Twenty-seven healthy volunteers participated in this study. Seventeen subjects were allocated to the active neurofeedback group (10 male, mean age 27.8±4.2 years) and 10 subjects to the sham neurofeedback group (5 male, mean age 26.2±2.9 years). All had a normal or corrected-to-normal vision and no history of neurological or psychiatric diseases. All participants except one were right-handed. All gave informed written consent before participating, in accordance with the declaration of Helsinki, and the study complied with the safety guidelines for magnetic resonance imaging (MRI) research on humans. The work was approved by the Ethics Committee of the Faculty of Medicine of the University of Coimbra.

The experimental session was composed of an anatomical run—where structural brain information is acquired—followed by six functional runs—task-related runs where brain activation is inferred by measuring the blood oxygenation level-dependent (BOLD) signal (Figure 1). These runs consisted of a localizer run followed by five imagery runs. The localizer run was used to functionally map the DLPFC spatial mask used in the following runs as the neurofeedback target region. The first and the last imagery runs were performed without providing feedback information to the participant. The scanning session lasted approximately 1.5 h, followed by a debriefing questionnaire.

MRI data acquisition was conducted on a 3T Siemens Magnetom TrioTim scanner with a 12-channel head coil. For the anatomical runs we used a high-resolution magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence (176 slices; echo time (TE): 3.42 ms; repetition time (TR): 2,530 ms; voxel size: 1 × 1 × 1 mm; flip angle (FA): 7°; field of view (FOV): 256 × 256 mm). Functional imaging was acquired with an echo-planar imaging (EPI) sequence with 32 slices, in-plane resolution: 3 × 3 mm, FOV: 192 × 210 mm, matrix 64 × 70, slice thickness: 2.5 mm, FA: 75°, TR = 2,000 ms and TE = 30 ms. Functional runs included two “dummy scans” at the beginning of the acquisition (discarded and not stored) that allowed the magnetization to stabilize to a steady state.

During the neurofeedback runs, apart from the first baseline block, visual feedback was provided in the form of a thermometer that was updated every TR based on the mean ROI activation of the neurofeedback target selected during the localizer run. In order to correct for head movements across runs, we performed an intra-session alignment (six degrees of freedom), using the first volume of the localizer as a reference for all runs.

The thermometer was divided into 10 discrete levels with a maximum value of 2.5%, where each level represented a given range of percent BOLD signal change (0 for an empty thermometer and 0.25% for each level). The feedback value fb for the current time point n is calculated within each block given the current value val, a baseline level bl (mean BOLD value in the target region, during the previous “baseline” block) according to equation 1:

Offline fMRI data analysis was performed using BrainVoyager QX 2.8 (Brain Innovation, Maastricht, The Netherlands). Preprocessing steps included slice scan time correction, 3D motion correction (6 degrees of freedom), temporal high-pass filtering (GLM Fourier method, two cycles, i.e., a GLM with predictors that accommodate sine and cosine functions with two cycles over the entire time-course of the run), spatial smoothing using a 3D gaussian kernel (FWHM = 6 mm), and normalization to Talairach (TAL) coordinate space.

First-level analysis was performed using a GLM for each run. The design matrix included a predictor for each experimental condition convolved with the BrainVoyager’s default two-gamma HRF, and confound predictors for the six motion parameters (three translational and three rotational) and motion spikes (volumes with a root mean square displacement greater than 0.25).

Second-level analysis was based on a random effects (RFX) GLM, correcting for multiple comparisons with false discovery rate (FDR; q = 0.005).

We examined the event-related average (ERA) responses of functionally relevant ROIs extracted from the previous analyses. To this end, the mean BOLD signal time-courses of the DLPFC and ventral striatum were converted into PSC (signal variation relative to the average BOLD value during the “baseline” condition). Then, the time-course was segmented based on the onset and offset of each “Imagery” condition block (we considered two volumes before the onset and five after the end of each block to better understand the temporal profile of the response). Finally, we averaged these segments across trials and participants for each group, allowing for the between group comparison of the response of each brain region.

In the debriefing questionnaire, all the subjects in the active neurofeedback group perceived a correspondence between the given feedback and the imagery task. Some of these felt that this correspondence was independent of the strategies they used. Most participants followed the suggested imagery task for activating the DLPFC (inverted recall of self-generated numeric sequences), although some reported to rely more on sequence visualization and others on mental calculation. The reported number of digits varied from 4 to 20 and the number of sequences from 2 to 12, per block.

