2.1.1 Participants and Study Overview

Twenty healthy young- to middle-aged adults (range 18–45) participated (10 females; mean age = 25; standard deviation [SD] = 5.56). All participants were right-handed, as evaluated by the Edinburgh Handedness Inventory (Oldfield 1971); they were native German speakers and reported no history of neurological or psychiatric conditions. Participants also had no history of alcoholism, drug use, or contraindications for MRI or tDCS (Antal et al. 2017).

2.1.2 Baseline Assessments

Participants underwent an initial telephone screening to confirm their eligibility. Following this, a comprehensive neuropsychological assessment was conducted to ensure normal cognitive function. Participants were also screened for depression (German version of Beck's Depression Inventory, BDI-II; Kühner et al. 2007).

During the neuropsychological assessment, verbal memory was assessed with the German Auditory Verbal Learning Test (verbal learning and memory test [VLMT]; Müller et al. 1997). Visual–spatial memory was examined using the Rey–Osterrieth Complex Figure Test (ROCFT; Zhang et al. 2021). Short-term memory was assessed with the Digit Span (Woods et al. 2011). Executive functioning was assessed using the Trail Making Test (Parts A and B) (Tombaugh 2004), the Stroop Color–Word Test (Van der Elst et al. 2006), and the Regensburg Verbal Fluency Test (Harth and Müller 2004). Verbal intelligence and vocabulary knowledge were assessed with the German version of the multiple-choice vocabulary test (MWT; Lehrl 2018).

Participants also completed a short version of the OLM task, both as practice for the main intra-scanner task and to classify them as “high” or “low” performers in the learning task. This classification is relevant to tDCS outcomes (Meinzer et al. 2013; Perceval et al. 2020) and can serve as a predictor of effects in subsequent active stimulation phases of the larger study. Participants’ characteristics are provided in Table 1.

2.1.3 OLM Task Design

Each participant underwent two task fMRI sessions, separated by at least 1 week. During each session, participants performed one of two parallel versions of the OLM task, adapted from previous studies by our group (Flöel et al. 2012; Prehn et al. 2017; Antonenko et al. 2018; de Sousa et al. 2020). In brief, the paradigm involves learning associations between objects (houses on a two-dimensional [2D] map) and their locations through repeated presentations, utilizing both instruction- and feedback-based mechanisms. The task employed in the present study uses a block design with four learning stages, which minimizes the effects of confounding factors related to time on task, novelty of being in the scanner, and fatigue (Sliwinska et al. 2017).

An overview of the intra-scanner task design and the experimental procedures is provided in Figure 1. In the learning task, participants were required to indicate whether a house was correctly positioned on a schematic street map, receiving immediate feedback to assist in learning the correct locations of each house across repeated trials. Without prior training in the novel houses and location combinations, participants were asked to judge whether the house they saw was at the “correct” position on a screen. Following each trial, they received written feedback indicating whether their response had been correct or incorrect; regardless of their accuracy, this feedback also provided the correct location of the house. If participants failed to respond within the allotted time, they were informed that they had missed the trial but received the correct house's location information. When a house was presented for the first time, participants were encouraged to guess whether its association with the map's shown position was accurate or not, utilizing the subsequent feedback to learn about the correct position. A control task required determining whether a house was on the left or right side of the map displayed on the screen, controlling for perceptual and motor demands without involving memory associations.

Each learning block comprised 40 trials, repeated across six blocks (30 object–location associations total). In each stage, five houses were presented pseudorandomly in correct and incorrect locations (five of each). The same set of five houses was repeated across all four stages of a block, enabling participants to learn five unique object–location associations per block. The control task followed an identical structure, consisting of two blocks of four stages each, and utilized a total of 10 different houses for right/left decisions.

In both experimental conditions, the trial duration was 4.5 s (Figure 1). Each trial presented a house positioned at a specific location on the map, and the same house picture was never presented consecutively. The house–location association was presented for 2.5 s, and participants were required to respond during this time frame. An additional 2 s were included to provide trial-specific feedback. Therefore, participants had to provide their response as quickly and accurately as possible by pressing the response buttons of an MRI-compatible response grip with their left hand's index finger or thumb for “yes” and “no” responses, respectively. Participants were informed that only their first response was recorded to ensure a balance between accuracy and response latency.

