NON-CARTESIAN SELF-SUPERVISED PHYSICS-DRIVEN DEEP LEARNING RECONSTRUCTION FOR HIGHLY-ACCELERATED MULTI-ECHO SPIRAL...

Area of Science:

• Neuroimaging
• Magnetic Resonance Imaging (MRI)

Background:

• Multi-echo fMRI enhances brain function quantification by sampling multiple echo times.
• Non-Cartesian trajectories, like spiral acquisitions, offer denser sampling but require high acceleration rates.
• Current Cartesian trajectory methods limit spatiotemporal resolution in multi-echo fMRI.

Purpose of the Study:

• To develop a physics-driven deep learning (PD-DL) reconstruction method.
• To accelerate multi-echo spiral fMRI acquisitions by 10-fold.
• To improve spatiotemporal resolution and enable meaningful Blood-Oxygen-Level-Dependent (BOLD) analysis.

Main Methods:

• Utilized a physics-driven deep learning (PD-DL) reconstruction framework.
• Modified a self-supervised learning algorithm for non-Cartesian trajectories.
• Trained the PD-DL network using the modified self-supervised approach for 10-fold acceleration.

Main Results:

• Achieved high spatiotemporal resolution in multi-echo spiral fMRI.
• Demonstrated the effectiveness of the self-supervised PD-DL reconstruction for accelerated acquisitions.
• Obtained meaningful BOLD signal analysis results with the proposed method.

Conclusions:

• The proposed self-supervised PD-DL reconstruction effectively accelerates multi-echo spiral fMRI.
• This method enhances image quality and enables robust BOLD analysis.
• PD-DL offers a promising approach for advanced neuroimaging applications.