fMRI data analyses are often directed to disentangling and understanding the neural processes that occur among brain regions.
Interactions in the brain are expressed, not at the level of hemodynamic responses, but at the neural level. Thus, an intermediate step that estimates the underlying neuronal activity is necessary for such analyses.
Given the nature of the fMRI BOLD signal, the appropriate approximation of the neuronal activity can be obtained by means of deconvolution with an assumed hemodynamic response (Gitelman et al. (2003)).
Simply put, deconvolution methods are capable of blindly estimating neuronal activity with no prior information on the timing of the BOLD events.
Given the nature of the fMRI BOLD signal, the appropriate approximation of the neuronal activity can be obtained by means of deconvolution with an assumed hemodynamic response (Gitelman et al. (2003)).
We propose a new approach (LR+MV-SPFM) for the spatio-temporal deconvolution of fMRI data that can simultaneously estimate global signal fluctuations and neuronal-related activity based on the low-rank plus sparse matrix decomposition method (Otazo et al. (2015)).
We present a novel formulation for the deconvolution of BOLD fMRI data that captures global fluctuations due to motion artifacts or physiological signals.
Simply put, deconvolution methods are capable of blindly estimating neuronal activity with no prior information on the timing of the BOLD events.
More specifically, I am currently working on methods for the spatiotemporal deconvolution of fMRI data (both single-echo and multi-echo fMRI) and tensor decomposition methods for the denoising of multi-echo fMRI data.
These global events are difficult to compensate during data preprocessing (Power et al. (2018)) and can be misinterpreted as neuronally related since their temporal signature can closely resemble the hemodynamic response function (HRF) assumed by the deconvolution model to describe neurovascular coupling.
The performance of existing deconvolution approaches can be hampered considerably in presence of global, widespread signal changes due to head jerks, hardware artifacts, or prominent non-neuronal physiological events (e.g. deep breaths) (Power et al. (2017)).