Fully 3D Unrolled Magnetic Resonance Fingerprinting Reconstruction via Staged Pretraining and Implicit Gridding

Magnetic Resonance Fingerprinting (MRF) enables rapid quantitative imaging, but high-resolution 3D reconstructions remain computationally expensive due to the NUFFTs required at every iteration, and the commonly used Locally Low Rank (LLR) regularization becomes ineffective at high acceleration. Learned 3D priors could address these limitations, but training them at scale is challenging due to memory and runtime constraints. This work proposes SPUR-iG, a fully 3D deep unrolled subspace reconstruction framework that provides fast, high-quality reconstruction for high-resolution non-Cartesian 3D MRF, while keeping training time computationally tractable.

SPUR-iG leverages implicit GROG-based data consistency (DC), which grids non-Cartesian k-space using a learned family of kernels, enabling efficient FFT-based DC with minimal artifacts. To make 3D unrolled training more efficient, we introduce a staged training strategy that keeps computation tractable while progressively improving reconstruction quality. We evaluate the method on a large in vivo dataset, as well as on cross-vendor out-of-distribution data.

At 1 mm isotropic resolution, SPUR-iG outperforms LLR and a state-of-the-art hybrid 2D-3D unrolled baseline in subspace coefficient quality and $T_1$ / $T_2$ accuracy. Whole-brain reconstructions complete in under 15 s, providing up to a 111 $\times$ speedup for 2-min scans relative to LLR. Notably, SPUR-iG reconstructions from 30-s acquisitions achieve mean $T_1$ accuracy that matches or exceeds the mean accuracy of LLR reconstructions from 2-min acquisitions.

SPUR-iG introduces a fully 3D unrolled reconstruction framework for MRF that improves both reconstruction speed and accuracy, making high-resolution accelerated 3D MRF more practical for research and clinical use.