Denoising Diffusion Probabilistic Models for Magnetic Resonance Fingerprinting

TL;DR

This work addresses fast MRF reconstruction by introducing MRF-IDDPM, a conditional diffusion model that reconstructs high-quality MRF time series from undersampled data. It leverages subspace compression in the diffusion process, batch-wise patch training, and an IDDPM-based fast sampling scheme to produce accurate T1 and T2 maps while offering uncertainty estimates through sample variance. Across in-vivo brain data with 5× acceleration, MRF-IDDPM outperforms SVDMRF, LRTV, and DRUNet baselines in TSMI and parameter-map quality, as evidenced by MAPE, RMSE, NRMS, and SSIM metrics. The approach achieves practical runtimes on modest GPUs and opens avenues for 3D extensions and integration of Bloch-constraint losses during reconstruction, enabling more reliable quantitative MRI in clinical workflows.

Abstract

Magnetic Resonance Fingerprinting (MRF) is a time-efficient approach to quantitative MRI, enabling the mapping of multiple tissue properties from a single, accelerated scan. However, achieving accurate reconstructions remains challenging, particularly in highly accelerated and undersampled acquisitions, which are crucial for reducing scan times. While deep learning techniques have advanced image reconstruction, the recent introduction of diffusion models offers new possibilities for imaging tasks, though their application in the medical field is still emerging. Notably, diffusion models have not yet been explored for the MRF problem. In this work, we propose for the first time a conditional diffusion probabilistic model for MRF image reconstruction. Qualitative and quantitative comparisons on in-vivo brain scan data demonstrate that the proposed approach can outperform established deep learning and compressed sensing algorithms for MRF reconstruction. Extensive ablation studies also explore strategies to improve computational efficiency of our approach.

Figures (5)

• Figure 1: Reconstructed T1 (left panel) and T2 (right panel) maps using different methods (columns) for three representative brain slices from the evaluation set. Every three rows correspond to one brain slice, displaying (from top to bottom): the full reconstructed map, a zoomed-in view of a region of interest (green box), and the corresponding percentage error map. Electronic zooming is recommended for detailed inspection.
• Figure 2: Magnitudes of the reference and reconstructed TSMIs (SVD compressed to $s=5$) from one representative slice in the evaluation dataset, using different methods. Methods aim to restore the corrupted input (condition image) from a gridding reconstruction.
• Figure 3: Effect of stride size on the reconstruction time (black curve), as well as the reconstruction NMRSE (left) and SSIM (right) metrics for TSMI (bottom) and averaged T1 and T2 maps (top, labelled as (T1+T2)/2). Metrics are averaged over the test dataset.
• Figure 4: Effect of timesteps (K) on the reconstruction time (black curve), as well as the reconstruction NMRSE (left) and SSIM (right) metrics for TSMI (bottom) and averaged T1 and T2 maps (top). Metrics are averaged over the test dataset.
• Figure 5: Average, standard deviation (STD), and absolute error maps for T1 (left panel) and T2 (right panel) reconstructions of three representative brain slices in the evaluation set. Each row corresponds to a different brain slice. The MRF-IDDPM algorithm was used to generate 10 reconstructions per slice, each initiated with a different random noise $\mathbf{x}_T$. Average and STD maps show the pixel-wise mean and standard deviation across these 10 reconstructions. Absolute error maps display the absolute difference between the average reconstruction and the ground truth reference from Figure \ref{['fig:tissue_maps']}.