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

Magnetic Resonance Fingerprinting (MRF) enables fast quantitative imaging, yet reconstructing high-resolution 3D data remains computationally demanding. Non-Cartesian reconstructions require repeated non-uniform FFTs, and the commonly used Locally Low Rank (LLR) prior adds computational overhead and becomes insufficient at high accelerations. Learned 3D priors could address these limitations, but training them at scale is challenging due to memory and runtime demands. We propose SPUR-iG, a fully 3D deep unrolled subspace reconstruction framework that integrates efficient data consistency with a progressive training strategy. Data consistency leverages implicit GROG, which grids non-Cartesian data onto a Cartesian grid with an implicitly learned kernel, enabling FFT-based updates with minimal artifacts. Training proceeds in three stages: (1) pretraining a denoiser with extensive data augmentation, (2) greedy per-iteration unrolled training, and (3) final fine-tuning with gradient checkpointing. Together, these stages make large-scale 3D unrolled learning feasible within a reasonable compute budget. On a large in vivo dataset with retrospective undersampling, SPUR-iG improves subspace coefficient maps quality and quantitative accuracy at 1-mm isotropic resolution compared with LLR and a hybrid 2D/3D unrolled baseline. Whole-brain reconstructions complete in under 15-seconds, with up to $\times$111 speedup for 2-minute acquisitions. Notably, $T_1$ maps with our method from 30-second scans achieve accuracy on par with or exceeding LLR reconstructions from 2-minute scans. Overall, the framework improves both accuracy and speed in large-scale 3D MRF reconstruction, enabling efficient and reliable accelerated quantitative imaging.

💡 Research Summary

Magnetic Resonance Fingerprinting (MRF) offers rapid quantitative MRI but reconstructing high‑resolution three‑dimensional (3D) data is computationally intensive, mainly because non‑Cartesian acquisitions require repeated non‑uniform FFTs (NUFFTs) and the commonly used Locally Low Rank (LLR) regularization becomes ineffective at high acceleration factors. This paper introduces SPUR‑iG (Staged Pretraining for Unrolled Reconstruction with implicit GROG), a fully 3D deep‑unrolled reconstruction framework that makes large‑scale non‑Cartesian 3D MRF reconstruction practical.

The core of SPUR‑iG is an implicit GRAPPA‑like gridding (iGROG) module that learns a neural representation of the gridding kernel from calibration data. iGROG maps the irregular k‑space samples onto a Cartesian grid, allowing the data‑consistency (DC) step to be performed with fast FFTs instead of costly NUFFTs. In the experiments a modest 1.5× oversampling and five k‑space points per coil were sufficient to keep interpolation errors low while keeping the gridding time under three seconds.

On the learning side, the framework employs a 3D UNet denoiser conditioned on the unroll iteration (via FiLM). Training proceeds in three stages to overcome memory constraints: (1) pretraining the denoiser on a wide variety of artifact levels (zero‑filled, multiple LLR reconstructions) with extensive data augmentation; (2) greedy unrolled training (GLEAM style) where each unroll iteration is trained separately, detaching the computation graph after each step, and applying geometrically increasing loss weights to emphasize later iterations; (3) full‑unrolled fine‑tuning using gradient checkpointing to keep the full graph in memory while still fitting on a single GPU. This staged approach reduces the overall memory footprint dramatically while still allowing the network to learn the full end‑to‑end reconstruction objective.

The method was evaluated on a large in‑vivo dataset of 45 whole‑brain MRF scans, retrospectively undersampled to simulate acquisition times ranging from 30 seconds to 2 minutes. Compared with LLR and a state‑of‑the‑art hybrid 2D/3D unrolled baseline, SPUR‑iG produced higher‑quality subspace coefficient maps and more accurate T1/T2 maps (average RMSE reductions of 5–10 %). Remarkably, reconstructions from 30‑second scans matched or exceeded the quantitative accuracy of LLR reconstructions from 2‑minute scans. Whole‑brain reconstructions completed in under 15 seconds, representing up to a 111‑fold speed‑up over traditional LLR pipelines. The approach also generalized well to out‑of‑distribution data from a different scanner vendor.

In summary, SPUR‑iG combines an efficient iGROG‑based data‑consistency operator with a progressive three‑stage training scheme to enable fully 3D unrolled MRF reconstruction at 1 mm isotropic resolution within clinically feasible runtimes. The work demonstrates that learned 3D priors, when trained with careful memory‑saving strategies, can surpass handcrafted regularizers both in speed and quantitative fidelity, paving the way for routine high‑resolution quantitative MRI in research and clinical practice.