Progressive Changes Between Thalamic Nuclei and Cortical Networks Across Stimulus–Response Learning

ABSTRACT

The thalamus is connected to the cerebral cortex and subcortical regions, serving as a node within cognitive networks. It is a heterogeneous structure formed of functionally distinct nuclei with unique connectivity patterns. However, their contributions to cognitive functioning within networks is poorly understood. Recent animal research suggests that thalamic nuclei such as the mediodorsal nucleus play critical roles in goal‐directed behaviour. Our aim was to investigate how functional integration of thalamic nuclei within cortical and subcortical networks changes whilst transitioning from more controlled goal‐directed behaviour towards more automatic or habitual behaviour in humans. We analysed functional magnetic resonance imaging (fMRI) data from a stimulus–response learning study to investigate functional connectivity (FC) changes across learning between thalamic nuclei with cortical networks and subcortical structures in 52 healthy subjects. We also defined additional regions‐of‐interest (ROIs) individually in native space, segmenting the thalamus into 47 nuclei and segmenting 38 subregions within the basal ganglia and hippocampus. Additionally, we defined 12 cerebral cortex ROIs via maximum‐probability network templates. Associative S‐R learning‐related connectivity changes were examined via ROI‐to‐ROI functional network analysis. Our results showed that learning was associated with: (1) decreasing FC between the frontoparietal network and higher order thalamic nuclei; (2) increasing FC between the cingulo‐opercular network and pulvinar nuclei; (3) decreasing FC between the default mode network (DMN) and right mediodorsal nuclei; (4) increasing FC between the DMN and left mediodorsal nuclei; (5) changes in functional connectivity between thalamic nuclei and putamen subregions, and (6) increasing intrathalamic FC. Together, this suggests that several thalamic nuclei are involved in the learning‐related transition from controlled to more automatic behaviour.

We analysed functional magnetic resonance imaging (fMRI) data from a stimulus–response learning study to investigate functional connectivity changes across learning between thalamic nuclei with cortical networks and subcortical structures within healthy subjects. Our results show diverse functional connectivity changes between higher‐order thalamic nuclei and cortical networks that underlie learning.

1. Introduction

Learning novel associations between stimuli (S) and responses (R) to achieve desired outcomes is of importance for adaptive behaviour. Oftentimes, novel S‐R associations are shaped via feedback in the form of rewarding outcomes or absence of unpleasant outcomes. Across repeated learning and practice trials, the underlying processing characteristics are continuously changing: Whilst strategic exploration strategies (Mohr et al. 2018) and reinforcement learning mechanisms (Sutton and Barto 2018) are dominating initially, S‐R associations once established, can be further strengthened through simple repetition learning. These dynamics are often conceptualised as a transition from goal‐directed towards increasingly automatic behaviour, which ultimately leads to the development of habits (Ersche et al. 2017; Gardner 2015; Smith and Graybiel 2016; Yamada and Toda 2023). On the one hand, goal‐directed behaviour is characterised by actions performed to achieve a desired goal state in a given stimulus context. On the other hand, habitual or automatic behaviour is typically conceptualised as learned S‐R associations which persistently drive behaviour even in the absence of a desired goal state, or alternatively, even when the desired goal can no longer be achieved by the learned response (Gardner et al. 2012; Watson et al. 2022). Automaticity can therefore be conceptualised as learned associations between stimuli and actions which persist even in the absence of sustained interest or desired outcome (Gardner et al. 2012; Watson et al. 2022). This goal‐directed to automatic transition is known to be accompanied by changes in cerebral cortex networks and subcortical regions, such as the striatum (van der Straten et al. 2020). However, how the thalamus is involved in this transition is largely unknown. This is surprising, as the mammalian brain is a thalamocortical system, with every cerebral cortex region receiving afferents from, and sending efferents back to, the thalamus (Hwang et al. 2017). Additionally, the thalamus is extensively connected to the basal ganglia and other subcortical structures (Minagar et al. 2013) and is therefore in prime position to flexibly reconfigure large‐scale neural networks (Kawabata et al. 2021; Saalmann and Kastner 2015). As such, the thalamus might facilitate goal‐directed behaviour through coordination of cognitive networks (Hwang et al. 2017) and cortico‐striato‐thalamo‐cortical (CSTC) loops (Calzà et al. 2019) via its connections and cortical synchrony (Portoles et al. 2022) and might also contribute to automatic processes (Gremel and Lovinger 2017).

Despite its function as network hub (Kawabata et al. 2021), thalamic roles in goal‐directed behaviour within humans has been mostly ignored due to technological difficulties regarding small size and deep location (Calamante et al. 2013; Fama and Sullivan 2015). Furthermore, most human neuroscience research investigates the thalamus as a whole structure (Goldstone et al. 2018) despite comprising around 60 cytoarchitectonically and functionally distinct nuclei with unique connectivity patterns with cortical and subcortical structures (Fama and Sullivan 2015). There is a general consensus that higher‐order thalamic nuclei include mediodorsal (MD) and pulvinar subregions (Penner et al. 2018), alongside proposals to include anterior, lateral dorsal (Perry et al. 2021), ventral anterior, and ventral lateral nuclei (Sugiyama et al. 2018). Of special note here are MD nuclei, which are involved in executive functioning, decision‐making and cognitive tasks (Pergola et al. 2018). The MD nuclei are critical for maintaining task representations and action‐outcome contingencies (Fresno et al. 2019) and facilitate cognitive flexibility via initiating switching between cerebral cortex representations (Rikhye et al. 2018). This involves the frontoparietal network (FPN) (Li et al. 2022), but also default mode network (DMN) and salience networks (SN) (Niu et al. 2022). However, insight rests mostly on experiments in animals, not humans.

We investigated how integration of thalamic nuclei within networks might change across associative S‐R learning. Due to the prominent role of the MD nuclei within goal‐directed behaviour and executive functioning, we expected these nuclei in particular to contribute to the transition from goal‐directed towards automatic behaviour, potentially together with other higher thalamus regions. Furthermore, the MD nuclei are the largest thalamic nuclei in rodents, with even greater relative size within humans (Onishi et al. 2022; Pergola et al. 2018). In human participants, mean volumes of MD nuclei have been reported to be 981.3 mm³ on the left side and 1005.0 mm³ on the right side, having a combined size of 1986 mm³ in healthy human participants, and a standard deviation of 101.2 and 95.8 for left and right sides, respectively (Danos et al. 2003). These volumes, therefore, easily fall within the spatial resolution of standard human functional magnetic resonance imaging.

A recent fMRI study investigated transition from goal‐directed to automatic behaviour across learning of stimulus–response associations by measuring changes in local BOLD (blood‐oxygen‐dependent) responses and functional connectivity, and tested whether learning‐related changes predicted putative behavioural markers of habit strength (Zwosta et al. 2018). This revealed connectivity changes across the cerebral cortex and basal ganglia during the transition from more controlled to more automatic behaviour, alongside decreased thalamic responses and changes in functional connectivity between thalamus and putamen. However, functional contributions of distinct thalamic nuclei underlying these changes were not investigated not least due to a lack of analytical precision.

Here, we re‐analysed data obtained within this study (Zwosta et al. 2018), and focused upon functional connectivity changes between thalamic nuclei, cortical networks and subcortical regions that occurred progressively across associative S‐R learning trials. We employed the methodological means to optimise analytical precision to obtain as much anatomical detail from standard functional images as possible. The primary aim of this study was to identify relevant thalamus‐related functional connectivity changes during S‐R learning. We had three hypotheses: In light of the scarce prior work on thalamic learning‐related connectivity changes and the complex nature of learning‐related processing dynamics described above, our hypotheses were rather exploratory and general in nature. The primary aim of this study was, as a first step, to identify relevant thalamus‐related functional connectivity changes during S‐R learning in the first place. On the one hand, reflecting a decreasing gradient in higher‐level control processes such as feedback evaluation and goal‐directed control, we hypothesised that S‐R learning would be associated with functionally diverse connectivity alterations between higher‐order thalamic nuclei, higher‐order cortical control networks known to be important for goal‐directed control, such as the frontoparietal network (FPN), the cingulo‐opercular network (CON) and the so‐called default mode network (DMN), and subcortical regions such as the caudate and the nucleus accumbens. Here, based on previous work, we hypothesised MD thalamus subregions to be particularly involved in this transition due to their roles in goal‐directed behaviour and cognitive control (Parnaudeau et al. 2018; Parnaudeau et al. 2015). On the other hand, reflecting an increasing gradient regarding automatisation, we expected functional connectivity changes between thalamic subregions and more posterior putamen regions known to be relevant for habitual behaviour (Kim and Hikosaka 2015; Wu et al. 2010).

2. Materials and Methods

2.1. Subjects

The sample originally used by Zwosta et al. (2018) consisted of a total of 53 subjects (29 female, 24 male) with a mean age of 23.5 years and a range of 19–32 years. We excluded one subject due to missing data in the nucleus accumbens. All subjects were right‐handed and had normal or corrected‐to‐normal vision.

The experimental protocol was approved by the Ethics Committee of the Technische Universität Dresden [EK306082011]. All subjects gave written informed consent prior to the experiment and were compensated 8€ per hour, in addition to any money earned throughout the experiment.

2.2. Experimental Procedure

The experimental paradigm was formed of three consecutive phases, of which the second phase is of primary relevance for the present paper (Figure 1). In Phase 1, goal‐directed behaviour was established, followed by Phase 2, which was designed such that participants transitioned from initial goal‐directed towards more automatic behaviour. In Phase 3, goal‐directed behaviour established in Phase 1 was put into competition with more automatic behaviour established in Phase 2.

FIGURE 1.

The experimental paradigm within Zwosta et al. (2018). The experiment consisted of three consecutive phases. In the present paper, we focused on fMRI data acquired during Phase 2. (A) Phase 1 established goal‐directed behaviour directed towards different colour outcomes (outside the MRI machine). The panel shows examples of the instructed stimulus–response–outcome associations with two exemplary trials from Phase A. Subjects were instructed to learn the association between object categories (artificial, natural) and the respective colour and gave responses via button press. In a subsequent test, they were asked to either maintain or change the colour association. (B) During Phase 2, novel stimulus–response mappings were learned for a subset of the stimuli used in Phase 1. Learning occurred via approach or avoidance and was practiced inside the MRI machine. The panel shows exemplary trials from Phase 2, which consisted of approach and avoid trials. (C) Phase 3 placed goal‐directed responses from Phase 1 and habitual responses from Phase 2 in competition with each other in order to test habit strength. The panel shows examples for compatible, incompatible and free‐choice trials within Phase C.

