Exploring protocol development: Implementing systematic contextual memory to enhance real-time fMRI neurofeedback

We explored the development of a protocol to treat patients with ADHD and Tourette’s Syndrome, replicating the ROIs of [32] and [28] (rIFG and SMA, respectively) in the treatment. By changing the ROIs in the algorithms, the protocol can be tailored to enhance/decrease activity/connectivity in any subregion of the brain.

As studies on neurodevelopmental disorders often target a younger population and SiV specializes in adult patients, we chose to explore the development of this treatment to treat adult patients. Results from neurofeedback studies have shown promise for these diagnoses, and analyses of the effects of stimulants and network deviations for these conditions have highlighted deviations in specific regions that can be targeted by neurofeedback. Furthermore, access to all of the required expertise and instrumentation at OUH/UiO/SiV allowed this exploration. This form of brain training shares parallels with physical training; for example, more training yields larger effects. However, as MRI scanning is both difficult and expensive, we chose to include VR and contextual memory to test if repeated sessions in VR could yield similar effects. If so, this would greatly heighten the applicability of the treatment, as patients could use a VR headset and continue training at home.

ADHD is characterized by pervasive and impairing symptoms of inattention, hyperactivity and impulsivity (deviations with respect to (wrt) inhibition, Fig. 2a), and affects around 5% of children and adolescents, and 2.5% of adults worldwide [52]. Patients may benefit from a combination of pharmacological and psychosocial structured treatments, where psychostimulants (e.g. Fig. 2b) are the gold standard for reducing symptoms ([57]); nonstimulants are less effective, and psychosocial treatments are recommended to learn to master the symptoms and deal with them in a more adaptive way [52]. However, since many patients experience little benefit and/or severe side effects from pharmacological agents, there is a need for new treatment options. The evidence for the longterm efficacy of psychostimulants is limited, and brain adaptation may be related to psychostimulants losing their effectiveness over time [57].

Figure 2:

Tourette’s Syndrome is a neurological condition appearing in childhood recognized by motor and vocal tics. The earliest and most frequent tics include blinking of the eyes, facial movements, and neck movements [58]. The diagnosis requires multiple motoric tics and at least one vocal tic which have lasted for at least one year [59]. In Norway from 2000 to 2010, 0.43% of children received a diagnosis of TS by the age of 12, and the internationally estimated prevalence of TS in children aged 6 to 15 years is 0.77% [60]. For a sizable number of patients, the condition persists into adulthood, and it is estimated to affect 0.08% of adults [59]. (Regions of the brain associated with TS symptoms are seen in Fig. 2c.) Pharmacotherapy and Habit Reversal Training/Exposure and Response Prevention/Comprehensive Behavioral Intervention for Tics ([61, 62, 63, 64, 65, 66, 67, 68]), are strategies used for treating TS. However, many individuals experience insufficient relief from these options, and there is a pressing need to develop new therapies for adults with TS. A new therapy - for adults - based on attention distraction shows promise [56] (Fig. 2d). Many patients diagnosed with TS have comorbid ADHD, OCD, and/or anxiety. The prevalence of comorbidity is as high as 90%, and these comorbid conditions often have a more significant impact on the Quality of Life of the patients than the tics themselves [59].

Treatment of ADHD (children) and Tourette’s Syndrome (adolescents) through rtfMRI-nf has been studied previously in [32] and [28], respectively. The ROIs of interest in these studies were rIFG ([32, 31]) and SMA ([28]).

For Tourette’s Syndrome (TS), tics (”unvoluntary” movements, neither voluntary nor involuntary, [72]) may have an origin in SMA; [73] showed that 2 seconds in advance of a tic, SMA was activated in TS patients. For patients with ADHD, a general hypoactiation of rIFG wrt inhibition was found in a meta-analysis ([74], Fig. 2a) and a stimulant (methylphenidate) which relieves the symptoms stimulates rIFG ([54], Fig. 2b).

