Neuroanatomical predictors of real-time fMRI-based anterior insula regulation. A supervised machine learning study.

Increasing evidence showed that learned control of metabolic activity in selected brain regions can support emotion regulation. Notably, a number of studies demonstrated that neurofeedback-based regulation of fMRI activity in several emotion-related areas leads to modi�cations of emotional behavior along with changes of neural activity in local and distributed networks, in both healthy individuals and individuals with emotional disorders. However, the current understanding of the neural mechanisms underlying self-regulation of the emotional brain, as well as their relationship with other emotion regulation strategies, is still limited. In this study, we attempted to delineate neuroanatomical regions mediating real-time fMRI-based emotion regulation by exploring whole brain GM and WM features predictive of self-regulation of anterior insula (AI) activity, a neuromodulation procedure that can successfully support emotional brain regulation in healthy individuals and patients. To this aim, we employed a multivariate kernel ridge regression model to assess brain volumetric features, at regional and network level, predictive of real-time fMRI-based AI regulation. Our results showed that several GM regions including fronto-occipital and medial temporal areas and the basal ganglia as well as WM regions including the fronto-occipital fasciculus, tapetum and fornix signi�cantly predicted learned AI regulation. Remarkably, we observed a substantial contribution of the cerebellum in relation to both the most effective regulation run and average neurofeedback performance. Overall, our �ndings highlighted speci�c neurostructural features contributing to individual differences of AI-guided emotion regulation. Notably, such neuroanatomical topography partially overlaps with the neurofunctional network associated with cognitive emotion regulation strategies, suggesting common neural mechanisms.

However, several neuropsychological and methodological aspects related for instance to ROI or brain network selection, feedback modality and calculation, participants' instructions and strategies, control conditions and control groups as well as neurofeedback learning assessment, still remain unclear and sometimes controversial Thibault, MacPherson, Lifshitz, Roth, & Raz, 2018). In addition, although some hypotheses and empirical indications of the neural processes leading to changes at neural and behavioral level were recently described (Emmert, et al., 2016;Gaume, Vialatte, Mora-Sanchez, Ramdani, & Vialatte, 2016;Shibata, et al., 2019;Sitaram, et al., 2017), the current understanding of the neural mechanisms underlying learned brain regulation, in particular in relation to emotional brain regulation, is still limited. A previous meta-analysis of studies focusing on regulation of different brain areas delineated a possible "regulation network" that included the lateral prefrontal cortex, anterior insula, basal ganglia, temporo-parietal areas, anterior cingulate and visual associative areas (Emmert, et al., 2016). However, in this meta-analysis numerous studies that were included have targeted areas directly or indirectly involved in emotional behavior, making it di cult to conclude on the speci c role of the reported "regulation network".
A recent study aiming to identify neurofunctional predictors of successful neurofeedback-based self-regulation, on the basis of pretraining functional data in a cohort of heterogenous studies targeting several different areas in healthy participants as well as neurological and psychiatric patients and individuals with psychological disorders (Haugg, et al., 2020), did not result in common brain-based neurofeedback success predictors. This result might be ascribed to the heterogeneity of neural processes mediating learned brain regulation within distinct functional domains. On the other hand, Zhao and colleagues exploring the potential associations between brain structure and real-time fMRI neurofeedback learning showed that the gray matter (GM) volume of the right putamen could signi cantly predict learning success across the three different data sets including distinct experimental designs and targeted training regions (Zhao, et al., 2021). This evidence con rmed previous indications pointing to the striatum, and its well-recognized role in instrumental and associative learning, as key node for real-time based brain regulation (Sitaram, et al., 2017). In their analysis, Zhao and colleagues included previous real-time fMRI studies targeting the anterior insula, amygdala and anterior cingulate cortex, all regions relevant to emotional behavior, however, no additional regions implicated in general emotion regulation emerged (Etkin, Buchel, & Gross, 2015;Grecucci, Giorgetta, Van't Wout, Bonini, & Sanfey, 2013;Messina, Grecucci, & Viviani, 2021). On the other hand, we previously showed that self-regulation of AI is functionally mediated by regions in the cortical emotion regulatory network (Etkin, Buchel, & Gross, 2015), and also relies on core emotional and motivational centers in the upper mesencephalon (Caria, 2020). These results also corroborated indications of possible similarities between cognitive emotion regulation and self-regulation of emotional brain regions (Caria, 2020;Emmert, et al., 2016;Etkin, Buchel, & Gross, 2015;Gross, 2014;Paret & Hendler, 2020).
