Subspace-constrained approaches to low-rank fMRI acceleration.

Area of Science:

• Neuroimaging
• Biomedical Engineering

Background:

• Accelerated functional MRI (fMRI) methods reconstruct high-fidelity images from under-sampled k-space.
• This allows for higher temporal resolution, reduced physiological noise, and increased statistical degrees of freedom in fMRI datasets.
• Existing methods like k-t FASTER exploit the low-rank nature of fMRI data.

Purpose of the Study:

• To present a reformulated k-t FASTER approach incorporating L2 constraints within a low-rank framework.
• To evaluate the impact of Tikhonov constraints, low-resolution priors, and temporal subspace smoothness on fMRI reconstruction.
• To assess the robustness of these methods to under-sampling and thermal noise.

Main Methods:

• Reformulated k-t FASTER with added L2 constraints (Tikhonov, low-resolution priors, temporal subspace smoothness).
• Tested on retrospectively and prospectively under-sampled finger-tapping task fMRI data.
• Evaluated reconstruction quality via low-rank subspaces and activation maps.

Main Results:

• L2 constraints consistently improved results, yielding high-fidelity reconstructions at higher acceleration factors and lower SNR values.
• The Tikhonov constraint demonstrated robustness across datasets.
• Temporal subspace smoothness achieved the best scores on prospectively under-sampled data.
• Reconstructions were achieved without model-based spatial constraints.

Conclusions:

• Regularized low-rank reconstruction effectively recovers functional information in fMRI at high acceleration factors.
• L2 constraints enhance reconstruction quality and robustness compared to existing low-rank methods.
• This approach offers a promising direction for advancing fMRI acquisition and analysis.