Automatic Detection of BOLD Activations using a Semi-supervised Bidirectional LSTM Neural Network

Motivation: Because there is significant variability in the hemodynamic response of individuals, fitting a simple function may lead to false negatives (i.e., a bad fit leads to larger residuals than the noise level would dictate). Goal(s): Develop a robust data-driven based approach for detecting BOLD. Approach: A semi-supervised automatic detection (SAD) method based on a bidirectional long/short-term memory neural network to find BOLD responses in the entire brain and assess classification performance on simulated fMRI data. Results: The proposed detection method exhibits robustness across various HRF shapes at realistic contrast-to-noise ratios. Impact: We proposed a method for detecting BOLD responses in high temporal resolution fMRI data that is based on a Bidirectional long/short-term memory neural network. Classification performance was excellent as tested using simulated data with different HRFs and contrast-to-noise ratios.