The impact of real-time fMRI denoising on online evaluation of brain activity and functional connectivity

Abstract

Objective: Comprehensive denoising is imperative in fMRI analysis to reliably evaluate neural activity from the blood oxygenation level dependent signal. In real-time fMRI, however, only a minimal denoising process has been applied and the impact of insufficient denoising on online brain activity estimation has not been assessed comprehensively. This study evaluated the noise reduction performance of online fMRI processes in a real-time estimation of regional brain activity and functional connectivity. Approach: We performed a series of real-time processing simulations of online fMRI processing, including slice-timing correction, motion correction, spatial smoothing, signal scaling, and noise regression with high-pass filtering, motion parameters, motion derivatives, global signal, white matter/ventricle average signals, and physiological noise models with image-based retrospective correction of physiological motion effects (RETROICOR) and respiration volume per time (RVT). Main results: All the processing was completed in less than 400 ms for whole-brain voxels. Most processing had a benefit for noise reduction except for RVT that did not work due to the limitation of the online peak detection. The global signal regression, white matter/ventricle signal regression, and RETORICOR had a distinctive noise reduction effect, depending on the target signal, and could not substitute for each other. Global signal regression could eliminate the noise-associated bias in the mean dynamic functional connectivity across time. Significance: The results indicate that extensive real-time denoising is possible and highly recommended for real-time fMRI applications.

A library for fMRI real-time processing systems in python (RTPSpy) with comprehensive online noise reduction, fast and accurate anatomical image processing, and online processing simulation

“…These processes can be completed in less than 400 ms for the whole-brain voxels on a current PC with a graphics processing unit (GPU) (Misaki and Bodurka, 2021). Thus, computation time does not limit today's real-time fMRI processing with a decent personal computer.…”

“…While the respiration volume per time (RVT) regressors (Birn et al, 2008) can also be made, we do not recommend using them in online processing. A previous study indicated that the online evaluation of RVT regressors could not be accurate, and its usage could introduce an artifactual signal fluctuation in real-time processing (Misaki and Bodurka, 2021). The RTP_RETROTS should receive cardiac and respiration signals in real-time.…”

“…The RTPSpy noise reduction is performed for the whole-brain voxels, which is advantageous in calculating the feedback signals from multiple regions, such as the functional connectivity and decoding neurofeedback (Watanabe et al, 2017). Misaki and Bodurka (2021) demonstrated the advantage of comprehensive noise reduction in connectivity neurofeedback signals. The calculated signal can be sent to an external application through a socket (Supplementary figures S1 and S2).…”

“…Nevertheless, a complex system setup specific to an individual environment and noisy online evaluation of neural activation due to a limited real-time fMRI signal processing have hindered the utility of rtfMRI applications and reproducibility of its result (Thibault et al, 2018). Indeed, the significant risk of noise contamination in the neurofeedback signal has been demonstrated in recent studies (Weiss et al, 2020;Misaki and Bodurka, 2021). These issues have been addressed with a community effort releasing easy-to-use rtfMRI frameworks (Cox et al, 1995;Goebel, 2012;Sato et al, 2013;Koush et al, 2017;Heunis et al, 2018;MacInnes et al, 2020) and consensus on reporting detailed online processing and experimental setups (Ros et al, 2020).…”

“…Instead, the goal of the RTPSpy is to provide building blocks for making a highly customized and advanced rtfMRI system. RTPSpy is a library package including 1) a fast and comprehensive online fMRI denoising pipeline comparable to offline processing (Misaki and Bodurka, 2021), 2) utilities for fast and accurate anatomical image processing to define target regions on-site, 3) a simulation system of online fMRI processing to optimize processing pipeline and target signal calculation, 4) interface to an external application to exchange real-time brain activation signals, and 5) a boilerplate graphical user interface (GUI) application integrating operations with RTPSpy. We believe that the independent processing components provided by RTPSpy can be more useful for developing a custom rtfMRI system than a tightly packed framework.…”