High-Quality 0.5mm Isotropic fMRI: Random Matrix Theory Meets Physics-Driven Deep Learning

Abstract

Submillimeter fMRI plays a vital role in studying the brain function at the mesoscale level, allowing investigation of functional activity in small cortical structures. However, such resolutions require extreme trade-offs between SNR, spatio-temporal resolution and coverage leading to numerous challenges. Therefore, interpretable locally low-rank denoising methods based on random matrix theory have been proposed and built into fMRI pipelines, but they require well-characterized noise distributions on reconstructed images, which hinders the use of emerging physics-driven deep learning reconstructions. In this work, we re-envision the conventional fMRI computational imaging pipeline to an alternative where denoising is performed prior to reconstruction. This allows for a synergistic combination of random matrix theory based thermal noise suppression and physics-driven deep learning re-construction, enabling high-quality 0.5mm isotropic functional MRI. Our results show that the proposed strategy improves on denoising or physics-driven deep learning reconstruction alone, with better delineation of brain structures, higher tSNR particularly in mid-brain areas and the largest expected extent of activation in GLM-derived t-maps.