Phase 1 took place outside of the MRI machine, and goal‐directed behaviour was established through the learning of 10 hierarchical stimulus–response‐outcome (S‐R‐O) associations involving 10 distinct stimuli, 2 distinct responses and 2 distinct colour outcomes. Phase 2 then took place within the MRI machine and included eight of the visual stimuli from Phase 1, but now required subjects to learn novel stimulus–response associations by trial and error and with reinforcement. Here, the crucial assumption was that behaviour would be goal‐directed initially (i.e., given a particular stimulus, select the action which will yield a rewarding outcome or enable the successful avoidance of an aversive outcome) and that this behaviour would gradually transition towards more automatic behaviour (i.e., the action would be generated ‘reflexively’ in response to the stimulus with increasingly less attention paid to the outcome). Finally, Phase 3 placed both goal‐directed behaviour established in Phase 1 and habits established in Phase 2 into competition with each other within incompatible trials (which had required different responses in Phase 1 than they did in Phase 2) in order to quantify the habit strength acquired in Phase 2.

The present paper primarily focused on the second phase with the aim to investigate thalamic contributions to the transition from goal‐directed behaviour to more automatic behaviour. Here, across associative S‐R learning and practice, subjects were supposed to gradually transition from more controlled or goal‐directed behaviour towards ever more automatic behaviour through learning trials and successful repetitions.

Phase 1: Establishment of goal‐directed behaviour, outside of the MRI machine.

This phase aimed at establishing a representation of hierarchical (here: stimulus‐dependent) response‐outcome (R‐O) associations as the foundation of goal‐directed behaviour (de Wit et al. 2009; Griffiths et al. 2014), whilst at the same time ensuring that no stable stimulus–response (S‐R) habits could be formed (in contrast to phase 2 where this was the case).

The task utilised 10 visual stimuli which were assigned into two categories: ‘artificial’ and ‘natural’. These consisted of black and white, vertically symmetrical icons of different objects. Artificial stimuli included a ball, car, computer mouse, cupboard and scissors, whereas natural stimuli included a cow, lungs, mushroom, snowflake and a tree. When each image was presented on a screen, subjects were requested to press a corresponding key on a keyboard, that is, by using the left index finger to press the key ‘D’ or using the right index finger to press the key ‘K’. Responding to a visual stimulus within one category with the right key would lead to a blue outcome colour, whereas responding to the same category of visual stimulus with the left key would lead to an orange outcome colour. This R‐O association was inverted for the other visual stimulus category, so that pressing the right key would lead to an orange outcome colour, whereas pressing the left key would lead to a blue outcome colour. This inversion was designed to exclude habitualisation processes, as for each stimulus, each of the two possible responses was equally often correct. Hence, the correct response was determined by the hierarchical relationship between stimulus category and intended goal (i.e., the anticipated colour outcome). Neither stimulus category (artificial or natural) alone nor intended goal alone was sufficient to determine the correct response on a given trial.

Phase 1 began with subjects looking at an instruction screen which displayed all visual stimuli belonging to each group and the assigned R‐O relationships. Next, the task was explained using text and illustrations. During this instruction phase, the experimenter stayed with each subject and during an additional 20 practice trials in order to answer any questions and ensure that the subject understood the instructions given.

Subsequently, subjects entered the testing phase. Here, they underwent a total of 240 trials, as outlined in Figure 1A. Each trial began with a cue on the screen, which was the German word for ‘change’ or ‘maintain’ surrounded by a square frame with either a blue or an orange colour. This colour would be the outcome colour achieved within the previous trial (or, in the case of the first trial, would be a random colour). This cue would then be followed 1200 ms later by a visual stimulus from either the natural or artificial image categories. Here, the subject would utilise their knowledge of the hierarchical S‐R‐O associations learnt during the training phase in order to select the appropriate key in order to either ‘change’ or ‘maintain’ the colour as given by the coloured frame. The subjects had a response window of 2000 ms. Responses would cause the background of the stimulus to turn to the respective outcome colour for 1000 ms. Incorrect responses would generate the word ‘Error’ displayed below the square and the incorrect outcome colour would be shown. The subject would then have to immediately repeat the trial until they gave the correct response.

Phase 2: Stimulus–response learning, within the MRI machine. In this phase, subjects were required to learn novel stimulus–response associations based on feedback in the form of either reward or punishment. Extensive practice of each S‐R association was supposed to establish habitual approach or avoidance behaviour, respectively. Four artificial and four natural stimuli were reused from Phase 1 (instead of the original 10 visual stimuli). At the beginning of Phase 2, subjects were informed that the categories of the stimuli were no longer relevant and that during this phase they were required to discover and learn the correct key responses by trial and error. For both of the stimulus categories, each of the four stimuli belonging to each category was associated with one of the four combinations of correct responses (left or right key press) and outcome type (reward or punishment avoidance). Subjects obtained points by pressing the correct key response to four of the stimuli, whilst for the other four stimuli they would avoid losing points by pressing the correct key. Subjects were also informed that a faster response speed would lead to gaining additional money to be obtained at the end of the experiment. Subjects received one cent per 10 points and one cent per millisecond of their average response time being below 550 ms.

In Phase 2, each trial began with a fixation cross on the screen which was framed by a black‐and‐white lined square. This fixation cross was displayed for a variable duration, which ranged from 500 to 6500 ms (Figure 1B). A visual stimulus from one of the two image categories would then be presented, for a maximum of 1500 ms. The point outcome would be displayed immediately upon the response given by the subject. Rewards were 10+ points which were printed on the screen in green colour, whereas punishments consisted of −10 points which were printed in red. Outcomes of zero points were printed in black. Failure to respond during the response window would generate the unfavourable outcome, whereby in approach trials they would gain zero points (‘0’) and in avoidance trials they would lose 10 points (‘−10’). This outcome would remain on the screen for 1000 ms. Subjects were able to view their current number of points during the whole trial as this information was displayed at the top of the screen.

Phase 2 consisted of seven task blocks with 112 trials each (14 per stimulus). This formed a total of 784 trials, which was 98 per stimulus. The total number of points collected during each block would be displayed at the end of each block, and the mean response time for correct responses was also displayed at the end of each block.

Phase 3: Goal‐directed versus habitual behaviour: testing habit strength within the MRI scanner. Both goal‐directed behaviour from Phase 1 and habitual behaviour from Phase 2 were put into competition in order to assess inter‐individual habit strength. Subjects were informed at the start of this phase that the rules of the prior Phase 2 no longer applied, in that they could no longer gain or lose points. In this way, the contingency between stimulus, response and (monetary) outcome was removed, whereas the contingencies within Phase 1 were re‐instructed. All 10 stimuli from Phase 1 were used in Phase 3 (Figure 1C). A fixation cross was shown on the screen, which was followed by a coloured frame. This time, the frame was either orange or blue as within Phase A but could also be purple to indicate a free choice trial, whereby subjects could select either the left ‘D’ or right ‘K’ key. One of the 10 visual stimuli then appeared on the screen. If the frame was blue or orange, then subjects were required to press the response that would lead to the outcome colour for the displayed stimulus in accordance with the R‐O contingencies learnt within Phase 1 during goal‐directed trials. A total of 6% of trials were catch trials, in order to prevent subjects from selecting a response prior to the presentation of the stimulus. Subjects were also told not to decide how to respond before they were presented with the stimulus.

For all trials, the fixation cross was displayed for a variable ITI from 500 to 6500 ms and followed by a cue frame for 700 ms. The stimulus was displayed for up to 2000 ms or until the subject responded. After a response was provided, the outcome colour would be displayed for 1000 ms.

Phase 3 consisted of 384 trials in total. There was a total of eight different trial types of interest, each appearing in a random order. All trials required goal‐directed responding as established in Phase 1. This included ‘compatible’ trials for which the habitual response from Phase 2 was identical to the required goal‐directed response because it had been previously rewarded (approach compatible, 48 trials) or not punished (avoid compatible, 48 trials). There were also ‘incompatible’ trials for which the habitual response from Phase 2 did not match the required goal‐directed response and thereby caused a conflict between the goal‐directed response and the trained habitual response (approach incompatible or avoid incompatible, both 48 trials). Lastly, there were trials for which no habitual response existed as the stimulus had not been used within Phase 2 (no training, goal only, 48 trials). Additionally, there were free‐choice trials. For one type of free choice trials, the stimulus for which a habitual response was trained during Phase 2 was displayed (approach free‐choice or avoid free‐choice, each 48 trials). For another type of free‐choice trials, the stimulus had not been used in phase 2 (no training free‐choice, 24 trials). The difference between response times in incompatible versus compatible trials was used as a potential index of habit strength during the connectivity analysis. We call this difference between incompatible versus compatible trials the compatibility effect. The data from no‐habit trials and the free choice trials were not used any further.

2.3. MRI Data Acquisition

MRI data were acquired on a Siemens 3T whole body Trio System (Erlangen, Germany) with a 32‐channel head coil. Earplugs were utilised in order to minimise MRI scanner noise. After the experimental session, structural images were acquired using a T1‐weighted sequence (TR = 1900 ms, TE = 2.26 ms, TI = 900 ms, flip = 9°) with a resolution of 1 mm × 1 mm × 1 mm. Functional images were acquired using a gradient echo planar sequence (TR = 2000 ms, TE = 30 ms, flip angle = 80°). Each volume contained 32 axial slices which were measured in ascending order. The voxel size was 4 mm × 4 mm × 4 mm (slice gap: 20%). The experiment was controlled by E‐Prime 2.0.

2.4. Data Analysis

2.4.1. Preprocessing

We performed preprocessing using fMRIPrep 22.0.2 (Esteban et al. 2019; Esteban et al. 2018; RRID:SCR_016216, n.d.) which is based on Nipype 1.8.5 (Gorgolewski et al. 2011; Gorgolewski et al. 2018; RRID:SCR_002502, n.d.).

