Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Real-time functional magnetic resonance imaging (rt-fMRI) is a powerful neuroimaging tool for monitoring brain activity and neurofeedback (NF) applications with promising therapeutic potential in psychiatric and neurological disorders. However, technical implementation of NF using acquired real-time fMRI and/or predicted real-time fMRI signals based on electroencephalographic (EEG) records remains restrictive and often lacks reproducibility. Here, a fully Python-based pyOpenNFT framework was designed for greater flexibility, modularity, and real-time processing efficiency. Its functionality was also extended with a ML-based prediction server for the fMRI NF signal using processed EEG records. The framework streamlines fMRI data acquisition and/or EEG-based prediction, NF signal estimation, and quality assessment (rtQA) without necessarily requiring a GUI. The FastAPI-based implementation for an EEG-based predictor integrates a Lab Streaming Layer (LSL) interface for processed EEG records and delivers real-time predictions of fMRI time-series for target brain regions. The system supports the visualization of additional NF sources by querying a RESTful interface, facilitating interoperability with external applications. Efficient real-time processing is achieved through parallelized workflows, optimized data handling, and shared memory buffers for seamless exchange of brain volumes, time-series data, and rtQA metrics. With open-source code available on GitHub, pyOpenNFT advances multimodal real-time neuroimaging and extends the platform for scientific, clinical and educational applications.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3982_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/OpenNFT/pyOpenNFT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{AntEka_pyOpenNFT_MICCAI2025,
        author = { Antipushina, Ekaterina AND Davydov, Nikita AND De Feo, Riccardo AND Prilepin, Evgeny AND Nikonorov, Artem AND Koush, Yury},
        title = { { pyOpenNFT: an open-source Python framework for ML-based real-time fMRI and EEG-fMRI neurofeedback } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {574 -- 583}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    5、This work presents a fully Python-based open-source framework, [AnonimSoft], designed to support real-time fMRI and EEG-fMRI neurofeedback applications. By integrating FastAPI and Lab Streaming Layer, it enables real-time processing of EEG data streams and fMRI signal prediction, enhancing the flexibility, modularity, and real-time efficiency of neurofeedback systems.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The authors made improvements based on OpenNFT, an open-source software for rt-fMRI NF training and quality assessment. Not only did they completely implement it in Python, replacing MATLAB, but they also introduced [AnonimSoft], which significantly optimized real-time data processing and machine learning applications.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1) Although the authors claim that their open-source code [AnonimSoft] has been released on GitHub, I was unable to locate it. No GitHub link or source code is provided. Moreover, there is a lack of detailed software architecture diagrams, training parameters, and descriptions of model saving/loading mechanisms, which significantly hinders reproducibility. 2) In the introduction, the authors state that their method aims to predict local fMRI activity using EEG signals. However, it is unclear whether this refers to predicting the BOLD signal of arbitrary individual voxels. According to recent work by Semenkov et al., EEG-to-fMRI approaches have so far only managed to estimate the average BOLD signal in seven bilaterally symmetrical subcortical regions. The description in the introduction is rather confusing and lacks clarity on this point. 3) The authors mention using a basic ridge regression model in the introduction. However, the use of ridge regression to reconstruct fMRI signals from EEG time and frequency features—especially in visual and subcortical regions—is an already well-established and dated approach. It is unclear why such a basic model is still being used as a primary example in this work. 4) There is a substantial discrepancy between the temporal resolutions of fMRI (with a TR of 2s) and EEG (sampled at 250 Hz). However, the manuscript lacks a clear description of how cross-modal alignment and delay modeling are handled. This omission raises concerns about the reliability of the prediction results. 5) While the Stockwell transform does retain some time-frequency localization, its resolution is fixed and may not be well-suited for highly non-stationary signals such as EEG. The choice of this transform appears limited, and alternative methods (e.g., wavelet transforms or empirical mode decomposition) may offer more adaptive representations. 6) The authors claim to have selected “experiments with better predictive performance” to construct high-predictability voxel clusters. Such post-hoc selection introduces a high risk of data leakage and overfitting, potentially inflating performance and compromising generalizability on unseen data. 7) The manuscript reports only Pearson correlation coefficients (r = 0.28 or 0.45) as evaluation metrics. Other important metrics such as Mean Squared Error (MSE), R², and Mean Absolute Error (MAE) are not included, limiting the interpretability of the results. It is also unclear whether any normalization, such as z-score standardization, was applied to the time series. 8) Although the authors compare processing time with OpenNFT, there is no comparison of model prediction performance against traditional neurofeedback decoding methods such as CCA or PCA-LSTM. Furthermore, the paper lacks ablation studies to evaluate the individual contributions of each component in the proposed framework. 9) The manuscript does not include visualizations comparing the predicted and ground truth fMRI time series (e.g., overlay plots or error curves). Additionally, there are no spatial activation maps showing whether the predicted signals correspond to task-relevant brain regions. These omissions significantly reduce the persuasive power of the presented results.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    12、Although the paper presents a promising method for real-time EEG-to-fMRI prediction using the [AnonimSoft] framework, the evaluation is limited to Pearson’s correlation and does not compare with other advanced methods, nor does it include ablation studies. Additionally, the post-hoc selection of experimental groups may lead to overfitting and data leakage. Finally, the paper does not show a visual comparison between the predicted and actual fMRI time-series, nor does it provide activation maps. These issues suggest that the paper needs improvement in clarity and robustness. Specific feedback can be referenced in the shortcomings mentioned above.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors developed the Python-based toolbox for real-time fMIR and EEG-fMRI neurofeedback and show its potential efficacy.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Overall, the paper is very well written, effectively conveying the work that the authors did.

