual feedback and explicit verbal feedback. Anatomical and resting-state fMRI was collected on days before and after training. To overcome the limitations of standard brain templates that fail when encountering severe anatomical abnormalities, we developed a specific template for comparing cerebellar patients with controls. Our new template reduced the spatial spread of cerebellar anatomical landmarks by 30% and tripled fMRI noise classification accuracy. Using this pipeline, we found that patients showed impaired connectivity between cerebellar motor regions and neocortical visuomotor and premotor regions at baseline compared with controls, whereas their corticocortical connectivity remained intact. Training with vision strengthened connectivity in the cortico-cerebellar visuomotor network contralateral to the trained arm in all participants. Cerebellar patients exhibited additional increased connectivity ipsilateral to the training arm in this network. Further, training with explicit verbal feedback affected connectivity between a cerebellar cognitive region and dorsolateral prefrontal cortex, although post hoc tests did not reach significance. These results document enhanced cortico-cerebellar connectivity as a neurophysiological response to visuomotor training in people with cerebellar degeneration.
- cerebellum
- functional connectivity
- motor training
- neurodegeneration
- resting-state fMRI
- spinocerebellar ataxia
Significance Statement
Despite cerebellar dysfunction, patients with cerebellar degeneration can still improve motor function with training. However, the neuroplastic changes in the brains of cerebellar patients underlying these training effects are largely unknown. Using a specialized brain template and a large patient-control sample, this study demonstrates that sensorimotor training increases connectivity in the cortico-cerebellar network of cerebellar patients, specifically between cortical and cerebellar regions where patients show impairments at baseline. Importantly, these changes were feedback-dependent, with visuomotor training increasing connectivity contralaterally in all participants and ipsilaterally in cerebellar patients. This work provides critical insights into how targeted interventions can engage residual brain function in cerebellar degeneration. Additionally, the novel brain template offers a valuable tool for future neuroimaging studies of cerebellar degeneration.
Introduction
Dysfunction of the cerebellum causes deficits in the coordination and control of gait, posture, speech, and fine motor movements. While the recovery of motor function after acute lesions of the cerebellum such as stroke typically happens within weeks and is often complete (Konczak et al., 2010), cerebellar neurodegeneration occurs over years and the ataxic motor symptoms worsen over time. For those affected by cerebellar degeneration, no medication has been proven effective, and potential gene therapies are still in the early stages of development (Manto and Marmolino, 2009; Marmolino and Manto, 2010; Ilg et al., 2014). Existing therapies in patients with cerebellar degeneration therefore focus on mitigating symptoms through physical rehabilitation (Ilg and Timmann, 2013; Ilg et al., 2014). Continuous and intensive motor training has been shown to improve motor function in cerebellar patients (Ilg et al., 2009, 2012; Miyai et al., 2012; Burciu et al., 2013; Keller and Bastian, 2014). However, no consensus exists on which type of training is most beneficial for cerebellar patients (Ilg et al., 2014). Understanding the neuroplastic changes that result from diverse types of motor training could help aid the development of tailored rehabilitation strategies for cerebellar patients.
Motor learning can be defined as change in movement parameters—speed, accuracy, and smoothness—due to experience or practice (Haibach et al., 2011). Motor learning can involve implicit learning processes that rely on processing movement-relevant information from multiple sensory modalities such as vision and proprioception. It relies on an intact network comprising the cerebellum, the basal ganglia, and sensorimotor cortex to modulate spinal or brainstem motor neurons. Given that the cerebellum is essential for motor learning and cerebellar degeneration impairs such learning, the utility of movement training programs has been questioned (Thach and Bastian, 2004; Bastian, 2006; Saywell and Taylor, 2008). Indeed, empirical evidence shows that the ability to adapt to external forces, a form of implicit motor learning, can become abolished in the severe stages of cerebellar degeneration (Maschke et al., 2004). However, recognizing the degenerative process typically spans over years, motor training has also been shown to improve motor function in cerebellar patients at mild to moderate stages of cerebellar degeneration (Ilg et al., 2009, 2012; Miyai et al., 2012; Burciu et al., 2013; Keller and Bastian, 2014). The extent of learning on a force adaptation task has been found to correlate with the severity of ataxia in cerebellar patients with the most severely affected patients showing the smallest amount of learning (Maschke et al., 2004). To date, no physical rehabilitation programs exist that take the amount of residual sensorimotor function present in the patient into account.
It is well established that proprioceptive afferent signals are vital for motor learning and that the cerebellum receives large proprioceptive afferents through the spinocerebellar tracts (Bloedel, 1973). In this context, it is noteworthy that proprioceptive perception is largely intact in people with cerebellar dysfunction as studies reported unimpaired limb position sense when the limb is passively rotated and no voluntary muscle activation is required (Maschke et al., 2004; Bhanpuri et al., 2012). However, cerebellar patients show impaired active position sense that relies on voluntary movement (Bhanpuri et al., 2012). A training protocol that focuses on proprioceptive cues could potentially drive learning in cerebellar patients, but the extent to which patients can use proprioceptive information remains elusive.
Explicit error feedback given by a teacher, for example, can aid motor learning. This indicates that additional cognitive processes involving additional networks occur in parallel and can contribute to motor learning outcomes (Taylor et al., 2014; McDougle et al., 2015). Research showed that cerebellar patients rely on explicit re-aiming strategies to counter rotations in a visuomotor adaptation task (Taylor et al., 2010), which implies that they may benefit from additional explicit feedback during motor training.