Here, in an attempt to delineate speci c neuroanatomical predictors of AI-guided emotion regulation, we used a kernel ridge regression model (Kong, et al., 2019) to assess whether regional gray and white matter (GM, WM) volumetric characteristics signi cantly impacted self-regulation of AI activity, an effective neuromodulation procedure that has been shown to elicit changes in emotional stimuli perception in both healthy individuals and patients with emotional disorders (Caria, Sitaram, Veit, Begliomini, & Birbaumer, 2010;Caria, et al., 2007;Linden, et al., 2012;Ruiz, et al., 2013;Sitaram, et al., 2014;Yao, et al., 2016;Zilverstand, Sorger, Sarkheil, & Goebel, 2015). In our study, exploiting data from two previous studies (Caria, Sitaram, Veit, Begliomini, & Birbaumer, 2010;Caria, et al., 2007), we separately estimated whether voxel-based (whole brain) and network-based GM volumetric information as well as voxel-based (whole brain) WM volumetric characteristics could signi cantly predict interindividual differences of AI regulation. In line with previous functional studies indicating a main involvement of regions in the central executive network (CEN) and salience network (SN) during both real-time fMRI-based regulation and emotion regulation (Buhle, et al., 2014;Caria, 2020;Emmert, et al., 2016;Etkin, Buchel, & Gross, 2015;Kohn, et al., 2014;Ochsner, Silvers, & Buhle, 2012) we expected that structural characteristics of some of these networks nodes, such as prefrontal circuits and basal ganglia, signi cantly contributed to successful AI regulation.
The real-time fMRI paradigm consisted of emotion regulation runs guided by online feedback of fMRI signal in the AI. Participants were provided with online continuous feedback through a visual display consisting of a graduated thermometer depicting changes of BOLD response with increasing or decreasing number of bars updated every 1.5s. Four emotion regulation runs composed of several blocks (16 blocks for n = 9 participants of Study 1; 20 blocks for n = 9 participants of Study 2) were performed in one day. Each run consisted of up regulation blocks (4 blocks of 22.5s each for n = 9 participants of Study 1; 4 5 blocks of 30s each for n = 9 participants of Study 2), cued with an arrow at the right side of the thermometer, alternating with down regulation blocks (4 blocks of 22.5s each for n = 9 participants of Study 1; 5 blocks of 30s each for n = 9 participants of Study2), cued with a cross hair at the right side of the thermometer. Participants were instructed that during up and down regulation blocks they had to attempt to respectively intensify or reduce the intensity of recalled emotional memories and imagery of personally relevant affective episodes guided by increasing or decreasing number of thermometer bars. The feedback represented the averaged BOLD signal in the AI normalized with respect to a reference region, calculated during up regulation with respect to down regulation. Three consecutive TRs were considered to reduce rapid signal uctuations; the rst ten volumes of each session were excluded to account for T1 equilibration effects. The target right AI (n = 9, from Study 1) was selected anatomically based on the high resolution T1 structural scan and consisted of a rectangular area of 4 x 5 voxels on a single slice. The target left AI (n = 9, from Study 2) was selected anatomically based on the high resolution T1 structural scan and functionally, through a localizer session consisting of ve alternating emotional recall and baseline blocks, and consisted of a rectangular area of 5 x 5 voxels on a single slice positioned manually around the peak of maximum beta estimate of the AI cluster. In both studies, a similar reference region of interest (ROI) consisting in large background area not encompassing emotion related areas was used to cancel out global effects and unspeci c activations. Participants were informed that the feedback information was delayed of about 1.5s due to online data analysis in addition to a physiological latency of the hemodynamic response of about 6s. Participants were instructed not to move during all the experimental conditions, and informed that physiological signals were monitored. During real-time fMRI training functional images were exported online from the MR console computer to a separate computer for real-time preprocessing and analysis with Turbo brain voyager (Brain Innovation, Maastricht, The Netherlands). Online preprocessing included incremental 3D motion correction and drift correction. Incremental statistical data analysis was based on recursive least squares General Linear Model (GLM).
