Rapid decoding of neural information representation from ultra-fast functional magnetic resonance imaging signals

Abstract

High spatio-temporal resolution is crucial for neuroimaging techniques to improve our understanding of human brain function. While the fMRI signal is slow and shows a spread in latencies over space, the precision of hemodynamic response latency for each voxel is preserved and has been shown to be able to detect oscillatory hemodynamic changes approaching 1 Hz, suggesting its potential to reveal rapid neural dynamics. To examine how fast neural information can be derived from fMRI signals, we performed experiments that acquire high-field (7T) fMRI signals at an ultra-fast sampling rate (TR = 125 ms) from the visual cortex while participants observed naturalistic object stimuli. We applied multivariate pattern decoders to extract presented object-category information from the acquired signals at each sampling time after stimulus onset. Results showed that decoding accuracy rose above statistical significance less than 2 s after signal onset, faster than the peak latency of the hemodynamic response. The peak latency of the decoding accuracy was independent of variations in the hemodynamic latency of voxels used for decoding. The application of sparse decoders further revealed that rapid and accurate decoding was possible by pruning vein-rich voxels off from the multivariate voxel input to the decoders. These results suggest that a combination of ultra-fast sampling and multivariate decoding allows fast and temporally precise analysis of neural activity using fMRI signals. Significance statementFunctional magnetic resonance imaging (fMRI) is the most successful method to evaluate human brain function at fine spatial scales but is thought to lack temporal resolution because of slow hemodynamics. We challenge this conventional notion by fast sampling of fMRI signals, combined with multivariate decoding to extract information content represented in the fMRI signals. Results showed that the information content can be extracted faster than the magnitude of hemodynamic response rises, independent of the large spatial variation of hemodynamic latencies across individual voxels, and even after removing venous voxels from decoders. Our method is thus effective at filtering out slow and variable hemodynamic components, leading to the extraction of rapid and temporally precise components reflecting neuronal activity in human fMRI signals.