Optimal compressed sensing reconstructions of fMRI using 2D deterministic and stochastic sampling geometries

Area of Science:

• Medical Imaging
• Neuroscience
• Signal Processing

Background:

• Compressive sensing (CS) offers a framework for accelerating fMRI acquisition by reconstructing images from limited frequency-domain samples.
• Existing CS studies often use stochastic sampling, which is insufficient for the rapid acquisitions required in fMRI.
• This study introduces a framework to optimize sampling geometry and reconstruction parameters for improved fMRI image quality.

Purpose of the Study:

• To develop a comprehensive optimization framework for determining optimal 2D sampling geometries (stochastic and deterministic) in fMRI.
• To identify optimal reconstruction parameters for guaranteeing high image quality in CS-based fMRI.
• To evaluate different k-space sampling strategies for accelerated fMRI acquisition.

Main Methods:

• Investigated deterministic (dyadic phase encoding, spiral low pass) and stochastic (random phase encoding, random samples on a PDF) 2D k-space sampling geometries.
• Evaluated over 36 sampling geometries at various sampling rates for reconstructing BOLD fMRI ON/OFF images and activity maps.
• Developed an optimization framework to determine optimal parameters and sampling geometry prior to scanning.

Main Results:

• Parameter optimization rapidly converged, yielding significant improvements in image quality.
• Spiral low pass (SLP) geometry achieved excellent reconstruction with only 20.3% of samples, showing high PSNR (57.58 dB) and SSIM (0.9747).
• Median parameter values with SLP geometry also produced excellent ON/OFF image reconstructions (mean SSIM >= 0.93), validated by leave-one-out analysis.

Conclusions:

• Compressive sensing parameter optimization substantially enhances fMRI image reconstruction quality.
• 2D MRI scanning using SLP geometries consistently provided the best reconstruction results, outperforming random sampling.
• Stable parameter regions were identified for achieving specific image quality levels with particular k-space geometries, enabling efficient fMRI acquisition.