A lot of different studies have shown the possible applications of real-time functional magnetic resonance imaging (rt-fMRI) and facilitate the importance of this method (Pandria et al.; Weber et al.; Lührs, Riemenschneider, et al.; Russo et al.; Ciarlo et al.; Sorger et al.; Tursic et al.; Weiskopf). The main prerequisite for successful rt-fMRI is the ability to get access to the fMRI data in real-time. Real-time can be seen as the time between two consecutive recorded images or in short, the repetition time (TR). One should, however, also note that this time interval should include both data export and processing to make sure

that the application keeps up with the acquisition. The beforehand mentioned experiments can be seen as separate classes of applications for rt-fMRI: Quality assurance (QA), rt-fMRI analysis, brain-computer interfaces (BCI) and neurofeedback (NF); which will be discussed separately to highlight out the importance of reliable data transfer in different scenarios.

Quality assurance

In a quality assurance design (Heunis, Hellrung, et al.; Lührs and Goebel; Heunis, Lamerichs, et al.), rt-fMRI data is used to verify that the recorded data has sufficient quality for subsequent use. This can be done in real-time and offline scenarios (fMRI experiments not requiring data in real-time). For offline scenarios, the rt-fMRI quality assurance can for example show that an experimental run needs to be repeated due to too much movement of the participant which can lead to a low sensitivity of the neural signal of interest. Another application would be detection of lack of activation in expected brain regions, which could be the result of not correctly understanding and/or following the instructions by the subject or because of other technical issues in the experimental setup (e.g., screen not fully visible, auditory stimuli not perceptible, loss of synchrony between the acquisition and the stimulus presentation). More details about potential applications of real-time quality control can be found in (Heunis, Lamerichs, et al.). Most of these measures don’t directly require that the data used for the measure is available in a stable real-time fashion since the data is not directly used for further real-time applications. Still the severity of a slow or unreliable real-time export can cause missing information which would potentially have helped to acquire better fMRI data.

rt-fMRI analysis

This design has the potential to get fMRI analysis results in real-time which could for example be used to stop an experiment earlier as soon as enough data for a respective task is available. Also, this design can be a pre-step for the BCI and QA applications. The importance of a reliable data transfer lies here in the processing times needed in general for real-time preprocessing and analysis methods which would benefit a lot from fast transfer time allowing for more preprocessing and analysis time.


The field of BCIs includes a variety of approaches that use the fMRI data in real-time to determine different measures that can be used to build a non-invasive interface to the subject’s brain. The potential applications can include assessing consciousness, communication, controlling dynamic experimental design, and controlling an external device (Sorger and Goebel). The speed of the data export is particularly crucial for communication and controlling because it strongly influences the feasibility of the application.


A special technique within the BCI field is Neurofeedback (NF) which, in short, presents the current brain activity (implicitly or explicitly) to the participant and thereby allows for modulation of the respective brain region or pattern. The main goal of a NF training is to induce neural plasticity changes, which can be maintained after the experiment. Different methods are used to visualize the feedback signal. In this article, we mainly focus on the time-dependent variables of these experiments and differentiate continuous and intermittent designs. In continuous design, the feedback is presented in real-time continuously giving most of the information directly to the participant (for example, a volume-by-volume motion information [Yang et al.] or moment-to- moment neurofeedback) whereas in an intermittent design the participant receives the feedback after performing the respective task for a certain amount of time without receiving real-time feedback (for example, the magnitude of the activation in a respective ROI after a block of pain stimulus [Chapin et al.]). The continuous feedback thereby requires that the respective fMRI data is available in real-time to be preprocessed and analyzed to generate the feedback signal. For intermittent designs, more time is generally available, but the duration of the intermittent designs can also be rather short still requiring reliable feedback presentation times.