2.4.2. Anatomical Data

The T1‐weighted (T1w) image was corrected for intensity non‐uniformity (INU) with N4BiasFieldCorrection (Tustison et al. 2010) distributed with ANTs (Avants et al. 2008); ARRID:SCR_004757), and used as T1w‐reference throughout the workflow. The T1w‐reference was then skull‐stripped with a Nipype implementation of the antsBrainExtraction.sh workflow (from ANTs), using OASIS30ANTs as target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM), and grey matter (GM) was performed on the brain‐extracted T1w using fast (FSL 6.0.5.1:57b01774, RRID:SCR_002823, (Zhang et al. 2001). Brain surfaces were reconstructed using recon‐all (FreeSurfer 7.2.0, RRID:SCR_001847, (Dale et al. 1999)) and the brain mask estimated previously was refined with a custom variation of the method to reconcile ANTs‐derived and FreeSurfer‐derived segmentations of cortical grey matter (RRID:SCR_002438, (Klein et al. 2017)).

Even though all our analyses were performed in native space, individual brains were additionally normalised to standard MNI space in order to be able to back‐transform cortical network ROIs defined in MNI space back into native space (see section Cerebral Cortex Networks). To this end, volume‐based spatial normalisation to standard space (MNI152NLin2009cAsym) was performed through nonlinear registration with antsRegistration (ANTs 2.3.3), using brain‐extracted versions of both T1w reference and the T1w template. The following template was selected for spatial normalisation: ICBM 152 Nonlinear Asymmetrical template version 2009c (Fonov et al. 2009; RRID:SCR_008796; TemplateFlow ID: MNI152NLin2009cAsym).

2.4.3. Functional MRI Data

The MRI data from experiment Phase 2 of each subject were preprocessed as follows. First, a reference volume and its skull‐stripped version were generated using a custom methodology of fMRIPrep. Head‐motion parameters with respect to the reference (transformation matrices and six corresponding rotation and translation parameters) are estimated before any spatiotemporal filtering using mcflirt (FSL 6.0.5.1:57b01774, (Jenkinson et al. 2002). MR images were slice‐time corrected to 0.972 s (0.5 of slice acquisition range 0 s–1.95 s) using 3dTshift from AFNI (Cox and Hyde 1997); RRID:SCR_005927).

The BOLD time‐series (including slice‐timing correction) was resampled onto its original, native space by applying the transforms to correct for head motion. These resampled BOLD time‐series will be referred to as preprocessed BOLD in original space, or just preprocessed BOLD.

Several confounding time‐series were calculated based on the preprocessed BOLD: framewise displacement (FD), DVARS and three region‐wise global signals. FD was computed using two formulations following Power (absolute sum of relative motions) (Jenkinson et al. 2002; Power et al. 2014) (relative root mean square displacement between affines, Jenkinson et al. (2002)). FD and DVARS are calculated for each functional run, both using their implementations in Nipype (following the definitions by Power et al. 2014). The three global signals are extracted within the CSF, the WM, and the whole‐brain masks. Additionally, a set of physiological regressors was extracted to allow for component‐based noise correction (CompCor, Behzadi et al. 2007). Principal components are estimated after high‐pass filtering the preprocessed BOLD time‐series (using a discrete cosine filter with 128 s cut‐off) for the two CompCor variants: temporal (tCompCor) and anatomical (aCompCor). tCompCor components are then calculated from the top 2% variable voxels within the brain mask. For aCompCor, three probabilistic masks (CSF, WM and combined CSF + WM) are generated in anatomical space. The implementation differs from that of Behzadi et al. in that instead of eroding the masks by 2 pixels on BOLD space, a mask of pixels that likely contain a volume fraction of GM is subtracted from the aCompCor masks. This mask is obtained by dilating a GM mask extracted from FreeSurfer's aseg segmentation, and it ensures components are not extracted from voxels containing a minimal fraction of GM. Finally, these masks are resampled into BOLD space and binarised by thresholding at 0.99 (as in the original implementation).

Components are also calculated separately within the WM and CSF masks. For each CompCor decomposition, the k components with the largest singular values are retained, such that the retained components' time series are sufficient to explain 50% of the variance across the nuisance mask (CSF, WM, combined or temporal). The remaining components are dropped from consideration. The head‐motion estimates calculated in the correction step were also placed within the corresponding confounds file. The confound time series derived from head motion estimates and global signals were expanded with the inclusion of temporal derivatives and quadratic terms for each (Satterthwaite et al. 2013). Frames that exceeded a threshold of 0.5 mm FD or 1.5 standardised DVARS were annotated as motion outliers. Additional nuisance timeseries are calculated by means of principal components analysis of the signal found within a thin band (crown) of voxels around the edge of the brain, as proposed by (Patriat et al. 2017). The BOLD reference was then co‐registered to the T1w reference using bbregister (FreeSurfer), which implements boundary‐based registration (Greve and Fischl 2009). Co‐registration was configured with six degrees of freedom.

The BOLD time‐series were resampled onto Free Surfer's ‘fsnative’ space. All resamplings were performed with a single interpolation step by composing all the pertinent transformations (i.e., head‐motion transform matrices and co‐registrations to anatomical and output spaces). Gridded (volumetric) resamplings were performed using antsApplyTransforms (ANTs), configured with Lanczos interpolation to minimise the smoothing effects of other kernels (Lanczos 1964). Non‐gridded (surface) resamplings were performed using mri_vol2surf (FreeSurfer).

Many internal operations of fMRIPrep use Nilearn 0.9.1 (Abraham et al. 2014) (RRID:SCR_001362), mostly within the functional processing workflow (for more details of the pipeline, see the section corresponding to workflows in fMRIPrep's documentation).

2.4.4. Thalamus Segmentation

We used FreeSurfer and the thalamic probabilistic segmentation algorithm by Iglesias et al. (2018) which is incorporated into the FreeSurfer software. Utilising a probabilistic thalamus atlas based upon histological data, this algorithm uses Bayesian inference to segment both the left and right thalamic nuclei of individual subjects into an overall total of 47 subregions (see Figure 2 for the results of an exemplary subject). This segmentation method was chosen due to it conformity to the traditional segmentation of the dorsal thalamus into ~50 nuclei (Roy et al. 2022). In a first step, we used SynthSeg (Billot et al. 2023) in order to obtain optimised whole‐thalamus segmentation. In a second step, the thalamus was segmented into 23 subregions for the left thalamus and 24 subregions for the right thalamus (with increased stiffness of the mesh from 0.05 to 0.25), as the paracentral nucleus on the left side was not available. Here, a subset of 16 subregions per hemisphere was included in our analysis: the anteroventral (AV), central medial (CeM), centromedian (CM), lateral geniculate nucleus (LGN), lateral posterior (LP), medial geniculate nucleus (MGN), mediodorsal lateral parvocellular (MDl), mediodorsal medial magnocellular (MDm), pulvinar anterior (PuA), pulvinar medial (PuM), pulvinar lateral (PuL), pulvinar inferior (PuI), ventral anterior (VA), ventral lateral anterior (VLa), ventral lateral posterior (VLp) nuclei and ventral posterolateral (VPL) nuclei. However, 7 subregions on the left side and 8 subregions on the right side could not be identified in all our subjects due to their extremely small volume in the probabilistic atlas. These included the bilateral central lateral (CL), laterodorsal (LD), limitans (suprageniculate) (L‐SG), medial ventral (reuniens) (MV‐re), parafascicular (Pf), ventral anterior magnocellular (VAmc), ventromedial (VM) and the right paracentral (Pc) nuclei. As a result, these subregions were excluded from our analysis in all subjects.

FIGURE 2.

Segmentation of thalamic nuclei using an exemplary subject and a probabilistic atlas by Iglesias et al. (2018), displaying 13 thalamic subregions. Subregions include the following: Anteroventral (AV), centromedian (CM), lateral dorsal (LD), lateral posterior (LP); mediodorsal lateral (MDl); mediodorsal medial (MDm), nucleus reuniens (Mv(Re)), pulvinar anterior (PuA), pulvinar lateral (PuL), pulvinar medial (PuM); ventral anterior (VA), ventral lateral anterior (VLa); ventral lateral posterior (VLp); and ventral posterolateral (VPL).

2.4.5. Basal Ganglia and Hippocampus Segmentation

In addition to the segmentation of thalamic nuclei, we were also interested in including additional non‐thalamic regions of interest that are implicated in learning processes. These regions principally included the putamen, hippocampus and amygdala, for which their own subregions were of interest to us due to their differing functional specialisations (Dall'Oglio et al. 2015; Roman et al. 2023; Xie et al. 2025; Zarei et al. 2013). The same also applied regarding our use of other regions such as the caudate nucleus and globus pallidus (Mattfeld and Stark 2015; Rudebeck et al. 2017; Seger and Cincotta 2005; Sheth et al. 2011). Therefore, for the analysis of changes in thalamic connectivity with subregions of the hippocampus and components of the basal ganglia, we utilised an atlas created by Tian et al. (2020) which segregated 38 bilateral regions (Figure 3). These ROIs were transformed from MNI space into subject‐specific native space via the antsApplyTransforms function based on the MNI152NLin2009cAsym_to‐T1w inverse transformation parameters created during fMRIPrep normalisation.

FIGURE 3.

Atlas of nuclei delineating hippocampus and basal ganglia subregions. Figure adapted from Tian et al. (2020). The thalamic subregions were not utilised in our study, as thalamic segmentation was performed using the atlas from Iglesias et al. (2018) due to its provision of subject‐specific segmentation and a greater number of thalamic nuclei. Hippocampus head medial subdivisions (HIP‐head‐m1 and HIP‐head‐m2), hippocampus head lateral subdivision (HIP‐head‐l), hippocampus body (HIP‐body), hippocampus tail (HIP‐tail), putamen ventroanterior (PUT‐VA), putamen dorsoanterior (PUT‐DA), putamen ventroposterior (PUT‐VP), putamen dorsoposterior (PUT‐DP), caudate ventroanterior (CAU‐VA), caudate dorsoanterior (CAU‐DA), caudate medial (mCAU), caudate posterior (pCAU), lateral amygdala (lAMY), medial amygdala (mAMY), nucleus accumbens shell (NAc‐shell), nucleus accumbens core (NAc‐core), globus pallidus posterior (pGP) and globus pallidus anterior (GPa).