A fixation cross was incorporated during the experiments to establish a low-level baseline condition. Each fixation cross was synchronized with the experimental trial duration (4.5 s) and was presented after each stage (10 trials) within each learning block. Additionally, after each learning block, a longer fixation cross lasting 18 s was introduced to facilitate the transition from the learning to the control task or the reverse. The order of the learning and control blocks was randomly assigned by the presentation software. At the beginning of each block, participants received a brief written speech bubble to introduce the respective task conditions.

Participants were shown instructions for performing the task before their first session. Task order was randomized, and two counterbalanced task versions (A and B) featured unique houses and rotated maps to maintain comparable difficulty and performance across sessions. Stimuli included 90 images of real-world buildings selected for recognizability and complexity (Flöel et al. 2012), paired with schematic street maps. Maps were divided into quadrants with assigned correct and incorrect locations, ensuring balanced and randomized placement over the entire street map across trials.

Stimuli were presented using Presentation software (Version 20.1, Neurobehavioral Systems Inc., Berkeley, CA, USA) and displayed via an MRI-compatible screen and mirrors. MRI-compatible response grips were used for behavioral decisions (NordicNeuro, Norway, www.nordicneurolab.com).

2.1.4 Focal Sham tDCS

The present study aimed to identify key brain regions activated during fMRI using a specific OLM paradigm. In addition, it aimed to assess the paradigm's TRR. Sham tDCS was administered during both imaging sessions to establish a reliable and feasible workflow for conducting tDCS-fMRI experiments, including neuronavigated electrode placement, intra-scanner tDCS procedures, task implementation during scanning, and optimization of fMRI acquisition parameters. Furthermore, the use of sham stimulation provided a methodologically comparable control condition for future phases of an overarching project that will involve active stimulation (Meinzer et al. 2024).

A focal 3 × 1 tDCS setup with a multi-channel direct current (DC) stimulator was used. Sham tDCS was administered during functional imaging sessions using MRI-compatible equipment, including a multi-channel DC stimulator (DC-STIMULATOR MC, NeuroConn GmbH, Germany), circular conductive rubber electrodes (diameter = 2 cm), and cables (for details, see Niemann et al. 2024). The sham stimulation protocol comprised 10-s ramp-up and ramp-down periods, with 20 s of DC application at 2 mA in between. This protocol is designed to mimic the initial physical sensation typically associated with active tDCS without inducing sustained alterations in cortical excitability (Nitsche et al. 2008). Sham tDCS was administered using the same montage and current intensity that will be used in the subsequent active and sham tDCS arms of the overarching project.

The right lateral occipito-temporal cortex (right occipito-temporal cortex [ROTC]) was chosen on the basis of a previous neuroimaging study (Gillis et al. 2016) to serve as a preliminary target at this preparatory stage. Recent evidence demonstrated structural and functional connectivity between the lateral occipital and temporal cortices and the hippocampus (Dalton et al. 2022; Rolls et al. 2024). Although the hippocampus is known to play a central role in OLM, it is not directly accessible by tDCS. Therefore, targeting functionally connected neocortical regions such as the ROTC may provide an effective, indirect approach to modulate hippocampal activity (Nilakantan et al. 2017; Tambini et al. 2018).

The anode was placed over the individually localized target region and was surrounded by three equally spaced cathodes (the distance between cathodes was kept constant between participants). The distance of the cathodes was determined by a simulation pilot study, which ensured that this setup would reach an intended threshold of stimulation intensity of 0.2 mA across participants and projects of the Research Unit MeMoSLAP. Electrode positioning was standardized using a 3D-printed thermoplastic spacer with a 3 cm radius (center-to-center distance between the anode and cathodes, Niemann et al. 2024).

The algorithm used for montage individualization aimed to ensure that the fields are robustly centered on the individual target and minimize co-stimulation of surrounding areas. Tissue segmentation of the head and brain (based on individual tetrahedral head meshes generated from T1- and T2-weighted images using “charm”) and electric field (EF) modeling was performed using SimNIBS v4 (simnibs.org; Thielscher et al. 2011; Puonti et al. 2020).

