A 3D Convolutional Encapsulated Long Short-Term Memory (3DConv-LSTM) Model for Denoising fMRI Data

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Signal Processing

Background:

• Functional magnetic resonance imaging (fMRI) data are susceptible to noise from head motion, physiological processes, and thermal factors.
• Existing denoising methods often process the entire fMRI time series, limiting their applicability to real-time analysis.

Purpose of the Study:

• To develop a generally applicable, deep learning-based method for denoising fMRI data in real-time.
• To generate noise-free, realistic individual fMRI volumes (time points) using an advanced AI approach.

Main Methods:

• A fully data-driven 3D convolutional encapsulated Long Short-Term Memory (3DConv-LSTM) model was developed.
• An adversarial network regularized the 3DConv-LSTM output, enhancing realism by deceiving a critic network.
• A gate-controlled self-attention mechanism was integrated to manage short-term dependencies and historical data within a memory pool.

Main Results:

• The proposed deep learning method successfully generated noise-free fMRI volumes.
• Evaluations on both task and resting-state fMRI data demonstrated superior performance compared to existing state-of-the-art deep learning techniques.
• Qualitative and quantitative analyses confirmed the effectiveness of the denoising approach.

Conclusions:

• The developed deep learning method offers a robust solution for real-time fMRI denoising.
• This approach significantly improves the quality and realism of fMRI data, outperforming current methods.
• The technique holds promise for advancing real-time fMRI data analysis in neuroscience research.