Decoding fMRI activity in the time domain improves classification performance.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning

Background:

• Current functional Magnetic Resonance Imaging (fMRI) decoding methods use statistical summaries, which are inadequate for analyzing events with short intervals.
• Standard deconvolution approaches struggle with closely spaced stimulation events, limiting decoding capabilities.

Purpose of the Study:

• To develop a novel framework for fMRI decoding that continuously analyzes time-series data, scan-by-scan.
• To overcome the limitations of standard methods in decoding events with short inter-stimulus intervals.
• To enable decoding of fMRI data not typically suitable for analysis and studies with reduced scan times.

Main Methods:

• A new framework was developed to separate spatial and temporal components for continuous fMRI time-series decoding.
• The proposed model decodes fMRI data scan-by-scan, reconstructing the time series.
• Stimulation events are identified via deconvolution of the reconstructed time series.

Main Results:

• The novel framework performs comparably to or better than standard fMRI decoding approaches.
• The model shows superior performance in decoding regimes with short inter-stimulus intervals (3-5s).
• Predictions generated by the new framework are highly interpretable in the time domain.

Conclusions:

• The proposed continuous, scan-by-scan fMRI decoding framework enhances analysis capabilities, especially for rapid events.
• This approach expands the scope of fMRI data that can be decoded and allows for shorter scan durations.
• The method offers interpretable temporal predictions, advancing neuroimaging analysis.