This approach ensured precise electrode placement during the fMRI-tDCS sessions. Although this study employed only sham tDCS to investigate potential target regions, these measures were implemented to simulate the planned, individualized modeling for the subsequent active stimulation phases of the larger tDCS study.

2.1.5 MRI Acquisition

MRI data were acquired at a 3.0 Tesla Siemens MAGNETOM Vida scanner (Siemens Healthineers, Germany) at the University Medicine Greifswald with a 64-channel head–neck coil. fMRI data were acquired using a multiband echo-planar imaging (EPI) sequence developed by the Center for Magnetic Resonance Research (CMRR) at the University of Minnesota (https://www.cmrr.umn.edu/multiband/). This sequence, incorporating simultaneous multislice (SMS) acceleration, was optimized for blood oxygenation level–dependent (BOLD) contrast to enable high temporal resolution while minimizing slice repetition time and maximizing brain coverage (Setsompop et al. 2012; Xu et al. 2013).

Task-based fMRI scans were acquired with a 110 × 110 matrix, a spatial resolution of 2 × 2 mm², and a slice thickness of 2 mm with no interslice gap. The imaging parameters were as follows: repetition time (TR) = 1000 ms, echo time (TE) = 30 ms, flip angle (FA) = 60°, field of view (FOV) = 220 mm, and a multiband acceleration factor of 6. Each session lasted approximately 30 min, during which 1800 volumes were acquired with phase encoding in the anterior-to-posterior (AP) direction.

To mitigate distortions and signal drop-out, the protocol incorporated an echo spacing of 620 µs and parallel imaging with GRAPPA (GeneRalized Autocalibrating Partially Parallel Acquisitions) at an acceleration factor of 2. Slice timing was interleaved to reduce motion-related artifacts. Additionally, field maps were acquired for geometric distortion correction. High-resolution structural images were also obtained, including T1-weighted images (0.9 × 0.9 × 0.9 mm³, TR = 2700 ms, TE = 3.7 ms, TI = 1090 ms, FA = 9°) using selective water excitation for fat suppression and T2-weighted images (0.9 × 0.9 × 0.9 mm³, TR = 2500 ms, TE = 349 ms).

2.2.1 Behavioral Data Analysis

Both response accuracy and response latency (mean reaction time [RT] of correct responses) were analyzed. To examine the effect of task type (learning vs. control) and stage (1–4) on participants’ performance accuracy in our OLM task, we employed a generalized linear mixed-effects model (GLMM) with a binomial link function. A GLMM was chosen to account for both fixed and random effects, accommodating the hierarchical structure of the data. The analysis was conducted in R (version 4.3.0) using the glmer function from the lme4 package, applying a logistic linear mixed-effects model (LMM) to model response accuracy as a binary outcome (correct vs. incorrect responses). Given the positively skewed distribution of RT, we used a GLMM with a gamma distribution and a log-link function for RT of the corrected responses.

The model included stages (1–4), task type (learning vs. control), their interaction, fMRI session (Session 1 or 2), task version (A or B), and block (six blocks in the learning task and two in the control task) as fixed effects. To account for individual variability, we included a random intercept for participants.

The statistical significance of fixed effects and their interactions was evaluated. To further explore the interaction between task type and stage, we conducted post hoc analyses using estimated marginal means (EMMs) via the emmeans package. EMMs were computed for each combination of task type and stage to examine how response accuracy and RT varied between learning and control tasks across the four stages.

2.2.2 fMRI Data Analysis

fMRI data were analyzed using statistical parametric map software SPM12 (version 7771) and MATLAB (R2021b). Structural T1 images underwent skull stripping, bias correction, and normalization to the MNI template while preserving the original voxel size. Functional BOLD images were preprocessed through realignment, unwarping, and slice timing correction (TR = 1 s, 72 slices). Coregistration of T1 and BOLD images was performed, followed by functional image normalization to MNI space using deformation fields. Lastly, spatial smoothing with a 6 mm FWHM Gaussian kernel was applied to improve the signal-to-noise ratio and meet GLM assumptions.

