StoDIP: Efficient 3D MRF image reconstruction with deep image priors and stochastic iterations

TL;DR

This paper addresses the challenge of reconstructing quantitative 3D MRF maps without ground-truth data. It introduces StoDIP, a ground-truth-free Deep Image Prior framework that enforces k-space consistency through a generator G_ heta applied to an initial estimate, hat{x}=G_\theta(\mathbf{x}^{(0)}), and minimizes \mathcal{L}(\theta)=||\sqrt{DCF}\cdot \mathbf{y}-\sqrt{DCF}\cdot A(\hat{\mathbf{x}})||_2^2 using A that includes coil sensitivities, NUFFT, and temporal SVD, with stochastic coil-wise updates to reduce memory demands. Key innovations include memory-efficient stochastic updates, evaluation of backbone architectures (DIP and DRUNet), and optional TV regularization to suppress artefacts, with per-coil losses defined as $L_c=||\sqrt{DCF}\cdot \mathbf{y}_c-\sqrt{DCF}\cdot A_c(\hat{\mathbf{x}})||_2^2$. Implemented in PyTorch on an RTX 4090, StoDIP uses DRUNet with an adaptive LR schedule and processes real-valued inputs of size $N\times 2K$ over 500 epochs, typically achieving competitive performance against baselines on whole-brain 3D MRF data accelerated by retrospective 2x subsampling; future work includes Bloch-consistency losses and leveraging multi-scan data for further gains.

Abstract

Magnetic Resonance Fingerprinting (MRF) is a time-efficient approach to quantitative MRI for multiparametric tissue mapping. The reconstruction of quantitative maps requires tailored algorithms for removing aliasing artefacts from the compressed sampled MRF acquisitions. Within approaches found in the literature, many focus solely on two-dimensional (2D) image reconstruction, neglecting the extension to volumetric (3D) scans despite their higher relevance and clinical value. A reason for this is that transitioning to 3D imaging without appropriate mitigations presents significant challenges, including increased computational cost and storage requirements, and the need for large amount of ground-truth (artefact-free) data for training. To address these issues, we introduce StoDIP, a new algorithm that extends the ground-truth-free Deep Image Prior (DIP) reconstruction to 3D MRF imaging. StoDIP employs memory-efficient stochastic updates across the multicoil MRF data, a carefully selected neural network architecture, as well as faster nonuniform FFT (NUFFT) transformations. This enables a faster convergence compared against a conventional DIP implementation without these features. Tested on a dataset of whole-brain scans from healthy volunteers, StoDIP demonstrated superior performance over the ground-truth-free reconstruction baselines, both quantitatively and qualitatively.

Figures (2)

• Figure 1: Model performance during training for k-space (measurement consistency, MC) loss, (solid lines) averaged across the eight coils, and the T1 + T2 MAPEs (dashed lines, using only for monitorning, not for training) for different subsets of experiments, each one showcasing a behaviour for a particular case of study: a) architecture choice, b) different $\textbf{x}^{(0)}$, c) stochasticity and choice of LR, and d) early stopping and the effect of TV spatial regularisation.
• Figure 2: Reconstructed T1 and T2 maps for the assessed approaches on data with (retrospective) scan-time acceleration factor of 2. All maps have been masked using BETsmith2002bet. Right panels show zoomed in regions (electronic zoom recommended).