2.4.6. Cerebral Cortex Network

To analyse changes in thalamic connectivity with the cerebral cortex, we utilised 12 functional network probability maps, which altogether were comprised of 153 spherical ROIs. These network ROIs were obtained within a study by Dworetsky et al. (2021) which created network templates based upon a winner‐takes‐all procedure from a total of six different datasets, available for download at https://github.com/GrattonLab/Dworetsky_etal_ConsensusNetworks. We preferred this network definition over alternative options due to its superior reproducibility across individuals. Given the relatively exploratory nature of our study, we included all 12 functional networks to cover the whole diversity of higher‐level networks (e.g., FPN) and lower‐level networks (e.g., SMN). Here, Dworetsky et al. (2021) used an average network assignment probability of ≥ 75% to assign the 153 spherical ROIs into the available networks. These networks included those known for their roles within higher‐order cognitive processes (Dworetsky et al. 2021; Hu et al. 2023; Zhou et al. 2018), such as the fronto‐parietal network (FPN), default mode network (DMN), salience network (SN), dorsal attention network (DAN), cingulo‐opercular network (CON), parieto‐medial network (PMN), parieto‐occipital network (PON), and language network (which corresponds to the ventral attention network in other work by Dworetsky et al. (2021)) (Figure 4). To be consistent with the original paper by Dworetsky et al. (2021), we have used the label of ‘language network’ for this ROI.

FIGURE 4.

A probabilistic representation of eight association networks from Dworetsky et al. (2021). Colours represent the respective networks; Auditory network, AN (orange), cingulo‐opercular network, CON (purple); default mode network, DMN (green), fronto‐parietal network, FPN (middle blue), parieto‐medial network, PMN (yellow), somatomotor dorsal network, SMd (red), salience network, SN (dark blue), and visual network, VN (light blue). These network voxels represent brain regions with highest confidence in network assignment across subjects in the Dworetsky study.

In addition to networks more directly involved in higher order processes, visual and somatomotor dorsal (SMd) and somatomotor lateral (SMl) networks were also provided and included in our re‐analysis, alongside an additional two networks, including the Temporal Pole (TPole) network and Medial Temporal Lobe (MTL) network. However, the TPole network did not reach the network assignment probability of ≥ 75% within the Dworetsky et al. (2021) study, and in many subjects within our own study, no voxels were found within the Medial Temporal Lobe (MTL) network. Therefore, neither of these two additional networks was included in our analysis.

The cortical network ROIs were transformed from MNI space into subject‐specific native space via the antsApplyTransforms function based on the MNI152NLin2009cAsym to‐T1w inverse transformation parameters created during fMRIPrep normalization.

2.4.7. Functional Data Analysis

2.4.7.1. Denoising

After preprocessing, we first performed denoising of the functional data in native space at the single subject level. To this end, we computed a General Linear Model (GLM) for each subject using SPM12 and MATLAB 2021a. The GLM included ‘nuisance’ regressors based on a subset of the confounding variables previously extracted by fMRIPrep. The residual timeseries were saved to disk for further processing. Consistent with the fmridenoise pipeline 24HMPaCompCorSpikeReg, we included the following confounding variables. First, we included the three translational and three rotational motion parameters and their quadratic terms as well as their derivatives and the quadratic terms of the derivatives. Second, we included the first five principal components of the CompCor timeseries within CSF (‘c_comp_cor’) and the first five principal components of the CompCor timeseries within white matter (‘w_comp_cor’). Third, we included the DVARS timeseries and the framewise displacement (‘FD’) timeseries.

2.4.7.2. Modelling Task‐Related Activity

In preparation for the functional connectivity analysis via generalised psycho‐physiological interaction (gPPI) analysis, we computed for each subject a new GLM on the previously generated denoised residual timeseries to estimate task‐related activity using SPM12 and additionally defined a 128 s high‐pass filter. We included model regressors for correct and incorrect approach trials as well as correct and incorrect avoidance trials. In order to capture associative S‐R learning‐related changes across correctly implemented trials, we additionally included parametric regressors that varied according to the incremental number of correctly implemented approach and avoidance trials per individual stimulus. Finally, after model estimation, we computed an ‘omnibus’ F‐contrast spanning all task regressors to be used by the subsequent gPPI for extracting seed region timecourses after regressing task‐related activity.

2.4.7.3. Generalised PPI Analysis (Whole Brain Parametric Connectivity Analysis)

Based on the GLM just described, we set up the gPPI analysis (McLaren et al. 2012) again using the GLM framework within SPM12 for each subject. To model mean task‐related activity, the gPPI again included all the task‐related regressors defined before. Most importantly, to capture task‐related functional connectivity between the seed region BOLD timecourse and the BOLD timecourses in each voxel of the brain, the gPPI model additionally included PPI regressors obtained by convolving the original task regressors with the eigenvariate timecourse determined across all voxels within the seed region. Hence, there were in total 8 additional PPI regressors for correct and incorrect approach trials as well as correct and incorrect avoidance trials, plus the respective parametric change regressors. Finally, the seed region eigenvariate timecourse was added as a model regressor to model task‐unrelated signal correlation between the seed region and target voxel timecourses. The effect of primary interest was the parametric connectivity change across correct learning trials (approach and avoidance alike). We therefore computed for each subject and each seed region a whole‐brain contrast for the mean learning‐related parametric change over approach and avoidance conditions. Here, the fMRI analysis was therefore based on linear regression across all correct trials.

2.4.7.4. Functional Network Analysis (ROI‐To‐ROI Connectivity)

For each subject, we took the parametric gPPI contrast images for each seed ROI and extracted for each remaining ROI (the ‘target ROI’) the mean contrast value across voxels within that target ROI. Overall, this resulted in an 81 × 81 ROI‐to‐ROI connectivity matrix for each subject (i.e., 6480 pairwise connectivity changes, excluding connectivity with own seed region) which was then used as inputs for the group‐level functional network analysis using the CONN toolbox (Nieto‐Castanon 2020). Note that due to the asymmetric nature of the gPPI, connectivity between ROI 1 as the seed region and ROI 2 as the target region was not identical to connectivity between ROI 2 as the seed region and ROI 1 as the target region. Hence, the ROI‐to‐ROI connectivity matrices were also asymmetric.

We used the CONN toolbox (Release 20.a) for the group‐level statistical analysis of these ROI‐to‐ROI connectivity using the default Functional Network Connectivity (FNC) option. Thereby, we were able to determine which of the 6480 pairwise connectivity changes were statistically significant. FNC first organises the ROIs into networks based upon functional connectivity similarity metrics between all ROI‐to‐ROI pairings using complete‐linkage clustering (Sorensen 1948). Once these networks have been defined, FNC analyses all connections between all ROI pairs, for both within‐ and between‐network connectivity sets (Jafri et al. 2008). FNC performs a multivariate parametric General Linear Model analysis for all connections, producing an F‐statistic for each pair of networks and an associated uncorrected cluster‐level p‐value. This is defined as the likelihood under the null hypothesis of a randomly‐selected pair of networks showing equal or greater effects than those observed within the current pair of networks. Lastly, a FDR‐corrected cluster‐level p‐value (Benjamini and Hochberg 1995) is defined as the expected proportion of false discoveries amongst all network pairs with similar or larger effects across the entire set of pairs (Nieto‐Castanon 2020). Here, a cluster threshold of p < 0.05 p‐FDR corrected was utilised, alongside a connection threshold of p < 0.05 p‐uncorrected. This method was chosen due to it being a conservative method, which was valuable due to our exploratory approach and desire to avoid false positives, and is the basis for our subsequent Focused Analysis, as described in detail in the subsection below.

Lastly, we performed an analysis whereby we investigated potential correlations between the compatibility effect within Phase 3 and functional connectivity changes across all regions of interest. To this end, we used the same CONN toolbox setup as before but this time added the subject‐specific compatibility effects as a covariate within the CONN toolbox multivariate GLM analysis and then tested the covariate for statistical significance regarding all connections.

2.4.8. Focused Analysis

Our primary analysis did not reveal significant connectivity changes involving the putamen. However, from a theoretical perspective, based on previous literature (hypothesis 3, see above) we had an a priori hypothesis that increasing automatisation involves the putamen. In addition, the correction method that was utilised within the primary analysis used a conservative correction method (Nieto‐Castañón 2020), and this method becomes increasingly conservative with larger numbers of included regions of interest and therefore has an elevated risk of generating false negatives. We therefore performed a ‘Focused Analysis’ with more lenient correction requirements, this time selectively focusing on changes in functional connectivity between thalamic nuclei and putamen subregions (thereby reducing the number of connections to be corrected for). Moreover, we utilised the ROI‐level p‐FDR corrected (ROI mass/intensity) option within the CONN toolbox. In contrast to the FNC approach where correction is performed on the level of identified function networks, the Focused Analysis applied correction on the level of ROIs. Here, a p‐uncorrected connection threshold of 0.05 was combined with a cluster‐level p‐FDR corrected threshold of 0.05. Compared to the primary analysis, this approach traded increased statistical power towards putamen‐related connectivity changes against increased risk of false positives. This enabled a greater chance to detect subtler connectivity changes which may have been missed with the stricter correction approach utilised for the primary analysis.

3. Results

3.1. Behavioural Results

For completeness, we include the original behavioural results as reported in Zwosta et al. (2018) for both Phase 2 and Phase 3.

In Phase 2, within‐subject ANOVAs were computed for both error rates and response times across seven learning blocks and for two motivation types, namely approach versus avoid categories (Figure 5). Overall, approach and avoidance learning conditions showed similar performance improvement in error rates and response times across the associative S‐R learning trials. Error rates and response times significantly decreased across blocks as associative S‐R learning progressed, with F(6,312) = 127.84, p < 0.001, η² = 0.71 and F(6,312) = 122.35, p < 0.001, η² = 0.70, respectively. Subjects showed generally faster responses in the approach condition than in the avoidance condition (main effect motivation type F(1,52) = 58.93, p < 0.001, η² = 0.53). A significant interaction between motivation type and learning block indicated a steeper decline in response times in the avoidance learning condition than in the approach learning condition (F(6,312) = 8.14, p < 0.001, η² = 0.14).

FIGURE 5.

Behavioural results for Phase 2. Error rates (A) and response times (B) across associative S‐R learning blocks for approach and avoidance trials. Taken from (Zwosta et al. 2018). Overall, approach and avoidance learning conditions showed similar performance improvements in error rates and response times.

In Phase 3, ANOVAs were computed separately for response times and error rates, including the within‐subjects factors motivation type (approach or avoid) and compatibility (compatible versus incompatible trials). Regarding error rate, there was a significant main compatibility effect in that error rates were higher for incompatible trials compared with compatible trials (F(1,52) = 6.58, p = 0.013, η² = 0.11). Motivation type did not significantly influence error rates (F(1,52) = 0.80, p = 0.375), and there was no reliable interaction between motivation type and compatibility (F(1,52) = 3.20, p = 0.079). Regarding response times, there was also a significant main effect of compatibility in that responses were slower for incompatible versus compatible trials (F(1,52) = 17.96, p < 0.001, η² = 0.26). There was no main effect of motivation type (F(1,52) = 2.66, p = 0.109), nor a significant interaction effect between motivation type and compatibility (F(1,52) = 1.29, p = 0.262).