For the first-level analysis of each participant, GLM was constructed to capture task-related neural activity. The model included regressors (trial onsets, durations) for each experimental condition (learning and control) across four stages (1–4). These events were modeled as stick functions and convolved with statistical parametric map's (SPM's) canonical hemodynamic response function (HRF) to reflect the delayed and dispersed characteristics of the BOLD response. Data from the two task fMRI sessions per participant were concatenated into a single design matrix. Nuisance regressors, including six motion parameters from realignment and framewise displacement, were added to control for head motion artifacts.

Contrasts of interest were defined to compare the task conditions, with the primary contrasts being the t-contrast of (learning > control and control > learning). Contrast images were generated for each participant and entered into the second-level group analysis.

The second-level analysis employed the participant-level contrast images using one-sample t-tests to examine overall task-related activation patterns. The statistical threshold for initial voxel inclusion was set at (p < 0.001, uncorrected), with a family-wise error (FWE) cluster-corrected p value. The resulting SPMs were overlaid on an MNI anatomical template, and anatomical regions were identified using the Harvard–Oxford cortical and subcortical atlases (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/). Activation maps were visualized using MRIcroGL software (http://www.mricro.com).

Region of interest (ROI) analysis: To investigate patterns of activity changes across learning stages, we identified ten ROIs based on peak activation clusters showing significant differences between the learning and control tasks. These ROIs were selected using a whole-brain analysis with an FWE correction (FWEc)-cluster-corrected threshold of p < 0.001.

We then used the Marsbar toolbox (http://marsbar.sourceforge.net/) (Brett et al. 2002) to extract beta values (parameter estimates) for each participant. The beta values were derived from a 5-mm sphere centered on the MNI coordinates of each peak activation cluster. This extraction process was conducted separately for each of the four stages of both the learning and control task blocks. This method enabled us to examine how neural activity within these key regions evolves across the different stages of learning and control tasks.

2.2.3 fMRI-Behavior Correlation

To investigate the relationship between behavioral performance and brain activity during the learning task, we conducted an ROI-based analysis guided by prior findings from a similar OLM study (Gillis et al. 2016). These ROIs were predefined, independent of our results, based on their established involvement in object–location associations, as per the validated ROI analysis framework (Poldrack 2007). In their comprehensive assessment of neuroimaging evidence supporting the cognitive model of OLM, they investigated the object–location binding process during new versus repeated object–location association (object + location novel > object + location repeated). The resulting activated regions included areas of the ventral visual stream (e.g., the fusiform gyri), the lateral occipital complex (LOC), and medial temporal structures (e.g., the hippocampus and parahippocampus).

In our study, we extracted beta values representing neural activation during the learning task for each subject from these predefined ROIs using the same learning contrast from the whole-brain first-level analysis. To assess the relationship between brain activation and performance accuracy, we performed a repeated-measures correlation analysis using the rmcorr package in R (https://cran.r-project.org/web/packages/rmcorr/). This method was selected because it effectively accounts for within-subject associations across repeated measures assessed on multiple occasions (Bakdash and Marusich 2017).

2.2.4 Test–Retest Reliability

To assess the reliability of both behavioral and fMRI results across two sessions with parallel task versions, we conducted an intraclass correlation coefficient (ICC) analysis for both behavioral performance and fMRI activation patterns (Caceres et al. 2009). The ICC is a statistical measure used to assess the reliability of measurements or ratings by evaluating the degree of consistency or agreement among multiple raters or measurements of the same subjects (Koo and Li 2016). The ICC was chosen as a suitable metric for TRR because it accounts for both within-subject and between-subject variability, offering a robust measure of consistency over repeated measurements (Weir 2005).

For behavioral accuracy, we calculated the ICC for the mean performance accuracy and latency in the learning task across the stages for each participant in both task sessions (Sessions 1 and 2). ICC estimates and their 95% confidence intervals were computed using the irr package in R, employing a two-way mixed-effects model with absolute agreement for single measurements (ICC[3,1]), as per McGraw and Wong (1996). This approach evaluates the reliability of individual scores by accounting for both systematic and random measurement variability, ensuring that absolute agreement between sessions is assessed.