    The toolbox appears to be systematically developed based on the description of the developed toolbox.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    For this kind of EEG-informed, predicted fMRI signal-based NF to be successful, the prediction performance is crucial.

    • Why do authors think that Pearson’s r = 0.45 was reported in [10], while Pearson’s r = 0.28 was achieved in the authors’ paper using the dataset in [9]?

    • Can authors elaborate and/or provide a clear idea on the performance of the currently state-of-the-art technique to get an informed/predicted fMRI signal using EEG?

    • For example, what has been achieved so far, and what still needs to be done?

    • What applications/use cases are applicable using currently developed techniques?

    • What is not possible and/or limited even using the most advanced technique?

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    I do not fully agree with the description, “… its complex experimental setup..” for fMRI study. I think the experimental setup is similar to that of alternative neuroimaging modalities, such as EEG and MEG.

    “Of note, we found that MATLAB and Python implementations of linear algebra … Cholesky decomposition of ill-conditioned matrices.”

    • Can authors elaborate on this a bit more?
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I am well aware of the topics and the techniques developed.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper develops a fully Python-based open-source framework called [AnonimSoft] for real-time fMRI data processing and EEG-based fMRI neurofeedback signal prediction. The framework is highly flexible, modular, and efficient in real-time processing. It achieves efficient real-time workflows through optimized data handling and shared memory buffers. Additionally, it integrates a machine learning-based prediction server to predict fMRI neurofeedback signals from EEG records, extending the scope of multimodal real-time neuroimaging and providing a platform for scientific, clinical, and educational applications.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Innovation：The framework is fully implemented in Python, which is more flexible and extensible compared to previous frameworks relying on MATLAB. It optimizes real-time data processing workflows, eliminates MATLAB dependencies while maintaining computational rigor. Multimodal Integration：By integrating EEG-fMRI machine learning models, it enables the prediction of fMRI neurofeedback signals from EEG data. This compensates for the low spatial resolution of EEG and leverages the high spatial resolution of fMRI, providing a more balanced signal for neurofeedback training and expanding the scope of neurofeedback therapy. Efficiency：It uses shared memory buffers and parallelized workflows to achieve efficient handling and processing of large-scale data, enabling real-time processing of fMRI and EEG data for neurofeedback training.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. Although the paper proposes a promising framework, the clinical validation is not sufficient. It only mentions performance testing on simulated data, but lacks real-world clinical application cases and effectiveness evaluations, failing to fully demonstrate its feasibility and effectiveness in clinical practice.
    2. The paper does not specify the compatibility of the framework with existing neurofeedback systems or clinical workflows. In practical applications, a newly developed framework needs to seamlessly integrate with existing equipment and technologies, but this aspect is not addressed in the paper, which may limit its adoption in clinical settings.
    3. In the section describing the use of machine learning models for EEG-fMRI signal prediction, although some related studies and model types are mentioned, the details of the model construction, training, and optimization processes are not adequately described. For example, the parameter selection of the model and the preprocessing of training data are not clearly explained, which may affect the understanding and reproducibility of this part by other researchers.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The major strengths of the paper are its innovation and multimodal integration capabilities. The fully Python-based framework developed for real-time fMRI data processing and EEG-based fMRI neurofeedback signal prediction represents a significant advancement in neurofeedback research. Additionally, the framework’s efficiency and modular design offer good scalability and application potential.