3.2. Learning Related Functional Connectivity Changes

Across associative S‐R learning trials, widespread significant changes in functional connectivity were found between numerous thalamic nuclei and cortical networks as well as with subcortical regions (Figure 6A,B) as learning progressed through the trials. For detailed statistical information regarding individual ROI‐to‐ROI and cluster functional connectivity changes, please see Table S1.

FIGURE 6.

Significant functional connectivity changes amongst thalamic nuclei, non‐thalamic structures, and cortical networks across associative S‐R learning during Phase 2 of the learning paradigm. See Table S1 for a comprehensive report of the statistical results. Blueish colours show decreasing functional connectivity, whereas reddish colours show increasing functional connectivity. Lighter colours indicate weaker effects and smaller T values (see colour bar for t‐value range). Thin black brackets indicate cluster boundaries, as determined by the initial hierarchical clustering step during the two‐stage statistical analysis within CONN. Shown are significant connectivity changes that survived multiple comparison correction based on the Functional Network Connectivity algorithm with a cluster threshold of p < 0.05 (FDR corrected) applied to individual connections thresholded at p < 0.05 (uncorrected). (A) Circular graph representation of the significant connectivity changes (B) Matrix display representation of the significant connectivity changes. Here, the asymmetric nature of gPPI‐based connectivity values becomes visible. Organised based on 3 sets of ROIs: (1) THA, thalamus: AV, anteroventral; LGN, lateral geniculate nucleus; LP, lateral posterior; MDl, mediodorsal lateral parvocellular; MDm, mediodorsal medial magnocellular; MGN, medial geniculate nucleus; PuA, pulvinar anterior; PuI, pulvinar inferior; PuL, pulvinar lateral; PuM, pulvinar, medial; VA, ventral anterior thalamus; VLa, ventral lateral anterior; VLp, ventral lateral posterior; VPL, ventral posterolateral. (2) SUBCORTICAL: AMY, amygdala; CAU, caudate; DA, dorsoanterior; DP, dorsoposterior; GP, globus pallidus; NAc, nucleus accumbens; PUT, putamen; VA, ventroanterior; VP, ventroposterior. (3) CEREBRAL CORTEX: CON, cingulo‐opercular; DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; HIP, hippocampus; PMN, parietal medial network; PON, parieto‐occipital network; SMd, somatomotor dorsal network; and SMl, somatomotor lateral network.

All results reported here are based on connectivity changes averaged across approach and avoidance learning conditions. The direct comparison between approach and avoidance learning conditions did not yield significant results. Also, we did not find significant correlations between the size of the compatibility effect and connectivity changes.

Whilst the functional connectivity changes between the ROIs in our study were clustered based upon similarity metrics within the present data set, this also coincided with known functional properties regarding cortical and subcortical regions in previous literature as described below (Figure 6A,B). All abbreviations are provided in Table 1.

TABLE 1.

List of subregions and corresponding abbreviations.

Regions of interest and abbreviations
Thalamus	Abbreviation	Amygdala	Abbreviation
Anteroventral	AV	Lateral amygdala	lAMY
Central medial	CeM	Medial amygdala	mAMY
Centromedian	CM	Caudate
Lateral geniculate nucleus	LGN	Ventroanterior caudate	VA CAU
Lateral posterior	LP	Dorsoanterior caudate	DA CA
Mediodorsal lateral parvocellular	MDl	Globus Pallidus
Mediodorsal medial magnocellular	MDm	Anterior globus pallidus	aGP
Medial geniculate nucleus	MGN	Posterior globus pallidus	pGP
Pulvinar anterior	PuA	Hippocampus
Pulvinar inferior	PUI	Hippocampus head, medial division, subdivision 1	HC head, medial 1
Pulvinar lateral	PuL	Hippocampus head, medial division, subdivision 2	HC head, medial 2
Pulvinar medial	PuM	Hippocampus head, lateral division	HC head, lateral
Ventral anterior	VA	Hippocampus body	HC body
Ventral lateral anterior	VLa	Hippocampus tail	HC tail
Ventral lateral posterior	VLp	Nucleus Accumbens
Ventral posterolateral	VPL	Nucleus accumbens, shell	NAcc, shell
Cortical Networks		Nucleus accumbens, core	NAcc, core
Cingulo‐opercular network	CON	Putamen
Default mode network	DMN	Ventroanterior putamen	VA PUT
Dorsal attention network	DAN	Dorsoanterior putamen	DA PUT
Frontoparietal network	FPN	Ventroposterior putamen	VP PUT
Language network	Language	Dorsoposterior putamen	DP PUT
Parietal medial network	PMN
Parieto‐occipital network	PON
Salience network	Salience
Somatomotor dorsal network	SMd
Somatomotor lateral network	SMl
Visual network	Visual

This means the ROIs within our study tended to form clusters with other subregions of the same region, such as thalamic nuclei clustering with many other thalamic subregions (especially those of a known similar functional role), hippocampal subregions having a tendency to cluster together (Strange 2022), as well as amygdala subregions (Keshavarzi et al. 2014). Furthermore, basal ganglia subregions also tended to form clusters amongst their own subregions, including the globus pallidus, caudate and putamen (Mattfeld et al. 2011). Furthermore, most cortical networks were clustered independently from each other. Most notably, higher‐order thalamic nuclei formed a cluster which consisted of many known CSTC loop nodes, namely MD, VLa and VLp which together comprise anterior cingulate, dorsolateral frontal, lateral orbitofrontal and motor CSTC loops (Sugiyama et al. 2018).

For convenience, we also display the connectivity changes for selective ROI clusters (defined by the initial hierarchical clustering step) in separate figures below (Figure 7A–D). Overall, the results were predominantly driven by decreasing functional connectivity involving three cortical networks (frontoparietal, cingulo‐opercular and default mode networks) with almost exclusively higher‐order thalamic nuclei, including mediodorsal, ventrolateral, lateral posterior, pulvinar and ventral anterior subdivisions (Figure 7A–C). In addition, associative S‐R learning was also associated with increasing intrathalamic functional connectivity amongst a cluster of higher order nuclei and decreasing intrathalamic connectivity between these higher order nuclei with LGN and CM regions (Figure 7D).

FIGURE 7.

Functional connectivity changes between thalamic regions and cortical networks and amongst thalamic nuclei across associative S‐R learning. This is a selective visual representation of the full results shown in Figure 6. The selected clusters include (A) functional connectivity changes between the frontoparietal network and thalamic regions, (B) functional connectivity changes between the cingulo‐opercular network and thalamic regions, (C) functional connectivity changes between the DMN network and thalamic regions, and (D) functional connectivity changes amongst thalamic regions. Red brackets indicate cluster boundaries which have been determined by the hierarchical clustering algorithm.

3.2.1. Changes in Connectivity During Learning in a Frontoparietal‐Thalamic Network

The FPN exhibited widespread decreasing functional connectivity changes with thalamic nuclei across associative S‐R learning trials (Figure 7A). Notably, all of these thalamic nuclei were higher‐order nuclei, most of which serve as proposed or established nodes within CSTC networks (Aton 2021; Provost et al. 2015), namely the mediodorsal, ventral anterior, ventral lateral, and lateral posterior nuclei.

3.2.2. Changes in Functional Connectivity During Associative S‐R Learning in a Thalamo‐Cingulo‐Opercular Network

In contrast, the CON network primarily showed increasing functional connectivity with various pulvinar nuclei across associative S‐R learning trials (Figure 7B). These changes occurred between CON and bilateral PuA, left PuI, and left PuL regions. Conversely, the only decrease in functional connectivity that was found occurred between CON and the right LGN. Notably, all of these thalamic nuclei are regions of the visual thalamus (Purushothaman et al. 2012), with the pulvinar being primarily a higher‐order nucleus, whilst the LGN is a first‐order nucleus (Sherman 2007).

3.2.3. Connectivity Changes During Associative S‐R Learning in a Thalamo‐Default‐Mode Network

The DMN showed widespread changes in functional connectivity with numerous thalamic regions (Figure 7C), almost all of which were either higher‐order nuclei or those involved in CSTC networks (Aton 2021; Provost et al. 2015; Saalmann 2014; Sherman 2007). These results were totally lateralised, whereby right MD, right pulvinar, and right CM nuclei showed decreasing functional connectivity with the DMN, whilst the left MD, left PuM, and left LP showed increasing functional connectivity with the DMN.

3.2.4. Changes in Intrathalamic Connectivity During Associative S‐R Learning

Finally, widespread changes in functional connectivity were observed amongst numerous thalamic nuclei (Figure 7D). Most notably, these changes featured increasing functional connectivity amongst the same higher‐order nuclei and CSTC nodes which exhibited changes in connectivity with FPN, CON and DMN networks. Specifically, these included MD, pulvinar, VLa, VLp and LP nuclei. Conversely, decreases in functional connectivity were observed between these higher‐order thalamic nuclei and the bilateral CM and first‐order LGN.

3.2.5. Connectivity Changes Between Thalamic Nuclei and Non‐Thalamic ROIs (hippocampus and Amygdala) During Associative S‐R Learning

In addition to functional connectivity changes between cortical networks and thalamic ROIs, numerous functional connectivity changes occurred between thalamic nuclei and non‐thalamic ROIs (Figure 8A,B). With the exception of two ROI‐to‐ROI connectivity changes between thalamic nuclei and the anterior globus pallidus, all functional connectivity changes occurred between thalamic nuclei with hippocampus and amygdala regions.

FIGURE 8.

(A, B) Functional connectivity changes between thalamic nuclei with hippocampal and amygdala subregions across associative S‐R learning. This is a selective visual representation of the full results shown in Figure 6. Shown are functional connectivity changes between thalamic and hippocampal subregions (A), and functional connectivity changes between thalamic and amygdala subregions (B).

First, functional connectivity changes between thalamic nuclei and hippocampal subregions revealed widespread decreasing connectivity across all higher‐order and CSTC node thalamic nuclei, namely CM, LP, MD, pulvinar, and VL subregions, spanning 19 ROI‐to‐ROI connections (Figure 8A). Conversely, only one increase in functional connectivity was observed and occurred between left PuI and right hippocampal head, medial division 2.