For the consistency of fMRI activation, we used the Python-based prelimit tool to calculate the ICC based on whole-brain activation patterns derived from the beta values of the learning task activation (Demidenko et al. 2024). The ICC estimates were computed both inside a whole-brain mask.

ICC values were interpreted as follows: Values below 0.5 indicated poor reliability, 0.5–0.75 indicated moderate reliability, 0.75–0.9 indicated good reliability, and values above 0.9 indicated excellent reliability (Koo and Li 2016).

3.1.1 Response Accuracy

Behavioral data analysis revealed significant improvements in performance accuracy across stages, with distinct trends based on task type (learning vs. control). The GLMM showed a significant effect of stage (estimate = 0.211, SE = 0.054, 95% CI = [0.105, 0.316], z = 3.89, p < 0.001) and task type (estimate = −2.03, SE = 0.168, 95% CI = [−3.365, −1.706], z = −12.11, p < 0.001), as well as a significant interaction between stage and task type (estimate = 0.70, SE = 0.06, 95% CI = [0.578, 0.827], z = 11.22, p < 0.001), indicating that the effect of stage on accuracy differed between the learning and control tasks, with greater improvement in performance observed across stages in the learning condition.

Post hoc pairwise comparisons were conducted for the response accuracy model to explore differences between task types (control vs. learning) across the four stages using EMMs. In Stage 1, accuracy was significantly higher in the control task compared to the learning task (estimate = 1.33, SE = 0.12, z = 10.90, p < 0.001). In Stage 2, the control task continued to outperform the learning task, although the difference decreased (estimate = 0.63, SE = 0.097, z = 6.47, p < 0.001). By Stage 3, there was no significant difference between the two tasks (estimate = −0.073, SE = 0.11, z = −0.66, p = 0.51). In Stage 4, accuracy in the learning task surpassed that of the control task, with a significant negative difference (estimate = −0.77, SE = 0.15, z = −5.21, p < 0.001).

Other tested fixed effects, including session (p = 0.614) and block (p = 0.576), were not significant. However, the task version showed a small but significant effect, with participants achieving slightly higher scores during Task B than during Task A (estimate = 0.13, p = 0.013).

3.1.2 Reaction Time

The GLMM revealed significant effects of stage and task type on RT. RT decreased across stages (estimate = −0.050, SE = 0.004, t = −11.30, p < 0.001), suggesting faster responses across stages. Participants in the learning task had slower RTs than those in the control task (estimate = 0.109, SE = 0.014, t = 7.49, p < 0.001), likely due to the higher cognitive demands of the learning task. The interaction between stage and task type was also significant (estimate = −0.105, SE = 0.005, t = −2.01, p = 0.044). Session significantly impacted RT, with faster responses in Session 2 (estimate = −0.021, SE = 0.005, t = −4.23, p < 0.001). Task version (A vs. B) did not affect RT (estimate = 0.009, SE = 0.005, t = 1.85, p = 0.064).

Post hoc pairwise comparisons showed significant differences in RTs between tasks, with learning trials having longer RTs than control trials at all stages. In Stage 1, RTs were slower in the learning task (estimate = −0.099, SE = 0.010, z = −9.86, p < 0.0001), with a similar pattern in Stage 2 (estimate = −0.088, SE = 0.006, z = −13.56, p < 0.0001) and Stage 3 (estimate = −0.078, SE = 0.006, z = −12.45, p < 0.0001). By Stage 4, the difference decreased but remained significant (estimate = −0.067, SE = 0.009, z = −7.08, p < 0.0001).

3.2.1 Whole-Brain fMRI Analysis

For the contrast learning > control, several significant clusters of brain activity were identified. We applied a cluster-based FWEc with a threshold of p < 0.001 and included clusters larger than 125 voxels. Task-related activations were found mainly in the temporo-occipital fusiform cortex, hippocampus, precuneus, precentral gyrus, and lateral occipital cortex. The significant clusters and their associated MNI coordinates, T values, and voxel sizes are provided in Table 2. Figure 3 displays the SPM overlaid on the MNI brain template.