Moreover, animal work suggests that training can slow down the process of degeneration (Fryer et al., 2011; Safdar et al., 2011). For people with cerebellar degeneration, motor training over several days led to structural changes in the neocortex, specifically gray matter increase in the premotor cortex (Burciu et al., 2013; Draganova et al., 2022). Documenting changes in functional and structural connectivity can provide an additional, more detailed description of the initial impairment and associated training-related changes of the sensorimotor network. Functional activity and connectivity alterations have repeatedly been found in cerebellar patients (Georgiou-Karistianis et al., 2012; Hernandez-Castillo et al., 2013; Zalesky et al., 2014; Harding et al., 2016; Tzvi et al., 2017). Functional connectivity measured by resting-state fMRI has been shown to reliably dissociate patients from controls at 92% accuracy (Hernandez-Castillo et al., 2014). Functional connectivity changes might therefore hold potential for monitoring disease progression in response to putative therapeutic interventions.
However, a major obstacle to capturing connectivity changes in people with cerebellar degeneration are inadequate neuroimaging analysis pipelines. Normalizing degenerating brain structures to brain templates that are developed on neurologically healthy brains is challenging (Ashburner and Friston, 2000; Good et al., 2001; Thompson et al., 2001; Evans et al., 2012), which limits our ability to detect patient-control differences at the group level. This is compounded by difficulties in aligning cerebellar structures (Diedrichsen, 2006), low sensitivity of MRI acquisition coils to cerebellar signal (Risk et al., 2018), and high noise levels induced by cardiac and respiratory cycles in areas close to the brainstem (Brooks et al., 2013). Hence, tailoring analysis pipelines to preserve cerebellar signal, particularly in patient populations, is crucial to revealing meaningful changes in connectivity.
In this study, we investigated the neuroplastic changes in cortico-cerebellar connectivity in a sample of patients with cerebellar degeneration induced by extensive motor training of their dominant right arm. The relevant behavioral and structural data have been published in Draganova et al. (2022). The results showed that participants improved performance during visuomotor practice that was accompanied by an increase in premotor cortex gray matter. However, a training focusing either on proprioceptive feedback or explicit verbal feedback did not yield additional benefits. To delineate changes in functional connectivity associated with motor training, we developed a neuroimaging pipeline that substantially improved spatial localization by building a specialized template for comparing cerebellar patient data to data from healthy controls.
Materials and Methods
Participants
Forty-one patients with cerebellar degeneration and 44 neurologically healthy individuals participated in this study. One patient and two controls dropped out before study completion due to acute illness unrelated to the study, and two control participants were removed from analysis due to incidental findings, resulting in a final sample of 40 patients (age, 55 ± 11.29 years; 21 females) and 40 age- and sex-matched controls (age, 55 ± 10.83 years; 20 females). For detailed study demographics, see Draganova et al. (2022). Patients were diagnosed with a pure form of cerebellar cortical degeneration. Predominant diagnosis were spinocerebellar ataxia type 6, autosomal dominant cerebellar ataxia type 3, and sporadic adult-onset ataxia of unknown etiology. Ataxia severity was assessed using the clinical Scale for the Assessment and Rating of Ataxia (SARA; Schmitz-Hübsch et al., 2006). Participants were pseudorandomly assigned to one of four motor training conditions. Training groups were matched for sex, age, and ataxia severity (SARA scores), and all participants were right-handed as assessed by the Edinburgh handedness scale (Oldfield, 1971). The study was approved by the Ethics Committee of the Essen University Medical Center, and all participants gave verbal and written informed consent prior to testing.
Motor training
During a 5 d motor training, participants performed elbow flexion movements in the horizontal plane using a single-joint manipulandum to targets arranged on a semicircular metal frame at 95 cm distance [Fig. 1A; manipulandum and task details can be found in Draganova et al. (2021, 2022)]. Participants began each trial in a start position of 90° elbow flexion. After a start command, participants performed a single, ramp-like movement without online correction to the target. Participants held their arm in the end position for 4 s before moving it back to the start position. Participants were instructed to perform swift and accurate movements. If movements were too slow or too fast according to previously calculated limits, the experimenter provided correctional feedback to reduce or increase movement speed.
Study design and motor training. A, Experimental setup of the single-joint manipulandum used for motor training. Start position was at 90° elbow flexion. A low-intensity laser attached to the manipulandum provided visual feedback of arm position online. An optical encoder embedded in the housing recorded forearm position. Figure reproduced from Draganova et al. (2022). B, Participants participated in a 5 d training with MRI sessions before (pre) and after training (post). Anatomical T1-weighted and resting-state fMRI data was collected during pre and post MRI sessions.
On each training day, participants performed 100 trials to targets located counterclockwise at 10 and 50°, in blocks of 25 trials that alternated between the two targets. The target of the first block was balanced across participants of each training group and alternated each day. Participants began each training day by performing three familiarization trials to the first target of that day. After each block, participants rested for 3 min, and after completing two blocks, participants relaxed their arm outside of the manipulandum.
To continuously drive performance, the width of the targets was reduced according to the movement accuracy achieved by the participant. Participants began at the lowest performance level with a target width of 4.5 cm, corresponding to 2.7° on the frame. After performing five consecutive movements where the beam of the laser pointer was on target, the target width was reduced to the next smaller width (4.5 cm/2.7°, 3.5 cm/2.11°, 2.5 cm/1.5°, 1.5 cm/0.9°, 0.5 cm/0.3°). Participants were instructed about the five levels prior to the experiment and notified by the experimenter as they advanced to the next level. Behavioral results have been reported previously (Draganova et al., 2021, 2022).
Conditions
Participants were pseudorandomly assigned to one of four motor training conditions, varying online visual feedback (Vision/No Vision) and postmovement verbal feedback (Expl. Feedback/No Expl. Feedback). Ten patients and 10 controls performed each condition (Vision | No Vision | Vision + Expl. Feedback | No Vision + Expl. Feedback). In conditions without vision, participants wore a mask that fully occluded vision. Here, the experimenter guided the arm of the participant from the start position to the target before each trial to enable the participant to memorize the target position. After moving back to the start position, participants performed the movement to the target. The experimenter informed participants verbally about the movement outcome, indicating whether the movement was “on target” or “target has not been reached.” In conditions with explicit feedback, participants received verbal feedback after each movement about the final joint position relative to the target and strategies to minimize future movement errors (i.e., “Target was undershot by 5 degrees. Increase movement by 5 degrees”).