O ine fMRI data analysis fMRI data preprocessing and analysis were performed using SPM12 (Wellcome Trust Centre for Neuroimaging, London, UK) and Matlab (The MathWorks, Inc., Natick, MA, US). For each participant, all functional images were rst realigned to the mean image using least squares and a 6 parameter (translations and rotations in space) and including resampling using 2nd degree B-spline interpolation, then unwarped and corrected for geometric distortions using the eldmap of each participant. The high-resolution T1 image was co-registered to the mean image of the EPI series using a rigid body model, estimated with mutual information. Segmentation parameters were used to normalize the functional images to the Montreal Neurological Institute (MNI) space. Last, normalized images were spatially smoothed with a 6 mm FWHM Gaussian kernel in order to balance effect size, spatial accuracy and statistical signi cance estimated using Gaussian random elds (Stelzer, Lohmann, Mueller, Buschmann, & Turner, 2014;Worsley & Friston, 1995).
A xed-effects general linear model (GLM) was used to perform rst-level statistical analysis. Hemodynamic response amplitudes were estimated using standard regressors, constructed by convolving a boxcar function, for up and down emotion regulation, with a canonical hemodynamic response function using standard SPM12 parameters. The time series in each voxel were high-pass ltered at 1/128s to remove low frequency drifts. An autoregressive AR(1) model was employed to address autocorrelation in the timeseries. Contrast images of up emotion regulation versus down emotion regulation were created for each block and run on each subject. Movement parameters were also included into the GLM as covariates to account for head motion artifacts. Estimation of BOLD signal changes in the AI during real-time fMRI training was performed using regions-of-interest (ROIs) analysis (Cremers, Wager, & Yarkoni, 2017;Poldrack, 2007) in the targeted ROIs (see further details in (Caria, Sitaram, Veit, Begliomini, & Birbaumer, 2010;Caria, et al., 2007)). AI regulation during real-time fMRI training was assessed by testing single subject's BOLD signal change across runs for the contrast up > down emotion regulation using a one-way repeated measures ANOVA with RUN as within factor. Post hoc paired samples t-tests and bootstrap analysis (1000 bootstrap samples, 95% bias corrected and accelerated con dence interval), as implemented in SPSS statistics software (v.24, IBM Corp. Armonk, NY), were used to test speci c differences between runs.
As indication of successful real-time fMRI-based regulation we considered the BOLD amplitude difference between last and rst run, and a learning index, calculated as the slope of the linear curve that tted % BOLD amplitudes during each run. These learning-related values, along with the averaged BOLD activity of all runs, were used in the subsequent multivariate pattern regression analysis. The averaged activity was expected to be mainly representative of the whole self-regulation session, and possibly also re ecting non-linear learning. Data from run 4 were not included in the analyses as not every participant underwent all four runs (14 out of 18).