A review on TS highlighted deviations in corticostriatothalamocortical (CSTC, [69]) circuits, where the direct, indirect and hyperdirect pathways dictate motor control (Fig. 3a). The same review showed a model of how the hyperdirect pathway bypasses striatum in inhibiting movement control ([69], Fig. 3b). Post-mortem studies have shown that TS patients can have a 50% reduction in GABAergic interneurons in striatum (part of the CSTC, [55]), and a recent PhD study demonstrated how attention distraction could be used as a treatment for adults with TS ([56], Fig. 2d).

Figure 3:

A common placement of Deep Brain Stimulation electrodes for treating movement disorders is in subthalamic nucleus (STN) ([70], Fig. 3c). An investigation with high spatial and temporal resolution through field potentials in the human cortex revealed that the hyperdirect pathway between inferior frontal gyrus and STN indeed exists in humans, and that it mediates rapid stopping ([75, 71], Fig. 3d). An investigation using diffusion tensor imaging to map the connections of right inferior frontal gyrus corroborated the connection between rIFG and STN [76].

Figure 4:

A meta-analysis of fMRI studies on inhibition (and attention) in patients with ADHD found a systematic reduction in activation in both SMA and rIFG concerning inhibition [53]. Based on these results, we aimed to enhance this pathway.

As rIFG may be trained as a consequence of the rtfMRI-nf method ([33]), we chose to provide feedback from SMA and measure resulting effects on both SMA and rIFG. Reinforcing both ROIs could aid both conditions; the primary interest is the relative importance of the ROIs for the benefit of the patients. Recent research has led to the hypothesis that ADHD and TS, and OCD and autism, may actually lie along a impulsivity-compulsivity spectrum [81].

The use of fMRI as a tool for brain training is expensive and time consuming both for the patients and health care providers. In need of a simplified brain training system, we used Virtual Reality (VR) goggles containing a visual experience mimicking their prior rtfMRI-nf training. VR based brain training lacks the feedback from the fMRI-feedback loop and is similar to sham-feedback sessions (which display prerecorded feedback), commonly used as a control in rtfRMI-nf studies ([28, 34, 82, 24], if not for ethical reasons [16, 3]).

Since many factors influence neurofeedback performance ([83, 84]), including psychological aspects ([85]), we aimed to test how episodic memory/VR could be used to enhance the reinstatement of the learning setting for rtfMRI-nf. Context to aid memory has been used since the ancient times (Fig. 4a), and VR induced context was shown to aid memory (Fig. 4b). Neurofeedback through VR has been used to test memory (Fig. 4c), reinstate context to aid memory ([86]), and experiments have been done to induce a mental context through rtfMRI-nf [30]. A model for how context may aid memory has been developed by deBettencourt et al. [80] (Fig. 4d).

We tested if VR simulation could trigger contextual/episodic memory ([87]) in a way that would allow us to extend the brain training, yielding results similar to rtfMRI-nf sessions. Our intention with using VR shares similarities with why participants in [32] were asked to practice daily between sessions of rtfMRI-nf; participants were given a cue-card depicting the feedback and asked to train by remembering the training in the MRI-scanner.

Our VR-based approach parallels the preparation of participants for MRI scans through exposure therapy using mock scanners and VR approaches [88, 46, 45]. Using VR during an MRI scan has been shown to reduce anxiety and improve results [47]. VR used to aid in reinstating context has been shown to enhance memory ([78]), and VR training has been shown to enhance attention in patients with ADHD [89]. Neurofeedback to practice motor imagery was also shown superior to repetitive practice ([90]), and a skill-based VR Cognitive Behavior Therapy has been shown to reduce chronic pain [11].

