Overcoming MRI data acceleration barriers in functional human brain mapping with physics-driven deep learning

Functional MRI (fMRI) data are typically obtained using highly accelerated acquisitions to provide sufficient coverage and resolution, particularly for whole brain coverage with high spatiotemporal resolution. However, high acceleration rates lead to residual aliasing and increased noise amplification. In the broader MRI reconstruction community, physics-driven deep learning (PD-DL) has gained immense interest to tackle this challenge and has shown promising results at higher acceleration rates where conventional methods struggle to maintain high-quality outputs. Yet, PD-DL reconstruction has several challenges that hamper its applicability in fMRI. PD-DL typically requires supervised learning, using fully-sampled training reference data that are not readily available for high-resolution fMRI studies. Furthermore, the impact of PD-DL reconstruction on fMRI activation maps remains to be assessed. In this work, we develop a multi-slice self-supervised PD-DL reconstruction to enhance accelerated fMRI acquisitions. We investigate this DL reconstruction with 20-fold retrospective and prospective accelerations. Our results show that PD-DL reconstruction allows reliable functional mapping at a previously unachievable accelerations, substantially improving on conventional reconstructions and attainable spatiotemporal resolutions. These results have the potential to have profound impact in the field of non-invasive neuroimaging by enabling highly efficient high resolution fMRI acquisitions that are otherwise unachievable. Ömer Burak Demirel, Luca Vizioli and colleagues developed a self-supervised deep learning method that improves highly accelerated functional MRI reconstruction without requiring fully sampled training data. The approach enables whole-brain fMRI at 20-fold acceleration with reduced noise and aliasing, while preserving image quality and functional activation mapping compared with conventional reconstruction methods.