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

TL;DR

SPUR-iG presents a fully 3D unrolled reconstruction framework for Magnetic Resonance Fingerprinting that integrates efficient iGROG-based data consistency with a learned 3D denoiser and a staged pretraining strategy. By decomposing the MRF signal into a low-dimensional subspace and replacing traditional priors with a scene-aware 3D neural prior, the method achieves dramatic speedups (up to ×111) and improved quantitative accuracy, enabling whole-brain 1 mm isotropic reconstructions in under 15 seconds. A three-stage training regime—denoiser pretraining, greedy per-iteration unrolled training, and full unrolled fine-tuning with gradient checkpointing—makes large-scale 3D unrolled learning feasible within reasonable compute budgets. Across in vivo data and cross-vendor tests, SPUR-iG consistently improves subspace coefficient quality and $T_1$/$T_2$ maps relative to LLR and 2D/3D baselines, highlighting its potential to make accelerated quantitative 3D MRI more practical for clinical and research use.

Abstract

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.

Figures (8)

• Figure 1: Illustration of inference (top) and training (bottom) pipelines of SPUR-iG. During inference, iGROG grids the non-Cartesian data onto a Cartesian grid using an implicit kernel representation learned from the calibration region. The gridded k-space is then passed to the unrolled reconstruction, initialized with ${\bf \cal A}^H {\bf b}$. Reconstruction proceeds for $N$ unroll iterations, each alternating between a DC update with a learnable step size $\mu_i$ and a step-conditioned denoiser. Training is performed in three stages: (1) a step-conditioned UNet is pretrained on inputs with varying artifact levels, using a heuristic mapping between input type and unroll step to promote robustness across noise/artifact conditions; (2) the pretrained UNet initializes GLEAM-style training, where losses are computed and backpropagated after each unroll iteration, detaching subsequent steps; and (3) full unrolled training, initialized from stage (2), fine-tunes the model under the exact inference setting, with gradient checkpointing (GC) across DC and denoising steps to reduce memory consumption.
• Figure 2: Example reconstructions of the first balanced subspace coefficient (a) and third unbalanced coefficient (b) for two test subjects. The top row shows the 6 min reference (left) and ${\bf \cal A}^H{\bf b}$ for the 30 s and 2 min acquisitions (middle and right) as initialization references. Subsequent rows correspond to reconstructions at 2 min, 1 min, and 30 s, with columns comparing our method, the hybrid 2D/3D variant and LLR. PSNR and SSIM are reported in the bottom right of each image. Error maps of the top-right quadrant (magnified $\times 10$) are overlaid to illustrate error levels.
• Figure 3: Qualitative examples of reconstructed $T_1$ (a) and $T_2$ (b) maps. The first row shows the 6 min reference (left) and maps from ${\bf \cal A}^H {\bf b}$ for the 30 s and 2 min acquisitions (middle and right) as references. Subsequent rows present reconstructions at 2 min, 1 min, and 30 s, with columns comparing our method, the hybrid 2D/3D variant, and LLR. Error maps of the top-right quadrant are overlaid on each image. Quantitative colormap (top) and error colormap (bottom) are shown to the right.
• Figure 4: Reconstructions of the first balanced subspace coefficient (a) and the third unbalanced coefficient (b) for two slices from the same subject, scanned on an out-of-distribution scanner vendor. The layout follows Fig. \ref{['fig:subspace_combined']}.
• Figure 5: Qualitative examples of reconstructed $T_1$ (a) and $T_2$ (b) maps from an out-of-distribution vendor scan. The layout follows Fig. \ref{['fig:quant_t1t2_ex']}.