Anatomical-based pattern regression analysis GM and WM MR data were used in a regression-based multivariate pattern analysis (MVPA), as implemented in PRoNTo toolbox (version 2.1, http://www.mlnl.cs.ucl.ac.uk/ pronto, (Schrouff, et al., 2013), to investigate volumetric patterns of white and gray matter predictive of real-time fMRI-based AI regulation (up vs down regulation) , considering % BOLD amplitudes during each run, average BOLD activity of all runs as well as the additional learning-related values. Quality check of the initial images was rst performed to exclude evident artefacts. The anatomical scan of one participant had to be excluded because of poor image quality. Images were segmented into GM and WM partitions using the segmentation procedure (Ashburner & Friston, 2005) as implemented in SPM12 (Statistical Parametric Mapping software, version 12; http://www. l.ion.ucl.ac.uk/spm; Wellcome Department of Imaging Neuroscience, London), normalized to MNI space and spatially smoothed (full-width at half maximum of Gaussian smoothing kernel [8,8,8]). Multivariate regression analysis was performed using Kernel ridge regression (KRR), a valid method enabling high predictive accuracy (Chu, Ni, Tan, Saunders, & Ashburner, 2011;Orsenigo & Vercellis, 2012), even in case of relatively small samples and single-subjects (Dadomo, et al., 2022;Schumacher, Halai, & Lambon Ralph, 2019). KRR is a multivariate regression method equivalent to maximum a posteriori approach to Gaussian process regression with xed prior variance and no explicit noise term. KRR is the dual-form formulation of ridge regression that enable to solve regression problems with high dimensional data in a computationally e cient way. In the feature extraction step, voxel-based values representing local GM and WM volume were extracted from raw brain scans as vectors, and then transformed into similarity matrices (kernels) to avoid the curse of dimensionality problem and to simplify calculations (Mourao-Miranda, et al., 2011). The KRR was trained to predict the percentage of BOLD change in the AI (target labels) from the GM and WM kernels.
To allow for generalization to new cases a Leave-One-Subject-Out (LOSO) cross validation method was used (1 structural image x subject). The dataset was partitioned into disjoint 'training' (n subjects -1 =16 subjects) and 'test' sets (the remaining n=1 subject) and analysis followed two recursive steps. During the training phase the KRR learned mapping between patterns and labels on the training set. During the test phase the learned function tried to predict the labels from the test set (Schrouff et al., 2013;Yang et al., 2016). This step was repeated for each of the 16 folds. Across all folds, predictive accuracy was estimated on the basis of Pearson's correlation coe cient, coe cient of determination (R2), and normalized mean squared error (nMSE) between predicted and actual real-time fMRI training effectiveness (Yang, et al., 2016). Hyperparameters were optimized as suggested by PRONTO developers with soft-margin C ranging from 0.0001, 0.01, 1, 10, 100, 1000, to compute the inner loop and the outer loop (model performance). Statistical signi cance of the classi cations was estimated using 5000 permutations with random assignment of group class to input image. For each iteration, the regression targets were randomly permuted across all participants and cross-validation procedure was repeated. Additional data operations included regressing out age and gender. GM regions were identi ed according to the Automated Anatomical Labeling (AAL) atlas (available on the WFU-PickUp Atlas toolbox of SPM12 (Tzourio-Mazoyer, et al., 2002), a manual macroanatomical parcellation of single subject MNItemplate brain consisting of 116 brain regions and additionally including brainstem regions. We also estimated predictability of AI regulation on the basis of well recognized macro-networks such as the default-mode network (DMN), central executive network (CEN), salience network (SAL), visual network (VIS) and sensorimotor network (SMN) (Damoiseaux, et al., 2006;Doucet, Lee, & Frangou, 2019;Smith, et al., 2009) (brain networks available in PRoNTo toolbox), by applying second level masks in the feature selection step. Network based analysis aimed to assess the overall involvement of speci c networks know to be involved in real-time fMRI-based emotion and cognitive emotion regulation (Buhle, et al., 2014;Caria, 2020;Emmert, et al., 2016;Etkin, Buchel, & Gross, 2015;Kohn, et al., 2014;Ochsner, Silvers, & Buhle, 2012). Furthermore, white matter regions predictive of AI regulation were assessed on the basis of the ICBM DTI-81 stereotaxic probabilistic white matter atlas, that fuses DTI-based white matter information with the anatomical template ICBM-152 Mori, et al., 2008). Finally, the weight maps for the KRR model showing statistically signi cant values of correlation and normalized MSE were obtained. The weight map spatially represents the model's weights in % by showing the contribution of each voxel in the image (Schrouff, et al., 2018;Schrouff, et al., 2013). The weight map of linear machine learning models is the average map across the cross-validation folds divided by its Euclidean norm (Schrouff, et al., 2013). In addition, the normalized weight for each brain structure was calculated as the average of absolute values of all voxel weights within each region de ned by the AAL divided by the number of voxels within the region. Labelled regions were then ranked according to the percentage of the total normalized weights they explained.