As shown above, the typical parts of the rt-fMRI procedure are the data acquisition, data transfer, preprocessing, QA, analysis, and BCI application, of which data transfer is often neglected due to a lack of options. However, the data transfer of the fMRI images from the scanner to an external processing platform for further processing of the data is crucial to guarantee a real-time scenario; therefore, it is the most defining part of the workflow. Since there are multiple MRI device manufacturers and no standard format does exist, best to our knowledge, a standardized approach for the real-time data export would improve the quality and reproducibility of rt-fMRI. This would address the missing standardization in rt-fMRI experiments and ensure reliable transfer which is one of the significant requirements for rt-fMRI experiments. Without a reliable, stable, and fast data transfer the fMRI images can’t be analyzed in real-time which would cause potential incremental delays or discrepancies in, for example, the stimuli presentation times. Furthermore, a non-reliable data transfer could result in missing data reducing the effectiveness of brain-computer interface (BCI) applications. Most of the current rt-fMRI studies do not give enough detail about the underlying technical aspects of how the data is accessed and analyzed in real-time (Thibault et al.). The same is true for any quality measures of the real-time data export concerning incremental delays or jittering during the data transmission or reconstruction. In this article, we investigate the potential difficulties and pitfalls that can occur during the export of fMRI data in real-time and point out the importance to resolve potential problems by providing a ‘direct’ export solution to the used procedures. This is done using a standard rt-fMRI setup by comparing different export procedures (‘indirect’, ‘direct’) used simultaneously.


To investigate the influence of data transfer times in real-time fMRI experiments 27 datasets were acquired using a 3T MRI (Siemens Prisma 3T, Siemens Healthineers, Erlangen, Germany) as well as five datasets using a 7T MRI (Siemens Magnetom 7T, Siemens Healthineers, Erlangen, Germany) including different TRs but always 100 volumes. Detailed information is available in the respective references (Lührs, Krause, et al.; Lührs, Poser, et al.). To measure the data transfer times, the start of the volume acquisition, as well as the receive time of the specific volume on an external computer were recorded and two different data transfer procedures were investigated simultaneously by using both transfer methods at the same time. The data was transferred to a real-time analysis computer running Turbo-BrainVoyager (TBV, Brain Innovation B.V., Maastricht, The Netherlands) to log the processing times or a self-build receiving script (Available in python and MATLAB [Heunis, Hellrung, et al.]). Example sequence parameters for the respective exported data can be found in table 1.

Table 1.Relevant parameters of an example 3T & 7T sequence
3T 7T
SequenceName epfid2d1 mbep2d
SliceThickness 2 1.5
RepetitionTime 2000 1400
EchoTime 30 21
SpacingBetweenSlices 2 0
NumberOfPhaseEncodingSteps 100 132
PercentSampling 100 100
PercentPhaseFieldOfView 100 100
AcquisitionMatrix 100\0\0\100 132\0\0\132
FlipAngle 77 60
Resolution 2mm isotropic 1.5mm isotropic
NrofSlices 32 72

The first procedure used a standard ‘indirect’ export of single mosaic DICOM files for each volume separately using the underlying server message block (SMB) network protocol. We call this method ‘indirect’ for easier differentiation between the two methods. This is provided by the MRI manufacturer without further tools needed. The (SMB) protocol was used since it is the default protocol implemented in Microsoft Windows desktop environments. Other protocols like SAMBA or NFS would result in a similar outcome. The second approach was a newly developed method based on a ‘direct’ TCP/IP-based connection between the MRI reconstruction computer and the receiving external real-time computer which allowed to send the volume data and specific header information for each volume. In this case, the data was exported directly using a custom real-time data export module (image calculation environment

(ICE) functor) that is appended to the end of the image reconstruction chain. An overview of the described procedure is shown in figure 1.

Figure 1
Figure 1.‘Direct’ real-time export procedure.

In the ‘direct’ real-time export procedure, the data is extracted using an ICE functor inserted (imagesend functor) as a last processing step in the reconstruction chain of the reconstruction computer (top left part). The data is then directly transferred to the receiving real-time computer via the MRI console computer using port forwarding. The reconstruction computer and the MRI console computer are connected using a direct network connection.

To connect the external real-time analysis computer to the image reconstruction computer a port forwarding tunnel was used which allowed routing the connection through the MRI host computer. The configuration of the port forwarding was performed using either the network shell (netsh [Microsoft]) from Microsoft Windows or a standalone port forwarding application (Putty) both running on the MRI host. In both procedures, the server that received either the DICOM or the pixel volume data was an external computer connected to the MRI host computer using a Gigabit-Ethernet connection. The measured times (volume acquisition trigger and volume receive times) were compared using the trigger at volume t+1 compared to receive time of volume t to calculate the pure data transfer time. Thereby the trigger was recorded separately to be most accurate.