Second, functional connectivity changes also occurred between thalamic nuclei and amygdala subregions. Consistent with findings regarding the hippocampus, widespread decreasing connectivity was also observed between thalamic nuclei and amygdala subregions. However, this did not primarily occur amongst higher‐order or CSTC node nuclei, as this instead occurred amongst CM, LGN and VPL, alongside two pulvinar subregions (Kim et al. 2023; Saalmann 2014). Lastly, the various amygdala subregions also showed increasing functional connectivity with each other as behaviour became increasingly automatic.

3.3. Lateralised Functional Connectivity Between the DMN and MD Subregions

Learning‐related functional connectivity changes included lateralisation between left and right MD subregions with the DMN, as described in subsection 3.2.3. However, as the MD thalamus is a particularly important thalamic nucleus within goal‐directed behaviour and cognitive control (Parnaudeau et al. 2018), and due to the well‐known lateralisation of brain function (Rogers 2021), a two‐tailed test was subsequently performed to investigate if this was indeed a robust finding.

Within this test, seed regions included each of the four MD thalamic subregions and the DMN. Here, we found highly significant values for left–right comparisons between MD subregions and the DMN. In order to save for selecting a certain direction, we utilised the means across them and found highly significant values of MDl‐DMN (p = 0.00007) and MDm‐DMN (p = 0.002). In addition, highly significant values were also observed across both directions, namely when taking MD subregions as seed regions and the DMN as the target region, and vice versa.

3.4. Focused Analysis

The Focused Analysis revealed significant changes in functional connectivity between numerous thalamic nuclei and putamen subregions, see Figure 9 and Table S2. Various thalamic nuclei showed functional connectivity changes with putamen subregions; however, these were restricted to two subregions, namely ventroanterior and dorsoposterior regions. The ventroanterior putamen showed increasing functional connectivity with bilateral MDm, MDl and VLp subregions, alongside increasing functional connectivity with left AV and right LP nuclei. Conversely, the dorsoposterior putamen exhibited both increasing functional connectivity with bilateral AV and decreasing functional connectivity with right CM and PuA nuclei.

FIGURE 9.

(A–C) Significant functional connectivity changes between thalamic nuclei and putamen subregions across associative S‐R learning. See Table S2 for a comprehensive report of the statistical results. All significant functional connectivity alterations between thalamic nuclei and putamen subregions (A), alongside selective visual representations of the full results within (A), for the ventroanterior putamen and thalamic nuclei (B), and for the dorsoposterior putamen with thalamic nuclei (C). Note that these results were obtained within the Focused Analysis which selectively focused on connections between thalamic nuclei and putamen subregions with more lenient statistical correction requirements as the primary analysis.

3.5. Summary of Overall Functional Connectivity Results of the Current Study

In summary, the main functional connectivity results that occurred across stimulus–response learning within the current study are represented within Figure 10. This schematic depicts overall functional connectivity changes that occurred between thalamic nuclei with cortical networks, non‐thalamic subcortical regions, and between thalamic nuclei.

FIGURE 10.

Summary of the results for the main overall functional connectivity changes across S‐R learning. These results depict the changes in functional connectivity that occurred between thalamic nuclei and between thalamic nuclei with cortical networks and non‐thalamic regions across correct learning trials, within our stimulus–response learning paradigm. The putamen results are utilised from our Focused Analysis and for simplicity, results for hippocampus and amygdala subregions are collapsed. Blue colours show decreasing functional connectivity, and red colours show increasing functional connectivity. Regions of interest depicted in black font represent thalamic‐nonthalamic functional connectivity changes, whilst regions in white font represent intrathalamic functional connectivity changes.

4. Discussion

Our results showed that during the transition from goal‐directed towards more automatic behaviour, numerous thalamic nuclei exhibited functional connectivity changes with cerebral cortex networks, hippocampus, other thalamic nuclei, and subcortical regions (Figure 10). These changes primarily included decreasing functional connectivity between FPN and higher‐order thalamic regions, increasing functional connectivity between CON and the higher‐order visual thalamus, and lateralised functional connectivity changes between higher‐order thalamic nuclei and the DMN, alongside increased intrathalamic functional connectivity. Lastly, increasing functional connectivity also occurred amongst subcortical subregions, with increased functional segregation of these same subcortical subregions as behaviour became increasingly automatic.

4.1. Functional Reorganisation Occurs Between Numerous Cortical Networks and Higher Order Thalamic Nuclei as Associative S‐R Learning Progresses

Goal‐directed behaviour is coordinated by a variety of cortical networks, which include executive networks such as CON and FPN (Wang et al. 2024; Wood and Nee 2023). These networks have diverse contributions to behaviour, suggesting not only that these networks should show altering levels of activation across the transition from goal‐directed to habitual behaviour, but also that these networks should show altering functional connectivity with thalamic nuclei, which also have specialised contributions to goal‐directed behaviour, namely higher‐order thalamic nuclei or CSTC nodes.

Accordingly, our results indeed showed that the FPN, CON and DMN exhibit learning‐related changes in functional connectivity with various thalamic nuclei, and that these primarily occurred amongst higher‐order thalamic nuclei. These changes were also consistent with previously reported findings regarding the roles of higher‐order thalamic nuclei within goal‐directed behaviour, as outlined in further detail in the following subsections.

4.2. Task‐Positive Networks Show Differential Functional Connectivity Patterns Across Associative S‐R Learning

The successful completion of a cognitive task requires the activation of various cortical networks, which are therefore known as task‐positive networks (Mills et al. 2018), and includes the FPN and CON (Yu et al. 2019). In accordance with this, our results revealed that a variety of task‐positive networks showed changes in functional connectivity with thalamic nuclei across associative S‐R learning. However, our results showed that these changes primarily occurred between higher‐order thalamic nuclei with FPN and CON networks. Notably, the thalamus is a known node of both the FPN (Camilleri et al. 2018; Dixon et al. 2018) and CON (Sadaghiani and D'Esposito 2015).

First, the FPN showed widespread decreases in functional connectivity with higher order nuclei or CSTC nodes, whilst the CON showed increasing functional connectivity with solely the higher order visual thalamus (namely pulvinar nuclei), whilst decreasing functional connectivity with the first order visual thalamus (the LGN). Here, the FPN serves as a crucial task‐positive network alongside the CON (Sadaghiani and D'Esposito 2015). The function of the FPN is to reorient and support attention processes and to facilitate flexible and goal‐directed behaviour using feedback from responses (Hausman et al. 2022; Liu et al. 2023). On the other hand, the role of the CON is to implement task sets (Marek and Dosenbach 2018) and to regulate actions. In this way, it has been suggested that the FPN initiates and adjusts cognitive control through conscious attention, whilst the CON maintains the task set (Zanto and Gazzaley 2013).

Consequently, the changes in functional connectivity between the FPN and CON with higher‐order thalamic nuclei within our results appear to align with the known functional roles of these regions across goal‐directed and habitual behaviour. It may therefore be the case that goal‐directed behaviour is initially driven by strong connections between the FPN and higher order thalamic nuclei, enabling responses to be outcome‐driven and conscious, and that this high connectivity becomes progressively diminished as automaticity forms. Additionally, the increasing functional connectivity between the CON and pulvinar during our visual task may have aided the successful implementation of the task set of visual S‐R associations, as this increase in functional connectivity between CON and pulvinar nuclei also aligns with the pulvinar's specialised role in visual attention and in the handling of visual input (Berman and Wurtz 2008). Here, implementation of the task set is sustained through the transition from goal‐directed to habitual behaviour (Schneider and Logan 2014), a function which is present even within habitual behaviour and in its early phases may reflect ‘strategic automaticity’ (Sheeran et al. 2006).

Together, this lends further support to the differential contributions of cortical networks within the coordination of goal‐directed versus habitual behaviour, and suggests that the generation of behaviour across this behavioural transition is mediated by functional reorganisation with select higher‐order thalamic nuclei. Consequently, it may be the case that executive cortical networks may have a hierarchical organisation within goal‐directed behaviour, whereby transition to habit is characterised by reduced contributions of cognitive control and executive networks, which first begin with the FPN, whilst the CON sustains its influence upon behaviour across a longer period and towards automaticity. This would be consistent with findings that higher‐order cognitive control has a gradient from higher to lower or rostral to caudal regions (Choi et al. 2018), such that within the frontal lobe, a hierarchy flows from schematic control towards contextual and finally sensorimotor control (Zhu and Han 2022). Accordingly, the nodes of the FPN and CON fall within this hierarchy, as does the rostral‐caudal organisation of higher‐order versus first‐order nuclei within the thalamus (Vertes et al. 2015).

Finally, regarding the decreasing functional connectivity between higher‐order thalamic nuclei and the FPN, we speculate that this striking result may also be explained by a decreasing demand upon the FPN once learning has been successfully achieved. Here, the FPN is likely to have exhibited elevated relevance whilst task sets and correct stimulus–response associations were being discovered during earlier trials (i.e., the current task set being assembled). FPN involvement then decreased once the task sets had been fully established and were now being ‘run’ by the CON. This would also explain the increasing connectivity of thalamic nuclei with CON throughout the experiment, as more task sets would have been set up by the FPN.

4.3. The DMN Shows Diverging Functional Connectivity Changes With Lateralised Thalamic Subregions Across Associative S‐R Learning

The generation of automatic behaviour is associated with DMN activation (X. Zhou and Lei 2018) and de‐coupling of the DMN with task‐positive cognitive networks (Baumann et al. 2023; Vatansever et al. 2017). Here, in contrast to task‐positive cognitive networks, task‐negative networks activate during internally‐directed attention, and include the DMN (S. Weber et al. 2022). The DMN is involved in internally‐focused thought processes, such as mind‐wandering, future planning, self‐reflection and memory recall (Menon 2023), and it is known to activate during distraction and boredom, especially within repetitive tasks (Danckert and Merrifield 2018). Repetitive tasks, by nature, foster habitual responding (Stojanovic et al. 2021). Consistent with this, our own results also show that during increasing automaticity, decoupling occurred between the DMN and task‐positive cortical networks, including CON, DAN and PMN.