Results
Self-regulation of AI activity As reported in our previous studies (Caria, Sitaram, Veit, Begliomini, & Birbaumer, 2010;Caria, et al., 2007), most of participants in the two groups achieved successful regulation of AI activity during real-time fMRI training as evidenced by signi cant differences between rst and last run in each group separately. This effect was con rmed when both groups were merged together as shown by repeated measures ANOVA (F 2 = 3.63 p = 0.038, with a signi cant linear trend F 1 = 5.02 p = 0.040) (Figure 1) and by post hoc analysis revealing a signi cant difference between rst and last run (run3) (paired-samples t test, t 17 = 2.24 p = 0.040) and between rst and second run (t 17 = 2.34 p = 0.030). As previously reported, successful regulation was attained through a combination of real-time fMRI feedback and emotion-related mental strategies, and differed from the effects observed in two control groups tested with either unspeci c real-time fMRI feedback or mental imagery alone(see (Caria, Sitaram, Veit, Begliomini, & Birbaumer, 2010;Caria, et al., 2007) for more complete information).

Pattern regression analysis
Whole brain KRR-based multivariate pattern analysis permitted to identify regional GM and WM volumetric features predictive of individual differences in self-regulation of AI activity during real-time fMRI-based emotion regulation.

GM whole brain results
In relation to Run1, multivariate regression analysis approached signi cance (r = 0.39 p = 0.051, R2 = 0.15 p = 0.257 nMSE = 0.10 p = 0.055); GM regions are reported in Table 1. No signi cant predictive regions were observed for Run2 (r = -0.16 p = 0.544, R2 = 0.02 p = 0.693 nMSE = 0.22 p = 0.841). A number of GM regions were instead signi cantly predictive of AI regulation during Run3 (r = 0.47 p = 0.028, R2 = 0.22 p = 0.257 nMSE = 0.07 p = 0.025), when all subjects successfully learned to increased AI activity, see Table 1. In this condition, a less distributed anatomical network with respect to Run1 involved the cerebellum and medial temporal regions, and included dorsomedial and dorsolateral prefrontal cortex. In addition, several GM regions successfully predicted the averaged AI activity of all runs (r = 0.49 p = 0.049, R2 = 0.24 p = 0.168 nMSE = 0.07 p = 0.049) -see Table 1 and Figure 2 -composing a large network that included portions of the cerebellum, frontal and prefrontal cortex such as the dorsomedial and dorsolateral areas, the right inferior and superior temporal gyrus and supramarginal gyrus, the occital cortex, hipoccampus and parahippocampal gyrus, and the globus pallidus. No signi cant results emerged when either the learning index or the difference between the last and rst run was considered.

GM network-based results
Multivariate regression analysis based on selected brain showed a predictive effect of VIS, CEN and DMN (r = 0.41 p = 0.038, R2 = 0.17 p = 0.298 nMSE = 0.10 p = 0.047) only for Run1 (see Table 2 Table 2). No signi cant results emerged when either the learning index or the difference between the last and rst run was considered. Table 2. GM brain networks predictive of learned AI regulation Normalized weight and number of voxels of GM brain networks, obtained by encompassing all regional-based predictive values in each network node.  Table 3 and Figure 3. The tapetum, fronto-occipital fasciculus, fornix and posterior thalamic radiation were among the regions showing the highest predictive value as indicated by the normalized weight. No signi cant results again emerged when either the learning index or the difference between the last and rst run was considered.