Each trial lasted 10 s in the vision-only condition and 15–20 s when verbal feedback was given in addition to vision. In the conditions without vision, trials lasted 50–60 s. Therefore, each training session with vision lasted 45–60 min and 90–100 min without vision.
MRI acquisition
Magnetic resonance imaging data was collected on the days before and after training (Fig. 1B) on a 3 T combined MRI-PET scanner (Siemens Healthcare) with a 16-channel head coil with the same scanning protocol. Structural MRI data was acquired using a magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence with the following parameters: TR, 2,530 ms; TE, 3.26 ms; inversion time, 1,100 ms; flip angle, 7; FOV, 256 × 256 × 176 mm3; bandwidth, 200 Hz/pixel; GRAPPA acceleration factor 2 and 48 reference lines; whole brain, acquisition time, 6:24 min. Resting-state fMRI data was acquired with the following parameters: TR/TE, 3,000/35 ms; FOV, 282 × 282 × 134 mm3; bandwidth, 2,312 Hz/pixel; voxel dimension, 3 × 3 × 2.7 mm3; GRAPPA acceleration factor 3 and 90 reference lines; whole brain, acquisition time, 5:17 min for a total of 100 volumes. During resting-state fMRI acquisition, participants fixated on a cross-hair image presented centrally on the screen. No functional data was acquired for one healthy control participant at both timepoints and for another healthy control participant at timepoint post due to scanner failure. The first subject was therefore excluded from functional analysis, while the second one was included for baseline connectivity but excluded for connectivity change analysis.
MRI analysis
All MRI data was processed using the functional MRI of the Brain Software Library (FSL; version 6.0.4; Jenkinson et al., 2012). Structural images were additionally processed using advanced normalization tools (ANTs; Avants et al., 2010) and ITK-SNAP (Yushkevich et al., 2006).
Structural MRI analysis
Structural images were reoriented to standard MNI orientation, cropped to remove superfluous neck tissue, and corrected for RF/B1 field inhomogeneities using FSL's fsl_anat (Jenkinson et al., 2012). Brain masks were generated using the Optimized Brain Extraction for Pathological Brains (optiBET; Lutkenhoff et al., 2014). To ensure precise tracing of the border between brain and nonbrain tissue, in particular for degenerated brains, brain masks were corrected by hand by an experienced technician otherwise blinded to the study. Using the hand-corrected masks, the structural images were skull stripped.
Template generation
Given the anatomical abnormalities present in brains with cerebellar degeneration and in scans of older adults (Good et al., 2001; Ge et al., 2002; Sato et al., 2003; Sullivan et al., 2004; Taki et al., 2004; Fotenos et al., 2005; Lemaître et al., 2005; C. D. Smith et al., 2007), it was important to use a registration template that is appropriate for patient brains and older adults. We therefore constructed a study-specific template from the scans of 40 participants. The other half of the sample was held out for validation purposes (see below, Template validation). Scans that showed highly abnormal anatomy or particularly severe pathology, such as enlarged ventricles or strongly degenerated cerebellum, were not used for template generation but included in the validation sample. To avoid biasing the template to a specific group or training condition, the template was constructed from 20 patient and 20 healthy control brains balanced across conditions (Table 1). Additionally, the scans were chosen to be representative of the mean sample age and sex distribution and balanced across scan timepoints.
Template generation and validation sample demographics
The study-specific template was generated from the brain-extracted structural images using the iterative process implemented by the advanced normalization tools (ANTs, antsMultivariateTemplateConstruction2; Avants et al., 2010). As a starting point for the template construction, an existing group template derived from brains aged 55–59 was selected, which presented the closest match to the mean age of our sample (obtained from Neurodevelopmental MRI Database; Fillmore et al., 2015). In brief, the structural images were rigidly, then linearly, and finally nonlinearly registered to the starting template. The resampled images were averaged and deformed using the inverse average deformation to obtain a true mean shape image. This image served as the new starting template for the next iteration. To optimize convergence, ANTs downsample and smoothen the images before registration and repeat the iterative process at progressively finer resolution and smoothing kernels. Our template underwent five refinement stages at 8 × 6 × 4 × 2 × 1 mm image resolution, 6 × 4 × 2 × 1 × 0 voxel smoothing kernels, and 500 × 200 × 100 × 70 × 20 maximum iterations. We hereby created a new template for comparing degenerated brains with control brains (DegenerationControl template, DeCon). To obtain a mapping between the DeCon template and MNI template, the DeCon template was nonlinearly registered to the MNI template using antsRegistrationSyn. The resulting nonlinear warp field was converted to an fsl-compatible format using the c3d_affine_tool from ITK-SNAP (Yushkevich et al., 2006) and FSL's fslulils (Jenkinson et al., 2012).
All structural images were nonlinearly registered to the template using antsRegistrationSyn (Avants et al., 2010). To enable comparisons between the DeCon template and existing templates, we additionally nonlinearly registered the structural images to the MNI template (Evans et al., 2012; nonlinear 6th generation MNI152 template or MNI152NLin6Asym available at https://www.templateflow.org/, hereafter referred to as “MNI template”) and the SUIT template (Diedrichsen, 2006) using antsRegistrationSyn (Avants et al., 2010), all at 1 mm resolution (with precision type set to float and histogram matching turned on and all other parameters set to default values). Registration to the SUIT template was performed after isolating the cerebellum from the surrounding tissue in the structural scan using the SUIT isolate function (Diedrichsen, 2006).