Furthermore, this functional decoupling occurred alongside changes in functional connectivity with lateralised thalamic regions. Here, left‐hemispheric thalamic regions showed increasing functional connectivity with the DMN, whilst right‐hemispheric thalamic regions showed decreasing functional connectivity. These changes occurred amongst higher order nuclei, particularly mediodorsal and pulvinar nuclei. Specifically, this included the left MDl, MDm, LP and PuL and the right MDl, MDm, PuA, PuI and PuM. This is aligned with a huge body of research which reports lateralisation across the cerebral cortex regarding many cognitive and executive functions, including decision making (Corser and Jasper 2014), attention (Bartolomeo and Seidel 2019), and in the access of semantic information (Liuzzi et al. 2021). Therefore, our results suggest that higher‐order thalamic nuclei also have lateralised functional contributions to goal‐directed and habitual behaviour, and are consistent with a growing body of research which reports thalamic functional lateralisation across sensorimotor functions (Herrero et al. 2002), language (Llano 2013), and speech (Wang, Lipski, et al. 2022). There is also evidence to suggest that the DMN shows leftward‐lateralisation in its own functioning, and that this lateralization is associated with healthy cognitive functioning (Banks et al. 2018). In addition, a study by Yuan et al. (2016) investigated resting‐state functional connectivity of thalamic nuclei with cortical networks and generated functional connectivity maps. They reported that the DMN has left‐lateralisation within the thalamus, including within the MD thalamus. Correspondingly, upon performing our own statistical test on these lateralised functional connectivity results, we found highly significant values for left–right comparisons between MD subregions and the DMN. This was observed when taking MD subregions as seeds with the DMN as the target region, as well as across the opposite direction. Here, in order to save for selecting a certain direction, we utilised the means across them and found highly significant values of MDl‐DMN (p = 0.00007) and MDm‐DMN (p = 0.002). Accordingly, we expect that our finding is of interesting functional significance due to its alignment with the aforementioned left‐lateralisation of the DMN, although further research would be beneficial for untangling its functional contributions.

4.4. The Mediodorsal Thalamus Is Involved in Switching From the FPN Towards the DMN Across Associative S‐R Learning

The MD thalamus has been identified as a special thalamic region within goal‐directed behaviour. Functionally, it appears to be central in coordinating flexible responses (Weber et al. 2023) through the updating of action‐outcome and stimulus‐outcome associations across rule or contingency changes (Wolff and Vann 2019), as well as in the maintenance of rule representations (Schmitt et al. 2017). This ability is essential within the flexible control of behaviour and within our own associative S‐R learning paradigm, in order for subjects to successfully discard associations learnt within previous phases and to maintain relevant and current task rules for successful performance across task phases and within each phase. Therefore, the MD nuclei appear to have critical roles in response‐outcome driven behaviour.

Here, the MD nuclei are known to connect with the FPN and support its functioning within cognitive tasks (Li et al. 2022; Shine et al. 2023), and they are also argued to be a central node of the DMN, thereby also influencing DMN functioning during internally focused cognition (Aguilar and McNally 2022). Furthermore, the MD nuclei are classically defined by their extensive connections with the prefrontal cortex (PFC) (Delevich et al. 2015), for which the PFC is a region which is also well known for being central in generating flexible goal‐directed behaviour (J. Weber et al. 2023). Therefore, as what might be expected, the MD nucleus accordingly projects primarily to the prefrontal cortex and is considered to be the ‘principal thalamic nucleus for’ (Zikopoulos and Barbas 2007) and ‘essential partner of [the] prefrontal cortex’ (Parnaudeau et al. 2018). Here, the MD nuclei are known nodes within both the prefrontal or executive CSTC loop (Sugiyama et al. 2018) as well as the limbic CSTC loop (Gibson et al. 2017), for which they also connect extensively with the basal ganglia (Smith et al. 2022). At the level of the thalamus, the MD nuclei serve as the ‘associative nucleus of the thalamus’ (Ji et al. 2019), consistent with the prefrontal lobe as an association area which mediates top‐down processing of sensory and motor information (Siddiqui et al. 2008; Spellman et al. 2021). Finally, extensive connections also run to and from other association cortices and the limbic system, enabling the MD nuclei to serve as an integration centre for somatosensory information (Hika and Khalili 2023). Here, cortical networks, including the FPN and DMN, form from these association areas (Yeo et al. 2011) in order to drive behaviour.

Therefore, the MD nuclei are in a unique and prime position to coordinate cortical network activity, including amongst task‐positive cognitive networks such as the FPN and the task‐negative DMN, and may thereby serve as a mechanism for switching behavioural control across task‐positive versus task‐networks. Here, through this dynamic interaction between MD nuclei with both the FPN versus DMN, the reorienting of attention between external versus internal loci of attention (V. Smith et al. 2018) could take place across shorter timescales or within early learning, facilitating flexible behavioural switching between response‐outcome versus stimulus‐outcome associations (Parnaudeau et al. 2018). Through this mechanism, the MD nuclei may therefore mediate the attention switching which occurs between the external task and internal phenomena during behavioural switching throughout learning, and ultimately, on a longer time frame, this could also underlie the transition from goal‐directed towards automatic control of behaviour, via reduced functional connectivity with the FPN and stabilised functional connectivity with the DMN, thereby driving goal‐directed versus automatized control of behaviour, respectively. Overall, this would also be consistent with the findings of Delavari et al. (2024), who investigated micro‐activation patterns (μCAPs) on a sample of healthy controls and patients with schizophrenia 22q11.2 deletion syndrome. This study found co‐activation patterns between MD nuclei and the DMN which suggested that a delicate balance between task‐positive and task‐negative network activation is required for typical functioning. Therefore, our results are also of significance within clinical research and inform potential thalamo‐network contributions to disorders featuring psychosis, which are known to exhibit reduced goal‐directed behaviour (Cooper et al. 2019).

As such, we propose that the MD nuclei may therefore serve as a critical mechanism for switching between goal‐directed versus automatized behaviour. Here, we speculate that the MD thalamus may act as a crucial thalamic node across cortical networks, through its increasing functional connectivity with the FPN during cognitive tasks and in its switching of connectivity instead towards the DMN during automaticity. This would thereby enable the MD nuclei to contribute to cognitive tasks and then to monitor automatic behaviours as behaviour becomes increasingly automatized.

4.5. Subcortical and Cortical Systems Become Increasingly Functionally Segregated as Automaticity Forms

A growing body of research is reporting that habitual behaviour is driven predominantly by subcortical regions of the brain, with cortical areas diminishing their influence as goal‐directed contributions decrease (Gillan et al. 2016; Liljeholm et al. 2015). In accordance with this, our own results show decreasing functional connectivity between several cortical networks and widespread regions of the thalamus and other areas, which included FPN, DAN, PMN, the hippocampus and amygdala. Together, this suggests that as behaviour becomes increasingly automatic, that cortical and subcortical regions become increasingly functionally segregated and isolated. Here, response‐outcome contingencies become progressively devalued or removed altogether as habits come to fruition (B. Gardner 2015; Schreiner et al. 2020), with behaviour becoming increasingly driven by stimuli instead of outcome. Overall, these functional connectivity changes appear to reflect decreasing cortical influence upon behaviour, and a shift towards subcortical control with the thalamus at its centre; a region which continues to receive bottom‐up information from the sensory periphery (Mashour and Hudetz 2017). Consequently, this may give rise to the impaired awareness and reflection upon behaviour which arises with automaticity (Gardner et al. 2024).

Furthermore, these same regions also became increasingly functionally segregated from each other, with the exception of the nucleus accumbens and caudate nucleus. This was also the case between the thalamus and hippocampus, and suggests that as behaviour becomes more automatic, the thalamus also becomes increasingly functionally isolated from memory processing areas, alongside cognitive areas. However, conversely, functional connectivity increased amongst these same regions, such that subregions of the caudate nucleus, hippocampus and amygdala became increasingly connected with subregions of their own anatomical region.

Altogether, this aligns with the vast body of research which suggests that the subcortical system becomes increasingly permitted to drive responses without moderation by cortical networks, driving the transition from outcome‐driven towards stimulus‐driven responses, even when doing so produces errors as a result of rule or contextual changes (Smith and Graybiel 2016). Here, this delegation of behavioural responding to the subcortical system in automaticity has been hailed as a ‘fundamental innovation in human cognition’ (Shine and Shine 2014) which frees the cerebral cortex to handle more complex and abstract psychological mentation, through this delegation of control to lower order areas. At the connectivity level, this behavioural change may therefore be mediated through a transition from task‐positive networks via elevated thalamocortical connectivity towards increased control amongst increasingly functionally segregated subcortical structures and decreased thalamocortical connectivity, giving rise to habitual control.

4.6. Functional Connectivity Between Thalamic Nuclei Is Associated With Increasing Automaticity

Not only does transition occur from cortical towards subcortical control, but this is mirrored by transition towards a less connected thalamocortical system, as the thalamus diminishes connectivity with FPN and DAN, and reorganises its activity with DMN. However, in addition to increasing functional segregation across associative S‐R learning and automaticity, higher‐order thalamic nuclei and CSTC nodes showed increasing functional connectivity amongst each other, namely between LP, MDl, MDm, PuA, PuL, PuM, VLa and VLp nuclei. Notably, these nuclei were those same nuclei which exhibited changed functional connectivity with cortical networks and putamen subregions. Consequently, it may be the case that as automaticity develops, thalamic nuclei thereby potentially form a functional network specific for mediating or assisting the generation of automatic behaviour. Here, the exception of increasing functional connectivity between the CON and pulvinar is consistent, due to the CON's role in implementing task sets (Dosenbach et al. 2006), and potentially within strategic automaticity before automaticity eventually solidifies and becomes controlled via thalamic rather than cortical processes.

The thalamus is therefore a prime candidate to assume control over behaviour in automaticity. It is a highly unique subcortical region which is formed from the posterior part of the forebrain within the embryo (Scholpp and Lumsden 2010), and has been termed ‘the principal information hub’ of the vertebrate brain (Govek et al. 2022). The thalamus receives bottom‐up input from almost all of the senses (Courtiol and Wilson 2014), integrates multisensory information (Tyll et al. 2011), generates consciousness (Whyte et al. 2024), and generates motor responses in coordination with other subcortical areas (Bosch‐Bouju et al. 2013), as cortical motor areas reduce their influence within habitual movements or after long‐term learning (Hwang et al. 2019).

Whether the functional connections between thalamic nuclei are also supported by direct intrathalamic connections, however, remains unknown. There are some reports about direct structural connections between several thalamic nuclei in non‐human animals (Rodrigo‐Angulo and Reinoso‐Suárez 1988; Perry and Mitchell 2019; Swanson et al. 2019), but also apparently a difficulty to detect them (Crabtree et al. 1998).