In the sham neurofeedback group, eight subjects (80%) reported no apparent association between feedback and imagery tasks. These participants tried different strategies, such as recalling numbers in a different language, repeating backwards the name of family members, and mentally playing an instrument. The reported number of digits varied from 3 to 16 and the number of sequences from 1 to 15, per block.

The number of digits and sequences did not differ significantly between groups (independent sample t-test, p = 0.496 for digits and p = 0.784 for sequences).

A table with participants’ answers to the debriefing questionnaire is provided as Supplementary Table 2.

The group activation maps showed the expected recruitment of regions of the FPN (including DLPFC and PPC), both during the localizer and imagery runs. When comparing these two tasks, the imagery results are left lateralized, as confirmed by the calculated laterality index (0, 26). This is consistent with the previous knowledge regarding the neural correlates of verbal working memory tasks. N-back task (localizer) elicits bilateral activation of lateral pre-frontal and parietal regions with lateralization possibly varying according to age and task difficulty (Owen et al., 2005; Rottschy et al., 2012; Emch et al., 2019). Backwards reciting task has a correspondent left lateralized functional map (Smith et al., 1996), regardless if the self-generated sequence is numerical or alphabetical, and with backward recitation requiring additional neural resources than forward recitation, mainly in parietal areas (Zhou et al., 2006), both used in our paradigm.

Additionally, functional maps showed (1) positive clusters in basal ganglia and thalamus, also known to be involved in working memory (Lewis et al., 2004); and (2) negative clusters in DMN representative of the well-known anti-synergic coupling (with negative correlation) of FPN-DMN. Both the interaction between FPN and DMN and the cortico-subcortical connectivity has already been shown, by Zhang’s group, to be modified by rt-fMRI working memory neurofeedback, with consequent performance modulation on post-training (Shen et al., 2015; Zhang G. et al., 2015, Zhang Q. et al., 2015; Zhang et al., 2016).

All participants from the active group and the majority from the sham group were able to increase DLPFC activity already in the initial training, before neurofeedback runs, indicating that the proposed imagery task, in a population without executive dysfunctions, is robust enough to recruit and maintain DLPFC activity without any type of feedback.

However, we also expected that the active NF group would be able to achieve greater control over signal modulation in the DLPFC when compared to the sham-control group, given the correspondence between feedback and mental strategy (Sorger et al., 2019). Indeed, our results show a significant difference in DLPFC activity during the neurofeedback runs, being higher in the active NF group. On the contrary, such differences were not found during the runs without feedback (train and transfer), denoting that closing the neurofeedback loop with valuable information increases the mean target ROI activation, a result that is consistent with improved self-modulation ability. Behaviorally, this was also implicit in the fact that only participants from the sham group tried different imagery strategies, while in the active group all maintained the provided suggestions, despite equal instructions for both groups. This is a reflection of the associative learning framework supporting neurofeedback, since reinforcement of the used mental strategy occurs when feedback acts as a reward.

In addition to previous neurofeedback studies engaging DLPFC, we demonstrate that active and sham neurofeedback groups differ not only in the modulation amplitude but also in the way DLPFC activity progresses along the NF run. Active neurofeedback participants were able to sustain higher levels of BOLD activity along all the run, while in the sham group it drops to null values after nine volumes from the imagination task onset. Again, this is probably due to the acknowledgement, as the run proceeds, of the non-contingency between their effort and the represented signal change.

Another study using DLPFC as a neurofeedback target, found differences in the dynamical regulation of physiological DLPFC activity after neurofeedback training, although without a modification of target activation levels (van den Boom et al., 2018). In this case, there was a reduction in time needed to return to baseline, suggesting that is possible to deactivate DLPFC in a deliberate way. They also stated that control of the elevation phase is more difficult to achieve with the neurofeedback since the execution of a WM task immediately increases to maximum the BOLD signal. In our data, during feedback runs, we also find an activity peak approximately at 6–8 s both for active and sham groups and the main difference arises after, on the ability to sustain it and not returning to baseline. However, van den Boom et al. (2018) only presented their data on the post-test, so we cannot compare the dynamics of DLPFC BOLD activity during feedback from this study with our own. Another important difference, since we are arguing on the contingency relevance, is that in their study sham group received feedback from another participant, while in our protocol feedback it was derived from white matter voxels.

