Intro

Note

This document is a personal notebook of my experience with OpenNFT. It is not the official tutorial or intended to be a tutorial, but rather a collection of notes that I took while learning how to work with OpenNFT.

Summary of the paper

This section of the notebook provides a summary of the paper titled “OpenNFT: An open-source Python/Matlab framework for real-time fMRI neurofeedback training based on activity, connectivity and multivariate pattern analysis”.

Introduction

The motivation of the paper is to leverage neurofeedback to help participants learn voluntary control over their own brain activity and connectivity via operant conditioning. The superiority of such approach lies in the fact that it is drug free, non-invasive, and effective in treating a variety of diseases.

The authors compares neurofeedback based on rt-fMRI to conventional neurofeedback study.

  • Conventional neurofeedback study

    • Performance of the participant not evaluated in real-time

    • Challenging to evaluate behavioral/therapeutic effect

    • Range from single-day session to long session over several days

  • Neurofeedback based on rt-fMRI

    • Demands high-performance computer

    • Lacks open-source GUI-based framework

And OpenNFT falls under the category of neurofeedback based on rt-fMRI.

Methods

  1. Neurofeedback data acquisition and transfer

    A data analysis workstation (where OpenNFT is installed) is added to the same network as the MR scanner and reconstruction console.

    • MR hardware and software export acquired and reconstructed data to a shared folder.

    • Data analysis workstation access the exported data via TCP/IP.

  2. Timing and synchronization

  3. Data preprocessing

    If precautions are not taken against movement, spatial alignment and reslicing (via SPM12) are recommended.

  4. Data processing

    • Incremental GLM (iGLM) is used to estimate whole-brian activation maps which is displayed in real-time.

    • Cumulative GLM is used to correct for linear drift and head motions.

    • Low-pass Kalman filter extracts desired signal and discards high-frequency noise.

    • First-order autoregressive model (AR(1)) applied on all voxels’ time-series prior to estimation of the GLM is used account for serial correlations in fMRI data due to non-neurophysiological fluctuations and non-modeled neuronal activity.

    • Dynamic ROI estimation.

  5. Feedback estimation

    1. Constantly displayed (continuous) and periodically displayed (intermittent) activation-based feedbacks

      Single/multiple ROI activity levels are estimated in terms of percent signal change (PSC) after preprocessing. Feedback estimation in this fashion usually involves:

      • Extracting voxel activity using ROI/pattern masks,

      • Obtaining average, weighted average, or eigenvector estimate from the spatial-temporal data sample within the ROI, and

      • Adaptively scaling (normalizing) the time-series (to prevent confusion for the participants).

    2. Intermittent effective connectivity feedback

      Connectivity can estimated as:

      • Temporal correlation (using Pearson correlation coefficient estimation) between two time-series, and

      • Functional connectivity networks using real-time Smooth Incremental Graphical Lasso (rt-SINGLE).

      Dynamic causal modeling (DCM) can also be used to estimate effective connectivity.

    3. Continuous classification-based feedback

      • A supporting vector machine (SVM) classifier is trained prior to neurofeedback run to discriminate between two attention states.

      • Accuracy of classification is determined using n-fold cross validation.

      • The trained classifier is used to provide neurofeedback.

  6. Feedback presentation.

    Feedback is presented visually as simple visual cues, e.g.

    • Thermometer,

    • Moving bar,

    • Avatar faces, and

    • VR-scenes,

    via

    • MR-compatible 2D screens/goggles,

    • MR-compatible 3D displays, and

    • Virtual-reality goggles.