Template validation
We evaluated the degree of anatomical correspondence between cerebella in group space by calculating the spatial spread of anatomical landmarks of the cerebellum, using a procedure established by Diedrichsen (2006). As landmarks, we chose two major fissures, the primary fissure separating lobules V and VI and the intrabiventer fissure VIIIa and VIIIb. Fissures were drawn as surfaces on the structural scans in native space of each participant by an experienced technician otherwise blinded to the study. The fissures were then resampled to DeCon space, SUIT space, and MNI space. Spatial overlap was calculated as the average distance between corresponding fissures of each possible pair of participants. Specifically, we computed the smallest distance of each point of fissure A to a point on the fissure B and averaged these first across all points of the fissure A and then across all participants. Fissure alignment was calculated both for the entire cerebellum, using a cerebellar mask in group space, and restricted to the vermal section, using a mask obtained from the SUIT atlas (Diedrichsen et al., 2009).
As a second evaluation criterion, we compared the ability of the automated noise classifier FIX (Salimi-Khorshidi et al., 2014) to identify noise components of functional data using structural-to-MNI registrations estimated either directly to MNI space or with the DeCon template as an intermediate step. Since noise driven by vein fluctuations is dissociable from neural signal only by voxel location in major draining veins, FIX uses MNI space masks of the major veins resampled to native functional space to identify vein noise (Salimi-Khorshidi et al., 2014). Thus, FIX's ability to detect noise relies in part on an accurate mapping between MNI space and native functional space. We therefore compared FIX's classification accuracy using functional-to-MNI registrations estimated directly or via the DeCon template.
Functional MRI analysis
The resting-state functional MRI data acquired before and after the training days was preprocessed using steps largely matched to the UK Biobank processing pipeline (Alfaro-Almagro et al., 2018). The data was motion corrected using MCFLIRT (Jenkinson et al., 2012), slice-timing corrected using Fourier-space time series phase-shifting, stripped of nonbrain voxels using BET (Brain Extraction Tool; J. K. Smith et al., 2002), and high-pass temporally filtered with a cutoff frequency of 100 s. No spatial smoothing was applied. Functional data were registered to the MNI space via the DeCon template and the structural scan of each subject and session. To this end, functional data were first linearly registered to the subject's structural image acquired in the same session using boundary-based registration (implemented in FLIRT, Jenkinson and Smith, 2001; Jenkinson et al., 2002; Greve and Fischl, 2009). The linear transformation was concatenated with the structural-to-DeCon nonlinear warp field and the DeCon-to-MNI nonlinear warp field into a functional-to-MNI space warp field.
FIX training
Structured noise was removed from functional data using a single-subject independent component analysis (ICA), implemented in MELODIC (Multivariate Exploratory Linear Optimized Decomposition into Independent Components) and the automated classifier FIX (Salimi-Khorshidi et al., 2014). Because existing labeled training datasets differed substantially in their study population and acquisition parameters from our data, we created our own training set by manually classifying a subset of scans (Griffanti et al., 2014; Salimi-Khorshidi et al., 2014). We selected 41 of the 157 scans in the dataset (26%), balanced across groups, conditions, and timepoints for hand classification. To avoid optimizing our analysis pipeline to any set of scans, we minimized overlap between scanning sessions selected for template generation and for hand classification, with scans from eight subjects and timepoints included in both.
FIX evaluation
Classification accuracy was evaluated using leave-one-subject-out cross-validation where ICA components from a single hand-labeled subject are left out of the training process and the classification accuracy is then tested on the left-out subject (Salimi-Khorshidi et al., 2014). Classification accuracy was summarized in terms of true positive rate (“TPR”), i.e., the percentage of correctly detected signal components, and true negative rate (“TNR”), i.e., the percentage of correctly detected noise components. A commonly used overall performance measure was calculated as the weighted average of TPR and TNR:
Functional connectivity
We quantified functional connectivity between neocortex and cerebellum using a region of interest (ROI)-based approach for motor regions and regions involved in cognitive aspects of motor planning. For the neocortex, we derived an ROI of the left and right primary motor cortex (M1) hand area based on previous functional MRI data of hand movements (Weinrich et al., 2017). A dorsal premotor cortex ROI (PMd) was defined for the left and right hemisphere from a tractography-based atlas (Tomassini et al., 2007). A posterior parietal cortex (PPC) ROI was defined as left and right area 7 of the Juelich histological atlas (Amunts et al., 2020) thresholded at 15. Finally we defined an ROI for the right dorsolateral prefrontal cortex (dlPFC) as area 46d of the Sallet atlas (Sallet et al., 2013) thresholded at 25.
To define cerebellar ROIs, we used a novel functional atlas of the cerebellum based on multiple fMRI datasets that parcellates the cerebellum into 32 symmetrical functional regions (Nettekoven et al., 2023). We defined cerebellar motor ROIs as the left- and right-hand motor regions M3L and M3R, which activate strongly in response to movements of the ipsilateral hand. To explore connectivity changes in cerebellar cognitive regions in response to explicit verbal feedback, we used regions D2L and D2R. D2 was chosen because it shows highest activity during rapid processing of instructive signals in a motor task (Nettekoven et al., 2024, their Supplementary Figure S2) and its specific involvement in working memory recall (Shahshahani et al., 2024), a cognitive function that deteriorates in cerebellar degeneration (Hoche et al., 2018). D2 also functionally connects to prefrontal regions of the neocortex (Nettekoven et al., 2023). Finally, as a control region thought to not be involved in motor processing, we chose region S1, which shows strong activation during tasks involving linguistic processing, such as verb generation or story reading, but no activity during left or right hand movements (Nettekoven et al., 2024).
We quantified functional connectivity as the Pearson's correlation coefficient between the resting-state time courses of two ROIs. To ensure normality, we used Fisher's r-to-z-transformation on the connectivity values. For visualization purposes, we used non-normalized connectivity values.