Overall, the mean data transfer time measured using the 3T MRI for the ‘indirect’ file- based DICOM export was 513.9ms (+/-std 171.7ms) whereas the data transfer time for the ‘direct’ TCP/IP-based export was 89.5ms (+/-std 76.9ms). For the 7T MRI, the ‘indirect’ file-based DICOM export took 301.03ms (+/-std 87.14ms) compared to the ‘direct’ TCP/IP-based export which needed 29.82ms (+/-std 18.29ms). The ‘indirect’ file-based data export also showed strong jittering between successive volume data transfers which was not visible or less prominent in the TCP/IP-based export. The jittering was stronger for the multiband sequence measured on the 3T MRI at one point in time only regardless of the transfer method (vol 34). A general overview of the recorded data over time is shown in figure 2.

Figure 2
Figure 2.Median data transfer times using ‘indirect’ files-based export and ‘direct’ TCP/IP-based export on 3T (blue) and 7T (green) MRI.

The error bars indicate the standard deviation of the transfer times.

Example data transfer times between individual volumes for different data sizes (200kb, 700kb, 1mb 812kb per volume) for standard and multiband EPI sequences are presented in figure 3. Different transfer patterns are seen for the ‘indirect’ approach whereas the ‘direct’ approach shows comparable results throughout each size.

Figure 3
Figure 3.Data transfer times as a measure between volumes.

In this plot the transfer times are measured as the time between individual scans for different volume sizes. The blue “TCP receiver” refers to the ‘direct’ export and in orange the “SMB receiver” to the ‘indirect’ export. The respective TR for each export size is shown using the grey line.


The ‘direct’ TCP/IP-based data transfer outperformed the ‘indirect’ file-based export both in data transfer times as well as in the robustness of the data transfer (i.e., reduced transfer time variability). This can be seen in the much smaller averaged standard error as well as in the combined data transfer plots (see figure 2). Also, the jittering artifacts were more prominent in the file-based data export which could potentially be caused by the underlying network protocol SMB, which was designed to allow massive data exchange for many clients at the same time but seems to be not perfectly suitable for real-time export of fMRI data (especially earlier versions of the protocol). With newly developed sequences that allow acquiring high-resolution images (i.e. large number of voxels) while keeping the repetition times below 500ms, the file-based data export may not be able to keep up with the acquisition and an incrementally growing delay might occur which would violate the real-time condition in which the data must be retrieved and analyzed during one TR. In addition, such cutting-edge applications also require the more general elements of the system, such as network architecture, to be considered and optimized for the continuous flow of large amounts of data. Care must be taken when performing real-time fMRI experiments with ongoing feedback and fast sampling with very low TR not to overlook delays in the processing pipeline. Here we only investigated the data transfer time but not the reconstruction time itself, which could also be a potential problem for extremely short TR values. Since we exported the data simultaneously using both methods, the reconstruction time would not influence our measures. It is reasonable to assume that use of both export methods in parallel did not by itself cause additional delays to either. Nevertheless, different sequences and acquisition protocols (especially spatial resolution and parallel imaging acceleration and multiband factors) result in different behaviors caused by the reconstruction itself. This was not systematically addressed in this work and should be further investigated.

General guidelines

Simplifying the pipeline by avoiding unnecessary file write, transfer, and reading as implemented in a direct connection between the image reconstruction computer and the real-time analysis computer can improve quality assurance and help to ensure proper rt-fMRI data transfer and analysis. We only applied the solution using the MRI devices that were available to us, but the proposed solution can be generally applied to other manufacturers as well without further limitations. Since the proposed method might not be immediately available or suitable for all MRI devices (e.g. clinical scanners) we propose some guidelines below to potentially improve the ‘indirect’ real- time data export. Data export should be tested for individual sequences and parameters as the size of the individual volumes has a direct influence on the resulting transfer times as shown in figure 3.

The direction of file sharing

In the file-sharing procedure, one computer usually serves as a sender and the other one as a receiver. This also applies when sharing files through a shared folder using

for example the SMB protocol. In this case, it is important to apply the correct order to prioritize file transfer. The computer that should perform the real-time computations on the fMRI data should create a shared folder on its system and the MRI console should connect to this shared folder and drop the fMRI images into the folder. This ensures that the file is listed immediately for transfer and acknowledged by the file transfer protocol.