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers for highlighting possible sources of confusion and providing the opportunity to clarify our text. Reviewer #1 To address Reviewer #1 and the ML-related comments from all reviewers, we clarified that EEG-based predictors of fMRI neurofeedback (NF) signals using classical ML—particularly the EEG Finger-Print (EFP) method—have already been successfully deployed in NF studies (Singer et al.: Development and validation of an fmri-informed eeg model of reward-related ventral striatum activation, 2023; Meir-Hasson et al.: One-class fmri-inspired eeg model for self-regulation training, 2016; Meir-Hasson et al.: An eeg finger-print of fmri deep regional activation, 2014). It demonstrates the feasibility of inferring spatially accurate signals from EEG with an estimator trained to predict the average signal in a target ROI (normalized signals). Specifically, each MRI scan is paired with spectral features extracted from the preceding 12 seconds of EEG signal. More details are provided in the Original description of the method [11]. However, further development is needed to improve clustering strategies, enhance generalization to new data, and investigate the potential of deep learning methods in NF contexts. Additionally, we rephrased “complex” as “restrictive” to more accurately describe the limitations imposed by strong magnetic fields and the requirement for physical immobility (see Introduction section). Regarding the discrepancy between the reported Pearson’s r = 0.45 in [10] and the Pearson’s r = 0.28 obtained in our work using the dataset in [9], we note that the higher correlation in [10] was achieved on a selected cluster of EEG-fMRI runs exhibiting high internal consistency, whereas our evaluation was performed on the full, more heterogeneous open dataset from [9]. Furthermore, while we were able to test the approach on open data, the dataset from [9] did not include the target ROI mask used to extract the fMRI NF signal, but only the full fMRI volumes. This required us to reconstruct the NF target ROI independently, leading to a potential domain discrepancy between our evaluation setup and the regressors originally trained on the dataset by Lioi et al. This likely contributed to the observed lower correlation. For detailed methodological choices, including metric definitions and alignment between modalities, we refer readers to De Feo et al. EEG-to-fMRI Prediction for Neurofeedback: Evaluating Regularized Regression and Clustering Approaches [biorxiv link pending as of 19.05.25] Reviewer #2 We now emphasize that the method used in our demonstration is predicting the mean signal in a target ROI. For the purpose of this demonstration, we used ridge regression for its data efficiency and interpretability, critical in neurofeedback where data availability is limited. Unlike other EEG-fMRI methods, which may show promising results but have not been deployed in NF studies, methods based on regularized regressors have been applied in a number of NF studies in recent years (Singer et al., 2023; Keynan et al., 2019; Meir-Hasson et al., 2016; Meir-Hasson et al., 2014). While we offer more details about ML models, for a full description of our reimplementation of the EEG-fMRI predictors, we should refer the reader to the original description of the method (Meir-Hasson 2014, Meir-Hasson 2016). We further introduced a reference in the final manuscript to a separate paper, including a thorough evaluation of our reimplementation of the method on freely available data, and open-source code. Reviewer #4 In response, we clarified in Section 3 “Simulation results” that we utilized real-world retrospective data to simulate a real-time setting. Additionally, we revised Section 2 “Materials and methods” to highlight our focus on openness, modularity, and flexibility, aiming to provide a highly adaptable software solution that can be integrated into a wide range of workflows, including but not limited to clinical application



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    Based on the comments from reviewer #1 and #4, I recommend acceptance of the paper. In particular, I would like to acknowledge the significant potential in terms of methodological innovation and multimodal data integration. However, I advise the authors to still address the comments raised by the two reviewers in terms of clinical validation to facilitate the application of this tool in clinical practice.



back to top