Discussion
The main objective of this study was to identify neuroanatomical features predictive of AI-guided emotion regulation. To this aim, we used a KRR model to estimate the predictive value of regional GM and WM volumetric features for learned regulation of BOLD activity in the AI. Our ndings revealed that volumetric differences of speci c neuroanatomical structures are signi cantly associated with interindividual differences in real-time fMRI based emotion regulation. Multivariate regression analysis showed that GM and WM volumetric characteristics of several cortical and subcortical areas successfully predicted regulation of AI activity in relation to both the most successful regulation run and average neurofeedback performance. Speci cally, numerous GM regions including areas in the prefrontal cortex, medial temporal cortex, lateral occipital cortex, globus pallidus, hippocampus, parahippocampal gyrus and the cerebellum, as well as WM regions including the fronto-occipital fasciculus, tapetum and fornix, were successful predictors of learned regulation of AI activity.
In addition, GM network-based analysis also showed signi cant predictive values of selected brain networks, such as the DMN, CEN and SN, for the averaged AI activity across all regulation runs. This latter result along with the overall small predictive value of single GM and WM regions indicated that successful prediction mainly relied on a collective effect of speci c anatomical networks.
The nature of neuroanatomical characteristics affecting self-regulation of brain activity is likely quite heterogenous, implying either increased or reduced volumetric values. Multivariate regression analysis is sensitive to any systematic volumetric difference contributing to prediction of behavioral data, independently of directionality. Consequently, our results are more likely re ecting the effect of both increased and reduced GM and WM volumetric characteristics that jointly impact AI self-regulation.
Previous studies exploring structural and functional predictors of learned regulation did not report consistent ndings (Haugg, et al., 2021;Zhao, et al., 2021). Haugg and colleagues, across a large cohort of different studies in both patients and healthy individuals, on the basis of ROI activity prior to neurofeedback training, observed no common brain functional predictors of successful regulation, where learning success was de ned either based on the slope of the regression line over brain-based successful regulation all neurofeedback runs or based on the difference between the last and the rst neurofeedback run (Haugg, et al., 2021). However, they measured a small positive correlation between pretraining activity during functional localizer and the difference between last and the rst neurofeedback run, suggesting the relevance of pre-activation of targeted ROI for consecutive neurofeedback learning success. On the other hand, Zhao and colleagues observed that among the few selected regions of interest the gray matter volume of the right putamen could successfully predict learning success as de ned by the difference of targeted brain activity between late and early stages of training (Zhao, et al., 2021). Our whole brain multivariate ML-based analysis did not show signi cant brain structures predictive of individual differences in neurofeedback learning-success, either considering the slope of the regression line over regulation runs or based on the difference between last and rst neurofeedback run. This result might be ascribed to our limited number of participants and the heterogeneity of their learning strategies possibly resulting in non-linear learning curves across runs. On the other hand, we observed signi cant structural predictors of the most effective AI regulation run. In addition, a number of brain structures signi cantly predicted average AI activity, that might alternatively capture, at least partially, individual regulation strategies differences as well as more complex learning processes. Notably, our results highlighted a major contribution of the cerebellum in relation to both the most effective regulation run and average neurofeedback performance, a brain region previously associated with fMRI neurofeedback and brain-machine interfaces and assumedly to underlie reinforcement learning mechanisms (Shibata, et al., 2019).
In line with the well-known functional effects in terms of brain network involvement during learned regulation (Lee, et al., 2011), analysis of single runs showed that the predictive anatomical network was less distributed in the last run, a phase when regulation is usually more consolidated, with respect to the initial phase. Involvement of frontal, prefrontal, occipital, medio-temporal and cerebellar regions was observed in both rst and last run, whereas the caudate and pallidum contributed only to the rst run.