In line with our hypothesis that was mainly the motivation differentiating both groups, when we performed an exploratory whole-brain analysis we found a single cluster in the ventral striatum (in particular, nucleus accumbens) with higher activation in NF group compared to sham group on NF runs. The ventral striatum has been pointed as responsible for unconscious reward processing in NF, with anterior cingulate cortex and anterior insular cortex being involved in the conscious counterpart of reward and NF perception, as reviewed by Sitaram et al. (2017). The nucleus accumbens, in particular, is recognized as key integrative region of both motivational and learning-memory circuits, receiving input from cortical areas (including DLPFC), limbic system, and midbrain (substantia nigra/ventral tegmental area) and, in turn, selecting appropriate responses (Camara et al., 2009).

The event-related time course analysis performed on this cluster during NF runs showed a positive response at the beginning of each block for the active NF group contrasting with a negative initial response and during almost all the block for the sham group during NF runs. In the active NF group, the BOLD signal follows the predicted pattern, peaking at approximately 6 s and then slowly returning to baseline, according to previous reports that nucleus accumbens activity is time-limited (i.e., not sustained) by dopamine metabolism (Knutson and Gibbs, 2007; Greer et al., 2014). In the sham group, although there is an initial overshooting (lower than the active group), it is followed by a negative response all along the block. Interestingly, this relationship inverts on down-regulation (baseline), where participants in the sham group may percept an apparent causality between their effort to zero the thermometer and a BOLD sign that actually does not represent cortical activity (with lower amplitude). When neurofeedback is taken from the equation (train and transfer runs) the difference between groups on time course analysis dissolves, as expected. Furthermore, subjectively, all the participants in the active group perceived a correlation between feedback and performance—against only 20% of the sham group, with subjects reporting frustration along NF runs.

Conversely, striatal and midbrain structures included in the reward network have been tested as the direct target in neurofeedback experiments, searching for therapeutic effects on diseases affecting the mesolimbic dopamine system. These studies offer a mirrored perspective which complements our results, showing how these structures contribute to successful learning in neurofeedback and how they are linked to other brain regions. The experiment of MacInnes et al. (2016) is relevant to understanding temporal BOLD signal dynamics during neurofeedback and how directly targeting the mesolimbic dopamine system contributes to learning volitional cognitive strategies. They also found sustained activation in the target region (ventral tegmental area—VTA) during all the 20 s of the trial as a proof of improvement, but which was only present in the post-test run, reflecting a learning effect (which was not present in our study). During neurofeedback runs (compared to pre-test) they showed higher connectivity VTA-caudate, VTA-hippocampus and NAcc-caudate, also emphasizing the role of NAcc in neurofeedback training targeting the reward network. A recent article (Hellrung et al., 2022) similarly targeting VTA/SN provides a more profound investigation of which specific neural mechanisms are related to the transfer success when training VTA, showing that the most successful individuals had stronger activation of cognitive control areas, mainly the prefrontal cortex, during transfer. That is, higher individual reward-related sensitivity in the DLPFC increases the chance of neurofeedback training success. Links between these networks are bidirectional and it has been proved that dopamine action in DLPFC sustains working memory performance (Arnsten et al., 2015). In conclusion, associative learning crucially contributes to real-time fMRI neurofeedback effects, as also suggested by our results.

Previous neurofeedback studies targeting DLPFC for working memory enhancement (Zhang et al., 2013; Sherwood et al., 2016a,b; van den Boom et al., 2018), and also our own, show that DLPFC readily activates with the explicitly suggested imagery task, with improved self-modulation ability provided by neurofeedback. However, in contrast to our study, none of the previous studies investigate motivational effects or the participant perception of feedback contingency and how it may influence NF outcomes. Directly linked to this issue, the choice of an adequate control condition is being highly debated in the neurofeedback community and also varied across studies. It is suggested that an ideal control must minimize the likelihood of placebo effects, while maintaining neurophysiological specificity and promote equal motivation/perception of success (Sorger et al., 2019). However, it is particularly challenging when targeting such a key core network and hub (as the executive network and in particular the DLPFC, which are efficiently activated by a working memory imagery task and by neurofeedback processing itself), to find a control ROI regulated as easily, with similar signal properties, while not being related to it. That is, both the sensitivity and the specificity of the control region are key aspects to consider (Sorger et al., 2019).

