Signal quality and Bayesian signal processing in neurofeedback based on real-time fMRI

• Published in: NeuroImage (Orlando, Fla.) Vol. 59; no. 1; pp. 478 - 489
• Main Authors: Koush, Yury, Zvyagintsev, Mikhail, Dyck, Miriam, Mathiak, Krystyna A., Mathiak, Klaus
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 02.01.2012; Elsevier Limited
• Subjects: Adult; Algorithms; Bayes Theorem; Bayesian; Brain; Brain - physiology; Brain Mapping - methods; Electrostatic discharges; Feedback; Female; Humans; Kalman filter; Magnetic Resonance Imaging - methods; Male; Neurofeedback; Neurofeedback - methods; Neurofeedback - physiology; Noise; Real-time fMRI; Signal processing; Signal Processing, Computer-Assisted; Signal quality; Signal-To-Noise Ratio; Time series; Signal processing; Real-time fMRI; Signal quality; Kalman filter; Neurofeedback; Bayesian
• ISSN: 1053-8119; 1095-9572; 1095-9572
• DOI: 10.1016/j.neuroimage.2011.07.076

Real-time fMRI allows analysis and visualization of the brain activity online, i.e. within one repetition time. It can be used in neurofeedback applications where subjects attempt to control an activation level in a specified region of interest (ROI) of their brain. The signal derived from the ROI is contaminated with noise and artifacts, namely with physiological noise from breathing and heart beat, scanner drift, motion-related artifacts and measurement noise. We developed a Bayesian approach to reduce noise and to remove artifacts in real-time using a modified Kalman filter. The system performs several signal processing operations: subtraction of constant and low-frequency signal components, spike removal and signal smoothing. Quantitative feedback signal quality analysis was used to estimate the quality of the neurofeedback time series and performance of the applied signal processing on different ROIs. The signal-to-noise ratio (SNR) across the entire time series and the group event-related SNR (eSNR) were significantly higher for the processed time series in comparison to the raw data. Applied signal processing improved the t-statistic increasing the significance of blood oxygen level-dependent (BOLD) signal changes. Accordingly, the contrast-to-noise ratio (CNR) of the feedback time series was improved as well. In addition, the data revealed increase of localized self-control across feedback sessions. The new signal processing approach provided reliable neurofeedback, performed precise artifacts removal, reduced noise, and required minimal manual adjustments of parameters. Advanced and fast online signal processing algorithms considerably increased the quality as well as the information content of the control signal which in turn resulted in higher contingency in the neurofeedback loop. ► Real-time fMRI can analyze and visualize localized brain activity online. ► The feedback signal is contaminated with linear and non-linear noise. ► Signal quality analysis emphasized the necessity for additional signal processing. ► Bayesian approach reduced noise and removed artifacts in real-time. ► Instantaneous CNR was around 1 even after filtering suggesting the need of averaging.