Highly-Accelerated High-Resolution Multi-Echo fMRI Using Self-Supervised Physics-Driven Deep Learning Reconstruction

Functional MRI (fMRI) is a critical tool for visualizing neural activities in the brain. fMRI analysis requires comprehensive coverage with high spatiotemporal resolution. To this end, a combination of simultaneous multi-slice imaging and in-plane acceleration is commonly used. However, conventional reconstructions are based on linear methods, leading to noise amplification and aliasing at high accelerations. In particular, the emerging class of multi-echo (ME)-fMRI techniques, which acquire the same imaging location at multiple echo times after a single excitation and offer the potential for further quantification, require higher acceleration rates, beyond what is achievable with conventional methods. In the broader MRI community, deep learning (DL) techniques have been proposed for improved image reconstruction at higher accelerations. While the conventional supervised training paradigms are not applicable to fMRI due to the lack of fully-sampled reference data, we have previously shown that self-supervised DL methods are feasible for highresolution accelerated fMRI. In this study, we adapt these strategies for ME-fMRI to enable prospective 20-fold acceleration with high-resolution and whole-brain coverage with 3 echoes at 7T. Our network leverages the T-2(*) correlations between multiple echoes. Results indicate the feasibility of high-resolution 20-fold accelerated whole-brain ME-fMRI, leading to neural activation maps consistent with the expected activation patterns.