When exploring the development of this protocol for VR enhanced rtfMRI-nf, we used the open-access software OpenNFT as the rtfMRI-nf platform [26]. For the initial in vitro testing, we used the open-access data in [91] accompanying [26]. Using this platform and this dataset, we collaborated with radiologists/radiographers and translated the in vitro setup to perform continuous classification-based feedback (Case study 3 in [91]). We wrote code/algorithms and developed a gamified feedback design (based on [32] and [28]), and incrementally learned to tune the protocol through pilots. We wrote protocols and code for preprocessing and analysis using Matlab and accompanying toolboxes when needed. Included in the development was the simulation of the rtfMRI-nf experience through VR, the sham-feedback itself was based on screen-recordings of participants, as used in e.g. [28, 33].

Consent was obtained in both verbal and written form. Participants were provided with information about the purpose and procedure of the study in written form, to which they agreed in advance of the initial trial. Further information was given verbally by the first author on each trial day, witnessed by the second author and radiographers. Radiographers ensured participants complied with safety precautions regarding MRI scans in advance of all trials.

The research related to human use has been complied with all relevant national regulations, institutional policies and in accordance with the tenets of the Helsinki Declaration, and has been approved by the authors’ institutional review board or equivalent committee.

As outlined in Sec.2.2, the developmental goal was to provide neurofeedback from SMA and rIFG. The ROI in the rtfMRI-nf procedure is the selected subset of the brain from which the activity is analyzed, feedback is calculated, and then presented to the participant. This study was an ongoing research project to develop both the scientific argumentation of the protocol and the competence to perform it. The different phases of the study thus targeted the given ROIs with increasing precision.

In Pilots 1-4, we used the mask in parietal lobe defined by [26, 91]. For a coarse restriction, WFU_pictatlas ([3, 96]) may be used, and this was how we defined the ROIs during pilots 5 and 6 (rIFG pars triangularis and pars opercularis through the Brodmann Areas 44 and 45 based on [97]). The ROIs were defined with a dilation factor 2 to ensure that the ROI contained the volume across participants. For pilot 7, we used the Julich Probabilistic Maps ([98]); which is probabilistic maps based on averaged/simulated dynamics in the human brain, defined in MNI-space. We chose this atlas as it is the foundation of the Multilevel Human Brain Atlas, is open to all neuroscientists via the Human Brain Project’s research infrastructure EBRAINS, and due to previous experience of one of the researchers in using regions defined through this atlas for neurofeedback. We reinforced SMA (based on e.g. [28, 5, 83, 37, 99, 33]) and rIFG ([32, 97, 54, 100, 101, 33, 56]).

In the EBRAIN database ([102]) we used the JULICH_BRAIN_CYTOARCHITECTONIC_MAPS_2_9_MNI152_2009C_NONL_ASYM to define the ROIs. The 3D reconstructed histological datasets in [102] are transferred into two reference spaces, the single subject MNI-Colin27 space, and non-linearly transformed into ICBM2009casym space; the latter being a compromise between the anatomically detailed MNI-Colin27 and the more generic but smoother MNI305 template [98]. Through the accompanying text-file one finds that SMA (in both hemispheres) is defined through areas 135 and 136, and that rIFG is defined through area 120 and 122.

As both rIFG and SMA are defined through two subregions, we combined these subsections using temporary variables and performed voxel-by-voxel summation in a custom-written Matlab script. For voxels where the ROIs to be combined were overlapping, we used the maximum value.

For the real-time fMRI calculations, we used OpenNFT [26, 91]. The rationale for choosing OpenNFT over e.g. Turbo Brain Voyager was financial, OpenNFT is open-access, which also induces easier communication and replication of the protocol by other scientists. The OpenNFT platform is a BCI that takes the ROIs to be investigated as input and analyzes the activity of this/these ROI(s) during the experiment. We wrote a gamified feedback where the output from OpenNFT was transferred into a game where a rocket was flying through space during activation-blocks (resembling [32]), and a submarine was sinking during deactivation-blocks, with a goal to obtain the maximum, summed score. As the output from Open-NFT was single number between 0 and 1, we reversed the feedback by using 1-OpenNFT feedback during downregulation blocks (thus, no rest-periods, [28]); maximum BOLD signaling in the chosen ROI then effectively yields a minimum score in the feedback during these periods.