In schizophrenia, alterations in functional connectivity between thalamic nuclei are observed between dorsolateral, ventral anterior and ventromedial portions of the thalamus (Gong et al. 2019). This is of particular interest given the impaired goal‐directed behaviour that can be associated with this disorder (Bowie and Harvey 2006; Cooper et al. 2019), including impaired action‐outcome learning, inflexible responses and reduced behavioural adaptation (Morris et al. 2018). Altogether, schizophrenia may therefore be indicative of crucial intrathalamic functional connectivity patterns which likely contribute to functioning within healthy subjects, and may contribute to the symptoms that are observed in schizophrenia. Further research could utilise the learning paradigm within our study, alongside a sample of patients with schizophrenia, to investigate functional connectivity alterations within this cohort. In this, it could be expected that alterations in functional connectivity patterns across the transition from goal‐directed behaviour towards automaticity will be found, given the impairments in task shifting in schizophrenia alongside its status as a disorder of connectivity amongst large‐scale networks, which specifically reflect reduced network integration and segregation (Wang, Hu, and Li 2022). Indeed, whilst such investigation is sparse, at least one study has found altered circuitry underlying habitual behaviour in schizophrenic subjects (Weickert et al. 2002), with others also proposing that these alterations in the habit system exist (Avery et al. 2019; Morris et al. 2018).

4.7. Thalamic Nuclei Show Changes in Functional Connectivity With Putamen Subregions

From a theoretical perspective, increasing automatisation might be expected to involve the putamen. However, our primary analysis did not reveal significant connectivity changes involving the putamen. In light of the theoretical relevance, we performed a post hoc analysis with more lenient correction requirements. This investigation of changes in functional connectivity selectively between thalamic nuclei and putamen subregions unveiled changes occurring primarily between higher‐order thalamic nuclei with ventroanterior and dorsoposterior subregions of the putamen. Here, the ventroanterior putamen exhibited increasing functional connectivity with bilateral MDm, MDl and VLp, alongside left AV and right LP nuclei. Moreover, the dorsoposterior putamen showed increasing functional connectivity with bilateral AV and decreasing functional connectivity with right PuA and CM nuclei.

The putamen is a component of the basal ganglia system, serving alongside the caudate nucleus as the dorsal striatum and main input region to the basal ganglia (Zheng et al. 2023). The putamen receives excitatory projections from cortical sensorimotor, midline thalamus and cerebellar areas (Kunimatsu et al. 2019; Simioni et al. 2016; Starr et al. 2011), and serves as a node within both direct and indirect pathways for motor generation and inhibition, respectively (Young et al. 2023). Therefore, the putamen acts as an action filter for behaviour within goal‐directed behaviour (Hikosaka et al. 2019) and is also well‐known for its importance in generating habitual behaviour (Balleine and O'Doherty 2010; Tricomi et al. 2009).

Results from a meta‐analysis reveal a range of functional differences between anterior and posterior subdivisions of the putamen (Pauli et al. 2016). Whilst both sub‐regions are related to sensorimotor processes, the anterior putamen is functionally connected not only with sensorimotor regions in the cerebral cortex, but also appears to have language‐related specialisations. In contrast, the posterior putamen is connected not only to cortical sensorimotor areas, but also with temporal cortex and insula regions. Furthermore, recent research finds that the anterior putamen is critical for newly acquired habits, whereas more extensively trained stimulus–response associations seem to rely more heavily on the dorsoposterior putamen (Guida et al. 2022).

Therefore, within our own study, transition from goal‐directed towards habitual control may have been facilitated by these thalamus‐putamen functional connectivity changes. In accordance with this, these nuclei included higher‐order thalamic nuclei, alongside thalamic nuclei which are established nodes of direct and indirect basal ganglia pathways (Mitchell 2015; Perry and Mitchell 2019; Rocha et al. 2023). Furthermore, the putamen subregions showing changed functional connectivity also have involvement in sensorimotor processing, the encoding of stimulus value, and stimulus‐driven motivation (Graff‐Radford et al. 2017). This suggests that thalamic nuclei may mediate the transition from goal‐directed to more habitual behaviour through their connections not only with goal‐directed cortical networks, but also through their dynamic connectivity with putamen subregions.

4.8. No Significant Association Between Functional Connectivity Changes and Behaviour

Previous results in humans showed a rather mixed picture regarding the association between neural and behavioural markers of habit learning (Gera et al. 2023; Tricomi et al. 2009; Wang et al. 2023; Zwosta et al. 2018). It has been notoriously difficult to reliably find such an association with respect to the putamen as one of the prime candidate brain regions typically assumed to be involved in automatic control based on animal research (Gera et al. 2023; Zwosta et al. 2018). The present study adds to this picture, as there was no significant correlation between connectivity changes and the size of the compatibility effect as a putative proxy of acquired habit strength. Overall, caution seems advised regarding the validity of behavioural markers of habit strength, at least in the context of human behaviour (Pool et al. 2022; Watson and de Wit 2018).

4.9. Limitations of the Current Study

In our study, we utilised a relatively low spatial resolution of 4 mm × 4 mm × 4 mm. This resolution is adequate to distinguish distinct thalamic nuclei across subjects, such as the relatively large MD and pulvinar (Byne et al. 2002; Highley et al. 2003). For some smaller nuclei, such as lateral dorsal or parafascicular nuclei (Mai and Majtanik 2018), or subsections within MD, this resolution may not allow for accurately distinguishing thalamic nuclei. Within our own study, the mean nuclei sizes for individual MD nuclei subregions ranged from 253 to 817 mm³. Furthermore, numerous other thalamic nuclei also ranged in size within these functional voxel limits, see Table S1. However, for clearer delineation of thalamic nuclei, and the inclusion of smaller thalamic nuclei, higher resolution MRI studies and techniques to deal with partial volume effects and lower signal‐to‐noise ratio are therefore necessary (Alemán‐Gómez et al. 2020; Tabelow et al. 2009). Also alternative parcellation methods, such as through white‐matter‐nulled MP‐RAGE imaging, can offer improved thalamic segmentation (Su et al. 2019). In addition, our mean LGN size was 250 mm, which is larger than typically reported, and may be due to a poor T1 image signal affecting the lateral portions of the thalamus.

Second, our study was relatively exploratory in nature regarding changes in functional connectivity between thalamic nuclei, cortical networks and non‐thalamic regions. Consequently, we deliberately avoided use of techniques such as Dynamic Causal Modelling (DCM) which require a much more directional and driven approach to processing and analysing the data (Zeidman et al. 2019). Therefore, future research may wish to take a more directional approach, by incorporating the findings from our study to inform initial hypotheses.

Third, there are several possibilities that can explain the lack of brain‐behaviour relationship in our study. First, the brain‐behaviour relationship may not be linear when it comes to functional reorganisation between thalamus nuclei and other brain areas or networks. Second, the goal‐habit paradigm (Zwosta et al. 2018) may render it difficult to induce or measure habitual or automatic behaviour, such as through a lower behavioural intensity or by a temporal lag between neural effects and corresponding behavioural changes. Third, although the participant sample in our study was relatively large, it might not be large enough for detecting putatively small effects because of the low signal‐to‐noise ratio in subcortical regions.

Fourth, our study was constrained in the number of brain regions and subregions that were included as regions of interest. This was due to the need to minimise the large numbers of networks, regions and subregions that were already deemed particularly important for inclusion due to their known roles within goal‐directed behaviour and associative learning. Consequently, further regions which may also be important within goal‐directed behaviour and learning have been omitted. One prominent example is the cerebellum, a region that is becoming known for its own contributions to cognitive processes, is known to have functional interactions with the thalamus, and boasts its own cognitive and limbic functional subdivisions (Rudolph et al. 2023; Schmahmann 2019). Future research would therefore be beneficial in order to explore potential interactions between the cerebellum and thalamus, amongst other areas, and their contributions to associative S‐R learning and goal‐directed behaviour.

5. Conclusions

Across associative S‐R learning and as automaticity increases, thalamic nuclei exhibit various functional connectivity changes with cerebral cortex networks and non‐thalamic regions. These alterations occurred predominantly amongst higher order thalamic nuclei, for which there are three key conclusions. First, we found that thalamic nuclei exhibit decreasing functional connectivity with cerebral cortex networks, especially the FPN as automaticity increases. Second, we observed that the DMN exhibit decoupling with other cerebral cortex networks, for which the mechanism for switching between FPN and DMN may be mediated by MD thalamic nuclei. Third, automaticity was characterised by increasing functional connectivity amongst higher order thalamic nuclei, which may therefore constitute an intrathalamic network. Altogether, this indicates that associative S‐R learning is characterised by massive functional reorganisation of the thalamocortical system, resulting in functional segregation of cerebral cortex versus subcortical systems as behaviour becomes increasingly supported by subcortical structures. Finally, this may underlie the decreased influence of cognition and increasingly stimulus‐driven mode of behaviour which characterises automatic action.

Disclosure

The manuscript has not been published elsewhere, but has been posted to the preprint server at bioRxiv.

Ethics Statement

The experimental protocol was approved by the Ethics Committee of the Technische Universität Dresden (EK306082011) and all subjects gave written informed consent prior to the experiment.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Detailed results for the group‐level functional connectivity analysis obtained using the Functional Network Connectivity multivariate parametric statistics implemented in the CONN toolbox. The table lists significant connectivity clusters for network pairs with p < 0.05 FDR‐corrected across clusters (in bold font) together with univariate statistics for individual connections within each cluster surviving an uncorrected threshold of p < 0.05.

Table S2: Within the Focused Analysis, this table shows detailed group‐level functional connectivity analysis results obtained using a ROI‐level p‐FDR correction (ROI mass/intensity) false‐positive control method implemented in the CONN toolbox, showing mass. The table lists significant results for ROI pairs with a p‐uncorrected connection threshold of 0.05 and a cluster‐level p‐FDR corrected threshold of 0.05.

Jarrett, C. , Zwosta K., Wang X., et al. 2025. “Progressive Changes Between Thalamic Nuclei and Cortical Networks Across Stimulus–Response Learning.” Human Brain Mapping 46, no. 15: e70382. 10.1002/hbm.70382.

Data Availability Statement

The single‐subject connectivity matrices together with the group‐level output generated by the CONN toolbox together with the relevant MATLAB code will be made available at https://osf.io.