Statistical analysis
Mean ± SD is presented throughout. Statistical analyses were conducted using R (version 4.3.1; R Core Team, 2017) and Python (version 3.9.10). We used two-tailed paired t tests to test for differences in fissure alignment and noise identification (scipy version 1.12.0; Virtanen et al., 2020).
To test for connectivity differences at baseline, we constructed a linear mixed effects model of baseline connectivity values using the lme4 package (version 1.1-35.1). We averaged connectivity values within subject across regions, such that each subject had one corticocortical, one cortico-cerebellar, and one cerebello-cerebellar connectivity value. We included fixed effects of group (patient, control), region (corticocortical, cortico-cerebellar, cerebello-cerebellar), and a group × region interaction term.
To investigate changes in functional connectivity and how these depend on group and training condition, we constructed separate linear mixed effects models for each pair of ROIs. We included timepoint, group, vision, and feedback as fixed effects, along with their higher-order interaction terms. We allowed intercepts for different subjects to vary in the linear mixed effects models, to account for covarying residuals within subjects. Separate models of connectivity with M1, PMd, PPC, and dlPFC were constructed, as the hypotheses regarding the specific fixed effects of interest were different for different regions. Residuals of all linear models were confirmed to be homoscedastic by visually inspecting the residuals plotted against the predicted values of each independent variable. Normality of fixed effects residuals and random effect residual was confirmed by visual inspection of the Q-Q plot using the ggpubr package (version 0.6.0). P values were obtained with Type III tests calculated with Satterthwaite's method (alpha = 0.05) implemented in the lmerTest package (version 3.1.3; Kuznetsova et al., 2017). In models with significant interactions, pairwise differences were assessed at each level using independent t tests for patient-control differences and paired t tests for comparing connectivity before and after training. P values for post hoc tests were corrected for multiple comparisons using the Benjamini–Hochberg method (Benjamini and Hochberg, 1995), where all patient-control comparisons or all pre-post comparisons formed one family of tests. For testing connectivity differences at baseline, region combinations formed a family of tests. Since tests between regions are not independent, we corrected for multiple comparisons using the Benjamini–Yekutieli method (Benjamini and Yekutieli, 2001), as this method does not assume independence of tests. We calculated effect sizes as standardized coefficients β along with their 95% confidence interval (CI) using the standardize_parameters function implemented in the effectsize package (version 0.8.6; Ben-Shachar et al., 2020). In some cases, standardized coefficient confidence intervals include zero, even when the corresponding p value is significant. This is because the p values from the linear mixed effects models (derived using Type III ANOVA with Satterthwaite's method) assess whether a factor as a whole explains significant variance in the model, accounting for its full structure. In contrast, the standardized coefficients and their confidence intervals (estimated via standardize_parameters) represent single model terms on a standardized scale. These two metrics are therefore not directly comparable (Greenland et al., 1991). To balance statistical inference with interpretability, we report both types of results: p values from the linear model for testing significance and standardized coefficients for estimating effect size, direction, and variability. Finally, we used a Grubbs test to detect potential outliers in the behavioral data. The Grubbs test identified five outlier subjects whose functional connectivity data was removed before repeating the main analyses to show robustness of findings against the influence of outliers.
Results
Behavioral results have been reported in Draganova et al. (2022). In brief, cerebellar patients and controls showed improvement in motor performance in relative joint position error (RJPE) after training. Training-related improvements were observed for all targets. Training with vision led to greater reductions in RJPE compared with training without vision, unless it was paired with explicit verbal feedback for the 50° target. Healthy control participants generally outperformed cerebellar patients, but no significant interactions between group, training condition, and time were found.
Study-specific template improves spatial overlap
Anatomical abnormalities present in older adults and brains with cerebellar degeneration (Good et al., 2001; Ge et al., 2002; Sato et al., 2003; Sullivan et al., 2004; Taki et al., 2004; Fotenos et al., 2005; Lemaître et al., 2005; C. D. Smith et al., 2007) pose a challenge for accurate registrations to standard brain templates which are based on young, healthy adults. To improve spatial registration, we therefore developed the DeCon template, a template derived from the brains of cerebellar degeneration patients and controls in our study (Fig. 2A; see Materials and Methods, Template generation). We used 40 anatomical scans, balanced across groups, timepoints, and conditions (Table 1). To validate the use of the DeCon template, we examined the spatial overlap of anatomical landmarks after registration to the DeCon template (see Materials and Methods, Template validation). We compared this to the spatial overlap after registration to two standard templates used widely in neuroimaging analysis of the cerebellum which are based on young, healthy brains: the MNI template (Evans et al., 2012) and the SUIT template (Diedrichsen, 2006). To ensure that our comparison is not biased toward the DeCon template, as it represents the average anatomy of those specific individuals, we repeated our analysis with anatomical scans from the 40 independent participants that were not used for template generation (validation sample).