Reinforcing both activation and deactivation was chosen to enhance conscious control of the ROI ([28, 5]). The feedback was mirrored onto a LCD screen behind the MRI-scanner, and the participant in the scanner only saw the rocket/submarine, and the updating score. In addition to the rtfMRI-nf sessions and a sMRI, we included a 12-volume pre-fMRI sequence to be used in adapting the ROI to the fMRI-space of the person in the scanner.

In our exploratory development we initially developed a procedure in line with the setup of OpenNFT, a block design with task and rest periods tunable through an OpenNFT configuration file ([32]). In this configuration file, output during rest periods equals zero. As we wanted to purposefully enhance deactivation of the ROI during rest periods ([28]) in our protocol, we chose to change the paradigm and developed a “block-design” consisting of 2 blocks. (The results from pilots 1-4 and pilots 5-7 are thus not comparable.) The rest block contained only the initial 10 volumes (for saturating the magnetic field), and the remaining 200 volumes constituated the task block. This task block was further subdivided in activation and deactivation periods, which were not visible in OpenNFT, but were presented in the feedback to the participant in the scanner.

The blocks were a compromise between [28, 32, 26], where we recorded 210 fMRI volumes (as in [26, 91]), designed a gamified feedback based on [32], and adapted the block length from [28] (20 volumes) to 18 to create an equal length of all blocks according to our repetition time (TR, 2.15 seconds).

The Julich Brain Atlas ROIs in Sec.3.1. are defined in MNI-space. To be able to track the ROI precisely during the rtfMRI-nf, a conversion of the ROI from MNI-space/coordinates to fMRI-space is required. We wrote an SPM Batch algorithm to adapt and convert the MNI-space-defined ROI. This conversion was done through a combination of an sMRI volume and an fMRI volume recorded first during the rtfMRI-nf training. The purpose of this conversion was to accurately target the specific region in the brain of the person in the scanner. Details of extracting the ROIs and of the SPM Batch process is described in SI.1-2; examples are shown in Fig. 5.

Figure 5:

Participants were given a general introduction to the principles of the procedure in advance of the experiments. The way this exploratory protocol is designed is with the goal of keeping the invasiveness to the participants to a minimum; therefore, our procedure did not require any additional information other than what was recorded during the rtfMRI-nf training sessions. Developing an objective, highly standardized protocol may circumvent the need for a localizer scan (as in e.g. [28, 103, 24]). The ROIs targeted were chosen due to anatomical deviations correlating with the diagnoses. Therefore, given that the patient is correctly diagnosed with the condition, they would benefit from our selective brain training when targeting the intended anatomical position. We chose to use a method to translate the anatomical MNI-space data of the regions to the participant fMRI-space to ensure individual adjustment (SI.2). This also removes the need for prerecorded fMRI data to be used, for example, in machine learning, to enhance precision (as in [104]). With regard to safety, in collaboration with professional radiographers, the preparation to and conduction of the rtfMRI-nf was similar to a normal MRI-session. Recruitment, information and preparation of the participants are described in SI.3.

While performing the rtfMRI-nf, participants in the MRI-scanner saw an animation of a spaceship flying upwards. In activation blocks, the speed of the rocket was dictated by the simultaneous activation of regions within their brain. Activation blocks were separated by periods of rest, during which no feedback was given. Starting from pilot 5, we modified the paradigm by replacing the resting periods with periods where active deactivation was reinforced. The VR simulation included a walk through the hospital toward the MRI-scanner, followed by a sham-feedback session based on a random sampling of screen recordings of earlier rtfMRI-nf sessions.