Overall, our neuroanatomical ndings are in accordance with previous functional studies as well as recent theoretical models on the general neural mechanisms underlying brain self-regulation, both indicating the dorsolateral prefrontal cortex and lateral occipital cortex associated with attentional processes related to feedback control (Caria, 2020;Paret, et al., 2018;Shibata, et al., 2019;Sitaram, et al., 2017), and pointing to the basal ganglia as mediators of core learning-related processes such as salience-based strategy selection and reinforcement assessment (Skottnik, Sorger, Kamp, Linden, & Goebel, 2019). In our analysis, the striatum contributed to the rst run but not the last, suggesting decreased relevance of strategies' individuation processes when regulation is more easily attainable.
In addition, the ventral striatum has been proposed to enable monitoring the implicit reward value of fMRI feedback (Paret, et al., 2018;Sitaram, et al., 2017), whereas the orbitofrontal cortex would process feedback failure-associated signals (Paret, Zaehringer, Ruf, Ende, & Schmahl, 2019). Other regions such as the insula and anterior cingulate cortex (Emmert, et al., 2016;Shibata, et al., 2019) would also participate to regulation processes, likely by supporting continuous error monitoring and evaluation (Gaume, Vialatte, Mora-Sanchez, Ramdani, & Vialatte, 2016). In our results, differently from previous studies, the interindividual structural differences of the SAL network, including anterior cingulate gyrus, amygdala, ventral striatum and SN/VTA, appeared less relevant for AI-guided emotion regulation.
Furthermore, our ndings indicate a substantial involvement of cerebellar structures in learned AI regulation. The cerebellum, besides its clear role in adaptive motor learning (Galea, Vazquez, Pasricha, de Xivry, & Celnik, 2011), is implicated in higher-level cognitive functions such as encoding of internal models of mental representations (Ito, 2008;Sokolov, Miall, & Ivry, 2017). In particular, it supports adaptive predictive mechanisms during error-based and reinforcement learning (Ito, 2008;Sokolov, Miall, & Ivry, 2017;Swain, Kerr, & Thompson, 2011 2020;Wagner, Kim, Savall, Schnitzer, & Luo, 2017). Existing interconnections of the posterior cerebellum with ventral and dorsal striatum (Bostan, Dum, & Strick, 2013) might be then implicated in exchanging of reward-and error-related information during real-time fMRI training. The cerebellum has been also recently described as part of an integrated network along with basal ganglia, and prefrontal cortex subserving multiple non-motor functional domains (Bostan & Strick, 2018) including affective and socio-cognitive behavior (Pierce & Peron, 2020;Sokolov, Miall, & Ivry, 2017;Van Overwalle, Baetens, Marien, & Vandekerckhove, 2014). The widespread cerebellar interplay with several regions of the limbic network might support modulatory functions during emotion regulation (Baumann & Mattingley, 2012;Pierce & Peron, 2020;Schutter & van Honk, 2009). Cerebellum-mediated adaptive predictive mechanisms might thus enable selection and adjustment of appropriate regulatory strategies. During our real-time fMRI neurofeedback training participants explicitly induced selfgenerated affective states through emotional recalling and autobiographical memory retrieval. The e cacy of different emotional strategies is typically continuously estimated and assessed so as to retain and ne-tune the strategies that more likely lead to successful regulation. A probabilistic representation of the adopted emotional strategies is then likely implemented and dynamically updated. Additional networks such as the DMN, might also support the instantiation of predictable future events (Spreng & Grady, 2010;Suddendorf & Corballis, 2007). The DMN is posited to constitute a reinforcement learning agent mediating higher order predictive control of behavior through decision process based on prediction error estimation and reward feedback assessment (Dohmatob, Dumas, & Bzdok, 2020). In line with this perspective, the DMN along with the CEN might play an important role in real-time fMRI-mediated emotion regulation.