In our study, we established a link between motivational aspects and DLPFC self-modulation through rt-fMRI neurofeedback. Notably, although sham participants could not identify a correspondence between used strategies and given feedback, all except one were able to provide an answer to the last two questions of the debriefing, discriminating better and worse strategies. Also, the number of digits and sequences did not differ significantly between groups, suggesting that both were equally engaged on the imagery task. Another critical point is that none of the participants was aware of a control condition, that is, they did not know that receiving false feedback was a possibility in this type of experiment (as they were naïve to neurofeedback experiments). In this sense, our results suggest that the perception of non-contingency did not affect effort, perseverance on the task, or the belief that the feedback was real, but had a motivational impact, which is supported by fMRI data.

Perception of contingency is vital for the operant conditioning that is a foundational principle of neurofeedback and we find motivational aspects difficult to be removed from the equation, at least in neurofeedback tasks targeting executive functions. Critically, they must be considered when reporting and interpreting neurofeedback results, having implications on possible adjustments to achieve optimized experimental design and, ultimately, in translation to clinical implementation.

We were not able to prove a learning effect from our neurofeedback training, with both groups equally performing in the transfer run. This was anticipated when we designed a single-session neurofeedback experiment, being our main goal to study the general capacity of individuals to regulate DLPFC (feasibility test). Although the optimal number of neurofeedback sessions is not yet established in the literature (possibly varying according to target and population), it is consensual that a single session will hardly induce changes in neural processes (Thibault et al., 2018; Paret et al., 2019; Fede et al., 2020). The fact that our participants were already able to modulate DLPFC activity in train run, alludes to a possible initial ceiling effect, precluding detection of within session learning and perhaps contributing to justify the equal performance on transfer. Furthermore, we did not control for fatigue or habituation, which may also interfere in the results of the final run.

Another limitation is that our paradigm was not specifically designed to unscramble task-dependent DLPFC BOLD-magnitude from the engagement on neurofeedback process itself, namely in the preparation and execution of mental strategies (Skottnik et al., 2019). Sherwood et al. (2016b) showed that left DLPFC activity evolves interchangeably in closed and open-loop neuromodulation across training. However, we argue that the cognitive control globally linked to neurofeedback training is equally present for sham feedback, as previously reported by Ninaus et al. (2013), so that the differences we found in DLPFC were related to the imagery task.

Here, we did not correct for physiological artifacts in feedback signal computation. As far as we know, there is no gold standard for the use of derivatives of cardiac and respiratory signals as confounds in real-time. However, the impact of these regressors has been studied in recent experiments performing connectivity-based neurofeedback (Weiss et al., 2020). With similar objectives, the use of a control region has been applied to account for global fluctuations (Dewiputri and Auer, 2013).

Finally, our results on reward network during neurofeedback training derived from a whole-brain exploratory analysis, demanding further corroboration in a larger sample and with different target areas. Here, we did not present a sampling plan. Since our study is a proof-of-concept, the sample size rationale was based on the literature of previous NF studies targeting DLPFC for executive functions enhancement (Zhang et al., 2013; Sherwood et al., 2016a,b; van den Boom et al., 2018).

The present results suggest that neurofeedback promotes not only greater signal modulation of DLPFC, but also a more sustained activity of this target region along the neurofeedback runs. This ability seems associated with motivation, since: (1) differences dissolved when neurofeedback is removed; and (2) when comparing functional maps from active NF and sham group, we found a single cluster located on the ventral striatum, responsible for unconscious reward processing.

DLPFC was easily activated by all subjects in most of the runs (with or without feedback) and is clearly modulated by neurofeedback, with the critical involvement of the reward network. It is also a large and superficial brain area, easily accessible by other technics (such as EEG, MEG or fNIRS), which makes it an ideal target for training transfer. Its clinical potential is reinforced by the important role that DLPFC plays in many high-level cognitive functions. Thus, alterations to its normative functioning are linked to the cognitive impairment found in a number of neuropsychiatric disorders, which ultimately largely impacts individual functional capacity and independence. Taking all together, DLPFC emerges as a clinically-relevant neurorehabilitation and/or neuroenhancement target, that deserves future translational research in the neurofeedback field.