3.2.2 ROI Analysis

In most of the ROIs analyzed, as shown in Figure 4, brain activity followed a declining trend, with a gradual reduction in activity as participants progressed through the stages. This pattern suggests that these regions play a critical role in the initial learning and encoding of object locations. However, they become less active as participants consolidate and refine their memory of house locations. The right hippocampus exhibited an opposing pattern of increased activity across the stages of the learning task, along with an inconsistent pattern in the control task, which may support its later involvement in memory consolidation as learning progresses.

3.2.3 FMRI-Behavioral Correlation

Figure 5 summarizes the correlation analyses, showing the relationship between neural activation (beta values) and behavioral learning accuracy across the four stages. Significant negative correlations were observed in bilateral LOC, right LOC (r = −0.57, p < 0.001), and left LOC (r = −0.51, p < 0.001), as well as in bilateral fusiform gyri; right fusiform (r = −0.61, p < 0.001) and left fusiform gyrus (r = 0.64, p < 0.001). In line with their unique activation pattern across the later stages of learning, as shown in the earlier analysis, the left hippocampus also exhibited a significant positive correlation with performance (r = 0.50, p < 0.001). The right hippocampus also showed a positive trend (r = 0.23, p = 0.07), but it did not reach statistical significance. These findings are consistent with the previously mentioned behavioral and ROI analysis pattern.

3.3.1 Behavioral Results

To assess the TRR of participants’ performance on our OLM task across sessions, an ICC analysis was conducted (Leyland and Groenewegen 2014). A two-way model (single score with absolute agreement) was applied in R to evaluate the agreement of mean response accuracy scores across both measurement sessions.

The analysis revealed an ICC value of 0.801 (95% CI [0.737, 0.85]), indicating good reliability. This suggests that participants’ accuracy in the OLM task was highly consistent across sessions. The F-test for the ICC was significant (F (159,159) = 9, p < 0.001). For the RT, the resulting ICC was 0.705 (95% CI [0.613, 0.777]), indicating moderate reliability across both sessions. This suggests that participants’ RTs demonstrated reasonable consistency between sessions, though with slightly lower reliability than accuracy scores. The ICC was statistically significant, F (159,133) = 6, p < 0.001. These findings support the stability of participants’ performance over time and underscore the reliability of our OLM task.

3.3.2 fMRI Results

To assess the TRR of fMRI activations during the learning task, ICC maps were generated for task-related activity, contrasting learning with control tasks across two sessions. The resulting ICC map, presented in Figure 6, reveals a pattern of moderate to excellent reliability in several key regions associated with learning and memory processes. Bilateral regions, including the fusiform gyri, middle temporal gyri, and lateral occipital cortices, demonstrated moderate-to-good reliability across both sessions.

Notably, the right hemisphere exhibited particularly high ICC values in several areas. The right middle temporal gyrus showed excellent reliability across its anterior division (MNI: 42, −55, 13; ICC = 0.902), posterior division (MNI: 41, −55, 12; ICC = 0.910), and temporo-occipital part (MNI: 42, −54, 12; ICC = 0.918). Similarly, the right lateral occipital cortex demonstrated good-to-excellent reliability in both its superior division (MNI: 25, −76, 24; ICC = 0.870) and inferior division (MNI: 47, −65, 0.5; ICC = 0.922). The right temporal–occipital fusiform cortex also showed good reliability (MNI: 30, −52, −8; ICC = 0.845).

Consistent patterns of reliability were observed in the left hemisphere. The left lateral occipital cortex exhibited good ICC values (MNI: −41, −79, −2; ICC = 0.803) alongside other regions, including the left middle temporal gyrus temporo-occipital part (MNI: −40, −64, 1.9; ICC = 0.843), superior temporal gyrus (MNI: −60, −33, 6; ICC = 0.804), lateral occipital cortex (MNI: −39, −75, 0; ICC = 0.857), and temporal fusiform cortex (MNI: −20, −72, −6; ICC = 0.824). The precuneus cortex has also shown moderate reliability (MNI: −6, −56, 50; ICC = 0.655).

Conversely, lower ICC values, shown in blue on the maps, were observed in regions, such as bilateral frontal regions, ventral temporal cortices, and white matter. These areas, unrelated to task-specific demands, likely reflect noise or non-specific activation patterns. These findings support the experimental design and the reliability of fMRI activations elicited during the OLM task.