Emotional imagery and affective memory recalling have been proved effective in supporting emotional brain regulation (Linhartova, et al., 2019), interindividual differences in the capacity to adopt such strategies are then expected to signi cantly in uence brain regulation and functional characteristics of the medial temporal lobe. On the other hand, additional mechanisms might be implicated during implicit and covert emotional brain regulation (Ramot, Grossman, Friedman, & Malach, 2016;Taschereau-Dumouchel, et al., 2018;Watanabe, Sasaki, Shibata, & Kawato, 2017).
Notably, our structural network associated with AI-guided emotion regulation partially overlaps with the neurofunctional network associated with cognitive emotion regulation (Gross, 2014). For instance, we also observed involvement of the dorsolateral and medial prefrontal cortex regions, known to mediate top-down emotional regulatory processes. Though, our results mainly emerged from the comparison of up versus down emotional brain regulation whereas these regions are typically associated with down-regulation of emotions. Future studies are then required to clarify whether morphological characteristics of frontal and prefrontal circuits can differentially impact real-time fMRIbased up and down emotional brain regulation as in cognitive emotion regulation. In fact, recent evidence indicated that while up-and down-emotion regulation are both supported by regulatory regions in the frontal and prefrontal cortices, they are instead differentially mediated by multiple interactions with distinct emotion-and interoception-related regions (Min, et al., 2022). In general, current and previous evidence indicate that real-time fMRI-guided emotion regulation similarly to cognitive emotion regulation relies on frontal and prefrontal regions-mediated top-down regulation processes but additionally exploits bottom-up mechanisms through direct modulation of activity in emotional brain nodes, and thus possibly involving differential brain circuits in relation to speci c targeted regions.
In summary, our neuroanatomical ndings corroborate previous ndings of brain networks involved in real-time fMRI-based self-regulation of brain activity and extend them by highlighting a structural topography relevant for AI-guided emotion regulation. However, because of our limited sample size, these results are still preliminary and need to be corroborated by further studies with larger samples possibly adopting homogenous experimental protocols. Future investigations might for instance explore brain structural landmarks at both macro-and mesoscale contributing to speci c emotional brain regulation strategies in contrast to more general regulatory mechanisms. In the eld of cognitive emotion regulation, it has been proposed a distinction, with some overlaps, between brain circuits speci cally supporting emotion regulation and those contributing to general neurophysiological regulation (Gross, 2014). Considering that real-time fMRI-based emotional brain regulation appears to have some commonalities with cognitive emotion regulation, an analogous differentiation is also plausible in the context of fMRI neurofeedback.
Declarations Figure 1 % BOLD signal changes during AI regulation of all partcipants.
Violin plots represent % BOLD signal change in the AI during real-time fMRI-guided regulation runs. White circles show the medians; box limits indicate the 25th and 75th percentiles as determined by the R software; whiskers extend 1.5 times the interquartile range from the 25th and 75th percentiles; polygons represent density estimates of % BOLD values and extend to extreme values. Violin plots were created with BoxPLotR (http://shiny.chemgrid.org/boxplotr). * p < 0.05

Figure 2
Voxel-based GM predictors of AI regulation Results of whole brain analysis resulting in GM weights map of KRR model indicating voxel-based contribution to predict average AI activity during all regulation runs. The color bar refers to the weight (%) of each voxel contributing to prediction of % BOLD signal change in the AI. Voxel-based predictive values were used to compute the total normalized weight of each anatomically labelled GM region, as reported in Table 1.

Figure 3
WM regions predictive of AI regulation Results of whole brain analysis resulting in WM weights map of KRR model indicating voxel-based contribution to predict average AI activity during all regulation runs. The color bar refers to the weight (%) of each voxel contributing to prediction of % BOLD signal change in the AI. Voxel-based predictive values were used to compute the total normalized weight of each anatomically labelled WM region, as reported in Table 3.