DeCon template improves anatomical alignment and functional localization. A, Primary (yellow) and intrabiventer (turquoise) fissures registered to the MNI, SUIT, and DeCon template show the highest spatial overlap in DeCon space. Only fissures of participants that were not used for template generation are shown (n = 40). B, C, Fissure overlap, computed as the average fissure distances between all possible pairs of participants, in MNI space (orange), SUIT space (magenta), and DeCon space (blue) is lowest in DeCon space. B, Primary fissure distance calculated within the whole cerebellum. C, Intrabiventer fissure distance calculated within the whole cerebellum. D, Primary fissure distance calculated within vermis. E, Percentage of correctly identified noise (solid line) and weighted ratio of correctly identified noise and signal (dashed line) when registering to the MNI template directly (orange) or via the DeCon template (blue). *** indicates significant p values at
Two major cerebellar fissures, the primary and intrabiventer fissures, were drawn on the native structural scan of each individual and transformed into DeCon, MNI, and SUIT space. To quantify fissure overlap, we calculated the average fissure distances between all possible pairs of participants. The primary and intrabiventer fissures showed a 30% higher overlap in DeCon space than in both MNI and SUIT space (Fig. 2A). The largest reduction of fissure distances was found in the vermal portion of the primary fissure (Table 2). The primary fissure in the vermal portion was ∼0.9 mm apart in the DeCon template, compared with 2–3 mm in MNI and SUIT space (Fig. 2D). Paired t tests demonstrated that the 70% reduction in fissure distances from MNI space and 60% reduction from SUIT space were significant for both the template generation sample (DeCon vs MNI:
Fissure distances
To ensure that no systematic differences were present in anatomical alignment due to pathological anatomy, we tested for differences in fissure overlap between patients and controls in DeCon space using independent sample t tests. We found no significant patient-control differences for the vermal primary fissure
Finally, we asked whether the improvement in anatomical alignment results in better localization of functional data. For this, we used the automated classifier FIX (Salimi-Khorshidi et al., 2014) to identify noise in the resting-state time series data (see Materials and Methods, FIX training). Among several features used for noise classification, FIX uses masks of the major draining veins to identify features of vein-driven noise (Salimi-Khorshidi et al., 2014). Because these masks are drawn in MNI space and resampled to functional space, FIX's ability to detect noise relies in part on an accurate mapping between MNI space and native functional space. We provided FIX with mappings between functional and MNI space estimated via the subject's anatomical image only or via the DeCon template (see Materials and Methods, Template generation) as an intermediate step between anatomical image and the MNI template. We then compared leave-one-out classification accuracy of FIX trained on hand-labeled signal and noise components across a range of classification thresholds in terms of correctly identified noise (true noise rate) and a weighted average between true noise rate and true signal rate (see Materials and Methods, FIX evaluation). Since FIX requires whole-brain registration, we only compared performance with the DeCon template, a whole-brain template. We did use the SUIT template as a comparison, since SUIT provides only a cerebellum template. Across classification thresholds, alignment via the DeCon template resulted in 46% more true noise components detected than alignment directly to MNI space (paired t test of true noise rate for DeCon vs MNI
Cerebellar degeneration impairs connectivity with cerebellar regions
To understand the effects of cerebellar degeneration on the cortico-cerebellar motor network, we first examined functional connectivity at baseline. We quantified connectivity in a region-of-interest (ROI) approach as the Pearson's correlation between the resting-state time courses of neocortical regions involved in motor control (primary motor cortex, M1; dorsal premotor cortex, PMd; posterior parietal cortex, PPC) and the contralateral cerebellar motor region (hand region M3) for both left and right (see Materials and Methods, Functional connectivity). We tested for significant effects of group (patient, control), regions (corticocortical, cortico-cerebellar, and cerebello-cerebellar), and a group × region interaction using a linear mixed effects model (see Materials and Methods, Statistical analysis). Patients showed significantly reduced connectivity at baseline (main effect of group:
Impaired connectivity in the visuomotor network of patients. A, Connectivity of the visuomotor network differs at baseline between patients (red) and patients (blue). B, Connectivity shown separately for corticocortical, cerebello-cerebellar, and cortico-cerebellar connectivity shows that connectivity reduction in patients is specific to connectivity with cerebellar regions. * indicates significant p values at
Training with visual feedback strengthens connectivity with bilateral premotor cortex in patients
Having established that cerebellar degeneration reduces connectivity in the cortico-cerebellar motor network, we then examined if prolonged training could reverse these impairments.
We first examined training effects on connectivity with the dorsal premotor cortex (PMd), an area which has been shown to increase in gray matter through training in patients with cerebellar degeneration (Burciu et al., 2013; Draganova et al., 2022). We modeled connectivity of PMd contralateral to the training hand with a cerebellar motor region ipsilateral to the training hand in a linear mixed effects model with fixed effects of group, session, vision (vision/no vision), and feedback (explicit feedback/no explicit feedback) and included all higher-order interaction terms, along with a random effect of subject. We observed a significant patient-control difference in connectivity (main effect group:
Connectivity change between dorsal premotor cortex (PMd) and cerebellar motor regions for training with and without vision. A, Left PMd connectivity with right cerebellar motor region. B, Right PMd connectivity with left cerebellar motor region. * indicates significant p values at
Previous investigations of training-related changes in PMd have reported gray matter increases ipsilateral to the training hand that were greater in patients than in controls (Draganova et al., 2022). We therefore assessed connectivity changes between right PMd and a left cerebellar motor region. We found that connectivity between right PMd and a left cerebellar motor region differed between groups (main effect group:
Finally, because the cerebellum also projects to ipsilateral neocortical regions (Sultan et al., 2012), we examined connectivity between the right dorsal premotor cortex (PMd) and the right cerebellar hand area. We observed a significant patient-control difference in connectivity (main effect of group:
To test the specificity of our findings to cortico-cerebellar connectivity, we performed a control analysis on interhemispheric connectivity calculated between left PMd and right PMd. As expected, we found no patient-control difference (main effect group:
Training with visual feedback increases cerebellar connectivity with contralateral posterior parietal cortex
In light of the key role the posterior parietal cortex plays in recalibrating visually guided reaching movements (Clower et al., 1996), we set out to test whether cerebellar connectivity with the posterior parietal cortex (PPC) increased when visual feedback was provided during training. We observed a significant difference in connectivity between patients and healthy control participants, with reduced connectivity in patients compared with controls (main effect group:
Connectivity change between posterior parietal cortex (PPC) and cerebellar motor regions when training with and without vision. A, Left PPC connectivity with cerebellar motor regions in the right hemisphere. B, Right PPC connectivity with cerebellar motor regions in the left hemisphere. * indicates significant p values at
Ipsilateral to the training hand, PPC connectivity with contralateral cerebellum showed a significant patient-control difference at baseline (main effect group:
Motor training increases connectivity with contralateral primary motor cortex
Motor training has been shown to change connectivity between left M1 and right cerebellar cortex (Jayaram et al., 2011; Schlerf et al., 2012). We tested whether practicing reaching movements, irrespective of training condition, changes functional connectivity between the right cerebellar and left M1 hand region. Indeed, we found that contralateral to the right hand, M1→cerebellar connectivity increased through training (main effect session:
Connectivity change between M1 and cerebellar motor regions. A, Left M1 connectivity with right cerebellar motor region. B, Right M1 connectivity with left cerebellar motor region. * indicates significant p values at
To test for the specificity of our findings to M1 contralateral to the training hand, we tested for connectivity changes from baseline of right M1 left cerebellar connectivity. We found no significant connectivity change from baseline overall, suggesting that the training-related increase was specific to connectivity between left M1 and right cerebellar hand region (main effect session:
Finally, we tested for a change in connectivity between left and a right cerebellar motor region through training or as a function of feedback. Our analysis showed a difference in connectivity between patients and controls at baseline (main effect group:
Explicit feedback affects connectivity in cognitive cortico-cerebellar network
Cognitive regions of the neocortex have been implicated in strategy forming and explicit learning during visuomotor reaching tasks (Anguera et al., 2010, 2011; Taylor et al., 2014). We therefore explored how explicit verbal feedback during training would affect connectivity between cognitive regions of the neocortex and the cerebellum. For this, we chose a region of interest from a novel functional atlas of the cerebellum, which was based on an aggregate of seven extensive functional datasets (Nettekoven et al., 2023). We chose region D2 from the cerebellar atlas, which is a core multiple demand region that activates during the rapid processing of instructive signals in motor tasks and is located in lobules VI and VII along the superior posterior and the ansoparamedian fissures. We examined connectivity of D2 with the right dorsolateral prefrontal cortex (dlPFC). Previous work has shown that excitatory noninvasive brain stimulation induces faster adaptation in a reaching task and greater corrections early in the training, as compared with no stimulation or stimulation of left dlPFC, suggesting that right dlPFC is specifically engaged during strategy-based learning in adaptive reaching (Song et al., 2020).
We observed a significant interaction between group, session, and feedback (group × session × feedback interaction:
Connectivity change between right dorsolateral prefrontal cortex (dlPFC) and right cognitive cerebellar region (cerebellar atlas region D2 in Nettekoven et al., 2024) training where explicit verbal feedback was given or not. Raw connectivity values are used for visualization purposes only and statistics are calculated on z-transformed values.
Finally, we wanted to test the specificity of any connectivity changes to cerebellar regions involved in the training. As an additional control analysis, we therefore examined connectivity of right dlPFC with a region that shows no involvement in hand movements (Nettekoven et al., 2024). We chose the region S1 from the cerebellar atlas, which has been shown to be functionally connected to right dlPFC and activates during tasks involving social-linguistic processing but does not respond to motor tasks or during rapid processing of instructive signals in a motor task (Nettekoven et al., 2024). We first tested whether right dlPFC and left cerebellar S1 were functionally correlated at baseline using a one-sample t test of baseline connectivity values. Indeed, the connectivity values were significantly different from zero (
Results were robust to outliers
Visual inspection of the behavioral data suggested potential outliers in the training-related improvement of the relative joint position error (Draganova et al., 2022, their Figure 3). To test whether these individuals may have driven changes in resting-state activity, we performed a Grubbs test for outlier detection on the difference in relative joint position error from pre to post training. The Grubbs test identified five outlier subjects whose functional connectivity data was removed before repeating the main analyses. Outlier removal did not change our findings, with the exception of right PMd→left cerebellar connectivity. Here, the three-way interaction between group, session, and vision was reduced to a trend (group × vision × session interaction:
Discussion
How cerebellar degeneration affects cortico-cerebellar connectivity and how therapeutic interventions can counter these effects are key clinical and neuroscientific questions that have yet to be addressed. In this study, we used a carefully validated patient-specific neuroimaging pipeline to demonstrate, for the first time, cortico-cerebellar connectivity alterations in cerebellar patients that normalize with motor training in a feedback-dependent manner. The main findings of the study are as follows: First, at baseline, patients showed reduced connectivity of a cerebellar motor region with contralateral regions of the neocortex and with the cerebellar region of the opposite hemisphere, while corticocortical connectivity was unimpaired. Second, visuomotor training led to increased connectivity between cerebellar motor and PMd contralaterally to the trained arm for all participants and ipsilaterally to the training arm for patients. Third, training with visual feedback increased connectivity with the PPC in all participants. These findings were highly specific to cortico-cerebellar connectivity. They were not driven by changes in neocortical connectivity, as left→right PMd connectivity and left→right PPC connectivity did not change overall.
To date, plastic changes after training interventions in cerebellar patients have only been demonstrated in regions not directly affected by degeneration (Burciu et al., 2013), including in a previous report of structural changes in this cohort (Draganova et al., 2022). These previous studies found gray matter increases in PMd but were unable to demonstrate changes in cerebellar gray matter. Increased temporal synchrony of resting-state fluctuations in distant brain regions are thought to reflect increased communication of these regions (Fries, 2015; Ivachtchenko et al., 2016). This connectivity increase could be driven by strengthening of connections between brain regions involved in training, which remain detectable at rest post-training. In line with this, connectivity increases between task-relevant regions have been demonstrated in the visual cortex, motor cortex (Sami et al., 2014), and cerebellar cortex (Albert et al., 2009; Nettekoven et al., 2020). Given that the cerebellum projects to PMd (Bostan et al., 2013), our findings indicate that visuomotor training can take advantage of residual connections in the cortico-cerebellar network and induce lasting changes in the communication between the training-relevant regions. Importantly, the unimpaired cortico-cortico connectivity in cerebellar patients at baseline indicates that cerebellar degeneration specifically affects communication between the cerebellum and neocortex. Further, the lack of significant change in cortico-cortico connectivity suggests that the observed training-related connectivity increases do not simply reflect a global increase in connectivity.