The experiments were performed at Oslo University Hospital, on a Siemens Magnetom Prisma 3T whole body MR scanner ([28, 91, 92]). For the fully developed protocol, each session consisted of the following: 1) a 12-volume pre-fMRI resting-state scan, 2) a T1-weighted sMRI scan, and 3) three rounds of rtfMRI-nf. The first two scans were used in a Matlab/SPM batch fMRI-space adaptation algorithm. The 12 volume pre-fMRI scans were recorded using a T2*-weighted EPI sequence TR/TE = 2150/28 ms, flip angle = 74°, 35 × 2.5 mm slices with a 0.6 mm slice gap, 22 cm FoV, and voxel size = 2.2 × 2.2 × 2.5 mm³. The T1 weighted scans were recorded with TR (repetition time)/TE (echo time) = 1.900/3.16ms, flip angle = 9°, 192 × 1 mm slices, matrix size 256 × 256, 256 mm FoV, voxel size = 1.0 × 1.0 × 1.0 mm³. The following three rounds of rtfMRI-nf were recorded with the same parameters as the initial 12 volume pre-fMRI resting state scan.

Our rtfMRI-nf protocol required the synchronization of the laptop performing the real-time analyses with the MRI scanner, enabling the software to expect and wait for incoming fMRI volumes. The repetition time (TR) of the MR-scan was set such that an entire loop of MRI-transfer, MRI-analysis/feedback calculation, and feedback display were conducted within each TR. (OpenNFT has additional safety checks to ensure analyzing the correct sequence.) The setup led to the raw MRI image being sent to a shared folder (shared between the MRI scanner and the laptop), whereas another program renamed and transferred the MRI image to a new folder. In this folder, OpenNFT expected the images to be analyzed to appear and analyzed all incoming files. OpenNFT was configured to analyze the BOLD data for specific ROIs within the fMRI volumes, and the calculated feedback was subsequently sent to our gamified feedback display, which is mirrored onto the LCD screen at the back of the MRI scanner. This loop was repeated for each fMRI-volume. We used OpenNFT version number 1.0.0rc0 and matlabengineforpython 2021a. The laptop-fMRI synchronization was based on [105]. The setup is further explained in SI.4.

Radiographers translated our experimental setup and rtfMRI-nf block-design and created a MRI-template. Our design included a 12-volume pre-fMRI resting state scan - with the same parameters as the later rtfMRI-nf sessions -, then subsequently a T1 weighted sMRI, in advance of three rounds of rtfMRI-nf. A 12-volume pre-fMRI was used to ensure that the magnetic field was homogenized/saturated at the 12th volume. This volume was then converted from dcm to nii format. The sMRI was first converted to axial slices, then converted to nii format, and subsequently combined to a stack through FIJI/ImageJ. The converted 12th volume pre-fMRI and sMRI files were then processed via the normalization/MNI-fMRI-space SPM Batch algorithm described in SI.2.

After the completion of the SPM Batch algorithm, the output files underwent further processing through another Matlab algorithm to create two masks: one weighted mask that was filtered to remove residuals (post-filtering/smoothing, with the deletion of voxel values under 0.1 and above 1.0), and a binarized mask. The binarized mask involved first deleting voxel values below 0.1, followed by setting all non-zero voxel values to 1.0.

To assess if we could utilize VR to enhance the efficiency of the treatment, we developed a VR platform with a sham replication of the rtfMRI-nf experience to be used for home training. The sham feedback was created through screen recordings of rounds of rtfMRI-nf run in vitro on data collected in earlier pilots. To aid the reinstatement of the mental context surrounding the experience, the VR experience included a walk through the hospital toward the MR-scanner. The MR-scanner noise was also added during the sham-feedback. The VR experience was made as simple as possible, such that the participants only had to press “Play” after mounting the VR-headset (SI.5).