SMB version

The version of the used network protocol is crucial since a lot of potential problems are resolved in later versions of the respective protocols. This is especially a problem for MRI devices since the protocols are part of the operating systems which are not always up to date for MRI console computers. If two computers are communicating the highest version available on both computers is used for file transfer. This could result in the use of an older version of the protocol, e.g. 2.0, which are slower and create more overhead per file transfer. To avoid any delay in name resolution, especially for older versions of the SMB protocol, the IP of the computer should be used to access the shared folder and write the file instead of the computer name.

Number of devices within the network

The more devices are in the local network the higher the chance that the file transfer might be delayed. Therefore, the network should in the best case only consist of the required computers. This could be achieved by using only the components in the network or creating a virtual network. Routing tables could also help to improve the transfer rates in the case of many devices in the same network.

Number of files to be transferred

Sending the whole functional data of one volume in one file is the optimal solution to reduce network delays. Splitting slices into separate files might reduce the performance and can cause additional delays. Still slice-by-slice transmission might be desirable depending on how the processing is performed (e.g. performing preprocessing on already available slices).

The developed ‘direct’ approach does not need to consider the above-mentioned parameters because a direct TCP/IP connection is usually not affected by these problems. The only potential problem would be the bandwidth of the network interface. This should be able to handle the traffic to ensure a real-time transfer. The developed functor, as well as the source code, is available upon request. An example receiver API for the fMRI images can be found on online (Fladd/RealTimeExportServer: Export Real-Time Pixel Data from a Siemens MR Scanner via TCP; Auer).

Impact of file size

Based on the results presented in figures 2 and 3 we assume a linear relationship between the data size and the transfer times, also because the speed is limited by the bandwidth of the ethernet connection. Using fiber optic connections could reduce the transfer times drastically. Therefore, different sequences and parameters should be tested individually for reasonable transfer times.

Quality assurance measures

To assure the data transfer in real-time is working correctly quality measures should be considered that point out the transfer time within each volume that is transferred. This could be implemented in many ways but would allow to interrupt the session and try to resolve the problem or to have the respective data transfer times available for post hoc evaluation. A standardized reporting of these transfer and processing times would be important for rt-fMRI experiments that use real-time data to present information to the participant. This is also true for intermittent designs since the delay of the data transfer can also reach several seconds (in the range of 10 to 20 seconds delay for a single volume).

Potential consequences for rt-fMRI experiments

Data transfer delays in rt-fMRI experiments can result in a variety of consequences. In case the transmission is jittered, and the data is not available regularly every TR, the participant might experience visual disturbances of presented feedback information, especially in continuous designs updating the display at every TR. Whether this influences the performance of the participant is unknown but should be investigated in more detail. For intermittent designs that need to display the data at a certain point in time, they could potentially fail to present the feedback since the data is not available. This could cause a general variability in the experiments and might be a problem in comparability across participants. Since the problem is not discussed widely in the community and not enough details are reported in the respective publications, the consequences are unknown but could potentially have led to suboptimal (behavioral) outcome measures in fMRI neurofeedback studies. As mentioned earlier, a reporting of these transfer and processing times would help to get more insights in the frequency of this problem. In some publications (Auer; Marxen et al.; Hellrung et al.; Krause et al.) a similar approach of direct data transfer has been mentioned indicating that the potential negative consequences of unrobust and slow data transfer have been fortunately avoided in a certain amount of real-time fMRI applications.


We showed that standardizing the data export for rt-fMRI experiments is important to decrease the probability of errors caused by the variability in data transfer times. Additionally, we showed that the proposed connection-oriented direct data transfer procedures of fMRI images lead to a decrease in transfer time variability which is especially important for NF experiments showing continuous feedback.


This research was financially supported by the European Commission’s Health Cooperation Work Programme of the 7th Framework Programme, under the Grant Agreement n° 602450 (IMAGEMEND) and n° 602186 (BRAINTRAIN). This article reflects only the authors’ view, and the funding sources are not liable for any use that may be made of the information contained therein.


Michael Luehrs and Rainer Goebel are working for the company Brain Innovation B.V. developing software for rt-fMRI applications.