The execution of visually guided movements involves the PPC (Desmurget et al., 1999). In line with this, we found that visuomotor training increases functional connectivity between a cerebellar motor region and PPC contralateral to the trained hand. Strengthening of a network involving cerebellar regions and contralateral PPC has previously been reported in response to motor adaptation in healthy participants (Della-Maggiore and McIntosh, 2005). To our knowledge, our study is the first to demonstrate connectivity changes involving the PPC in cerebellar patients as well as healthy controls.
Previous research has found strengthening of connectivity between cerebellum and M1 contralateral to the training hand, when adaptive learning takes place, but not when no learning was induced (Jayaram et al., 2011). This M1→cerebellar connectivity change, assessed through paired-pulse transcranial magnetic stimulation measures of cerebellar-brain inhibition, appears to rely on the presence of larger movement errors and does not manifest when introducing systematic errors gradually or randomly (Schlerf et al., 2012). Here, we found M1→cerebellar resting-state connectivity increased regardless of task condition. These findings are in line with prior work, since all task conditions in this study tended to induce similar rates of learning, with differences between conditions appearing only in some target directions (Draganova et al., 2022). Moreover, the training conditions in this study were matched in all aspects of training protocol apart from visual and explicit verbal feedback, which have so far not been reported to induce differential changes in M1→cerebellar connectivity.
When exploring changes in a cortico-cerebellar network involved in higher-order cognitive aspects of motor training, we observed changes that differed between patients and controls and depended on the presence of explicit feedback. While these differences appear to be driven by a connectivity increase in the patient group when training without explicit feedback, post hoc tests did not survive multiple comparisons. Although these findings might point to a potential malleability of cognitive cortico-cerebellar connectivity through targeted interventions, they will need to be replicated in an independent sample.
In this study, we used a correlation-based approach to analyze functional connectivity, a method that has been widely used in the field for its simplicity and interpretability (Lurie et al., 2020). Correlation-based functional connectivity measures have shown sensitivity to individual differences in health and disease (Greicius, 2008; Fox and Greicius, 2010; Hernandez-Castillo et al., 2014). Functional connectivity has also been shown to change in response to behavior (Albert et al., 2009; Lewis et al., 2009; Sami et al., 2014; Nettekoven et al., 2020) and may have powerful utility as clinical biomarkers (Drysdale et al., 2017; Reggente et al., 2018; Etkin et al., 2019). While more complex methods probing “effective connectivity,” such as dynamic causal modeling (Friston, 2003) or Granger causality (Goebel et al., 2003), allow inferences about directed connectivity, standard fMRI data is ill-suited for several of these methods due to its low sampling rate (Friston, 2009; Seth et al., 2015). Given the relatively modest sampling rate of the resting-state data in this study, we therefore chose an approach that ensured that results were robust and accessible. This simplicity does not detract from the validity of our findings. On the contrary, it allows us to capture patterns of connectivity change without the risk of overfitting or introducing unnecessary complexity. Furthermore, the specificity of the connectivity changes we observed underscores the utility of this correlation-based method in answering our research questions.
We were able to reveal highly specific connectivity changes by optimizing our analysis pipeline for maximizing overlap of cerebellar structures and thereby preserving cerebellar signal. By creating a novel template for the analysis of cerebellar patient data, the DeCon template, we reduced spatial spread of anatomical landmarks in the cerebellum by 30–70% from existing templates, including the MNI template and the SUIT template, which had so far been the gold standard in cerebellar neuroimaging analysis (Diedrichsen, 2006). When training a noise classifier on cerebellar resting-state fMRI data, mapping to MNI standard space via the DeCon template in a multistage deformation tripled classification accuracy. To enable other researchers to take advantage of this methodological advancement, we are making the template freely available under https://github.com/carobellum/DegenerationControlTemplate. We believe that it will improve anatomical and functional analysis of patient data and prove useful to the research and clinical community.
Conclusion
Our results show that altered cortico-cerebellar connectivity can be ameliorated in cerebellar patients with training in a feedback-dependent manner. We used a newly developed and carefully validated brain template to reveal these subtle, but robust, connectivity changes. These results demonstrate a fundamental step forward in understanding residual function in cerebellar degeneration patients and developing training protocols that take this knowledge into account. Further, the methodological advances presented here constitute a valuable addition to the tools available for cerebellar patient studies.
Data Availability
The study-specific template is publicly available at https://github.com/carobellum/DegenerationControlTemplate (DOI: 10.5281/zenodo.13273700). The template validation data is publicly available at https://github.com/carobellum/DegenerationConnectivity/tree/main/data along with the extracted time series data supporting the connectivity findings of this study. Due to data privacy restrictions, the raw MRI data has not been openly released but can be made available upon request to the corresponding author. The code for generating and validating the template, analyzing the functional connectivity data, and for recreating the results and figures in this paper is publicly available as the GitHub repository https://github.com/carobellum/DegenerationConnectivity (DOI: 10.5281/zenodo.13621539).
Footnotes
We thank Jacob Levenstein for providing advice on template generation and Joern Diedrichsen for helpful discussions about template validation. We thank Beate Brol for hand-correcting cerebellar masks and manually drawing cerebellar fissures. This work was supported by a research grant of the German Research Foundation awarded to J.K. and D.T. (DFG TI 239/14-1) and a research grant of the Bernd Fink Foundation awarded to J.K. and D.T. Part of the work was completed during a research visit to the Department of Neurology, University Hospital Essen, University of Duisburg-Essen, supported by a Travel Grant from ThinkGlobal RUB awarded to C.N.
The authors declare no competing financial interests.
This paper contains supplemental material available at: https://doi.org/10.1523/JNEUROSCI.1823-24.2025
- Correspondence should be addressed to Caroline Nettekoven at cr.nettekoven{at}gmail.com.