Analysis of the results was conducted using the Matlab SPM-based toolboxes CONN and REX ([106, 107]) for the analysis of the connectivity and activity of the ROIs, respectively. Details of this analysis are described in SI.6. We chose to register each round for each participant as an individual subject to prevent the automatic averaging of participants. For instance, a participant with two separate sessions, each containing three rounds of neurofeedback, resulted in six different sets of fMRI scans, but only one T1-weighted sMRI was imported.

To ease replication and enhance standardization, we conducted the analysis following the procedures suggested by CONN, only actively selecting the additional REX output. Through the standardized 4-step analysis setup, the CONN GUI culminates in an interface where one can e.g. choose to selectively investigate how the connectivity between two isolated ROIs - predefined from a Harvard- Oxford brain atlas - changed during the fMRI scans in the study. Through the extra REX output, the analysis also exported additional files (interrogated through the Matlab toolbox REX) where the activity pattern for each isolated subregion could be displayed.

In this section, an overview of the protocol is given in Fig. 6, and the details of the steps are described in the supporting information (SI). Links to respective sections describing the steps are included in the flow diagram.

Figure 6:

In this study, we explored the development of a protocol to conduct rtfMRI-nf. Additionally, we performed preliminary testing to investigate how VR may potentially enhance the treatment through sham feedback home-training. We performed pilot testing throughout the exploration of the different stages of the procedure. Pilot 1 was performed purely in vitro, pilots 2 and 3 were conducted on a singlesubject each, pilot 4 involved 10 volunteers, pilot 5 involved 2 volunteers, pilot 6 involved 3 volunteers, and pilot 7 involved 2 volunteers. Representative results from the 7 pilots are displayed in Fig. 7 and Fig. 8.

Figure 7:

Figure 8:

First we made the platform work in vitro, and Open-NFT as seen in [26] is displayed Fig. 7a. Fig. 7b shows the end display after an early pilot using our adjusted software; the center window is the feedback display, and the end score in this round was 520. In Fig. 7c-d, the rocket, which represents how the feedback was presented, is displayed in its minimum speed (Fig. 7c) and maximum speed (Fig. 7d), and the image used for relaxation blocks is shown in Fig. 7e. We initially used a block design with activation vs. rest ([32]), but later decided to reinforce activation and deactivation ([28]).

Connectivity-analyses of the MRI-Images from the rtfMRI-nf pilots 2 and 3 (tested on the first author, 2 sessions, 1x3 and 1x1rounds) through the Matlab toolbox CONN is shown in Fig. 7f-j. Furthermore, we tested whether the connectivity of our initial rIFG mask to the mask given feedback changed during the training, as indicated in the three sequential rounds (Fig. 7f). In Fig. 7g-h sections from the CONN definition of rIFG and SMA is shown, respectively. In Fig. 7i-j connectivity analyses from the CONN definitions of rIFG (both the pars triangularis and pars opercularis) are shown. Fig. 7j includes a refined definition of the combination of rIFG. Of primary interest in Fig. 7i-j is the indication of the effect of time and sequential training. As these were results from rtfMRI-nf of the same participant spread across two times, the heightened activity level in the fourth column (with respect to the first) may indicate that the participant learned/the brain matured from session 1 to session 2.

In pilot 4 we conducted a pilot of 10 volunteers, the results from which are shown in Fig. 7k-o. In Fig. 7k, the average total score for each of the participants are shown through the 6 rounds of rtfMRI-nf, divided into two sessions, separated by 1 week. In Fig. 7l-m the averaged developing activity level through REX analyses of the binarized mask given feedback and rIFG is shown. The mask given feedback (Fig. 7l) did not display an increasing trend, which was in contrast to the activity in rIFG (Fig. 7m), which indicated an increased activity level from start to finish. In Fig. 7n-o example images are shown from REX analyses of the activity in the binarized mask and the weighted mask, respectively, in the parietal lobe for one of the participants in the 4. pilot. Limitations of this pilot include e.g. the definition of the ROI masks, and the use of fMRI-space adapted masks to a different MRI-scanner and different brains.

