Deep learning based image enhancement for dynamic non-Cartesian MRI: Application to "silent" fMRI.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence

Background:

• Silent functional MRI (fMRI) is crucial for noise-sensitive environments but often suffers from reduced image quality due to undersampled k-space.
• Deep learning (DL) offers potential solutions for enhancing image quality in undersampled MRI data.

Purpose of the Study:

• To investigate the efficacy of DL-based image enhancement for silent fMRI, focusing on preserving temporal signal dynamics.
• To compare 2D and 3D UNet architectures for enhancing undersampled 3D radial fMRI data.

Main Methods:

• Constructed a ground-truth dataset using Human Connectome Project (HCP) resting-state fMRI data.
• Simulated undersampled non-Cartesian k-space data using a 'Looping Star' sequence trajectory.
• Trained and compared 2D-UNet and 3D-UNet models for image enhancement.

Main Results:

• The 3D-UNet model significantly outperformed the 2D-UNet, achieving a 97% reduction in mean square error.
• The 3D-UNet successfully preserved temporal signal variations in resting-state fMRI (rs-fMRI) data, including correlated activity in the posterior cingulate cortex (PCC).
• Noise was effectively mitigated, and image quality was improved.

Conclusions:

• Deep learning, particularly 3D-UNet, is effective for enhancing silent fMRI image quality and preserving temporal BOLD signal characteristics.
• 3D convolutions are advantageous over deeper 2D networks for handling global artifacts in fMRI.
• The developed DL approach shows promise for improving fMRI studies in challenging acoustic environments.