Plug-and-Play Reconstruction for 3D Non-Cartesian fMRI Data

Functional Magnetic Resonance Imaging (fMRI) captures a temporal sequence of 3D <tex>$T_{2}^{\ast}$</tex> weighted magnetic resonance volumes to study brain activity at rest or during task performance. Consequently, imaging at the required high temporal and/or spatial resolution necessitates highly undersampled k-space acquisitions, making the reconstruction of these volumes severely ill-posed. Yet, due to the lack of ground truth in brain activity, evaluating the impact of reconstruction methods on the statistical analysis of fMRI remains an open problem. In this work, we leverage the SNAKE-fMRI simulator to generate realistic artificial brain activity over a 3D brain template and synthesize the corresponding k-space data, providing a reference ground truth for evaluating reconstruction methods in the fMRI statistical analysis stage. Additionally, we introduce a Plug-and-Play (PnP) approach for image reconstruction that relies on an implicit 3D deep denoising prior. In a benchmark against compressed sensing strategies, we show that PnP-based reconstruction strategies produce high²-quality images but suffer from a loss in statistical sensitivity compared to compressed sensing reconstructions in simulated fMRI studies when a ground truth is available.