For pilot 5, we tested redefining the ROI masks given feedback on two volunteers and created a mask of rIFG for the first volunteer and a mask of SMA for the second volunteer (Fig. 8a-d). We also altered the block design of the neurofeedback, replicating the method described in [32] for rtfMRI-nf of the first volunteer and the method described in [28] for rtfMRI-nf of the second volunteer. The masks were created in WFU_pickatlas and output in MNI-space, but the reslicing of the masks was done based on earlier pilots. The first volunteer did show a small increase in the activity level of rIFG given feedback (Fig. 8a), but the connectivity between rIFG and SMA did not increase (Fig. 8b). The second volunteer did not show an increase in activity in either SMA or rIFG.

In pilot 6 we tested the use of support vector machine learning to enhance the efficiency of the neurofeedback protocol. Based on [26, 91], we used the PRoNTo toolbox ([108]) to test adapting the mask given feedback through the brain training, replicating [109]. Due to the time required for updating the masks, we only updated the mask once (during the sMRI), and did one session of rtfMRI-nf on three healthy volunteers, all training SMA. REX results from these pilots are shown in Fig. 8e-g. The results show that the activity in the binarized SMA increased from the start to the finish for all three volunteers (Fig. 8e). The activity also increased in the weighted SMA mask being PRoNTO updated (Fig. 8f), and similarly, for two of the three volunteers, the activity in rIFG also increased through the training (Fig. 8g).

In pilot 7 sham feedback was tested, both in the MRI-scanner, and during the VR-home training. Due to an unforeseen event, we were forced to conduct sham feedback on both volunteers in session 1, and the VR-home training therefore lost a lot of its purpose as we have no reference to which to evaluate the performance of the volunteers. Analyses of the neurofeedback training sessions through REX and CONN are shown in Fig. 8h-k. In this pilot the MRI-protocol was updated, and additional software was written to include adaptation of the mask to the fMRI-space of the volunteer in the MRI-scanner. In Fig. 8h and Fig. 8i the averaged activity of SMA and rIFG through the sessions are plotted, respectively; evident from these plots is that the activity of both the ROIs for the two volunteers increased through the brain training. Also of interest is the fact that there is a marked activity increase from session 1 to session 2. In Fig. 8j the connectivity between SMA and rIFG is plotted. Also evident from this plot is the general increase in the connectivity between the two ROIs. This increased correlation of activity may indicate that conscious use of SMA, reacting/adapting to the neurofeedback, may enhance the activity of rIFG. Feedback from the volunteers revealed that the VR-platform need further development. The similarity of each round of the VR-home training was less inspirational than the neurofeedback training at the hospital, resulting in them not performing much home-training.

In Fig. 8k results from an in vitro rerun of data from pilot 4 is shown, where feedback was shifted from the parietal mask from [91] to personalized rIFG. The result from this analysis also indicates that rIFG is enhanced through the rtfMRI-nf sessions, in accordance with [33].

Through implementing this targeted brain training on patients, we anticipate being able to measure increases in SMA and rIFG BOLD responses through fMRI analysis. We anticipate that these same increases will be paralleled by enhanced connectivity and activity, as demonstrated through MATLAB analyses using CONN and REX, respectively.

The extended implications of increased conscious control over these brain regions for patients with these diagnoses may reduce the severity of their conditions. However, to demonstrate this, long-term follow-up studies implementing the completed protocol on patient populations are necessary.

During the development, there were no blinding or other controls for biases. All the volunteers were recruited through ad hoc processes; when pilots were to be arranged and the MRI scanner was available, volunteers were contacted through personal networks of the employees at Nordic Neurotech. Therefore, there was no systematic control of the personal characteristics, homogeneity, or representativeness of the samples in the different pilots. But as e.g. different ROIs, different block-designs, and different feedback paradigms were utilized through the project, comparison between pilots can not in any regard be justified. Developing understanding of the software and instrumentation throughout the project also adds another variable dimension.