Physics Informed Guided Diffusion for Accelerated Multi-parametric MRI Reconstruction

Perla Mayo, Carolin M. Pirkl, Alin Achim, bjoern menze, Mohammad GolbabaeeLecture Notes in Computer Science 2025 · DOI · OpenAlex

Introduction

Magnetic Resonance Fingerprinting (MRF) represents a paradigm shift in quantitative MRI, enabling the simultaneous acquisition of multiple tissue properties, such as T1 and T2 relaxation times, within a single scan. Despite its efficiency, clinical adoption is often hindered by the long acquisition times required for high-resolution mapping. Accelerating MRF typically involves significant undersampling in both spatial and temporal domains, which introduces severe artifacts and noise.

Traditional reconstruction methods, such as compressed sensing or standard deep learning (DL) approaches like Convolutional Neural Networks (CNNs), often struggle to balance image fidelity with physical accuracy. While DL models excel at denoising, they frequently operate as "black boxes," potentially generating "hallucinated" features that lack physical basis—a critical risk in medical diagnostics. This paper presents MRF-DiPh (Physics-informed guided Diffusion), a novel framework that integrates the generative power of denoising diffusion models (DDMs) with the rigorous constraints of MRI physics.

Methodology

The MRF-DiPh framework treats the reconstruction of MRF time-series images as a constrained optimization problem. The approach decouples the reconstruction process into three distinct "anchors" using a Half Quadratic Splitting (HQS) algorithm:

  1. The Diffusion Prior: A pretrained DDM serves as a sophisticated image prior, capturing the underlying distribution of high-quality brain images to guide the denoising process.
  2. K-Space Consistency: This ensures that the reconstructed images remain faithful to the raw data acquired by the MRI scanner, maintaining measurement integrity.
  3. Bloch Model Consistency: To prevent physical hallucinations, the framework enforces the Bloch equations—the fundamental laws of nuclear magnetic resonance—via a dictionary-matching step. This ensures that the signal evolution in every pixel is physically plausible.

Technically, the integration is achieved through a single-step Alternating Direction Method of Multipliers (ADMM) within the HQS iterations. This allows the model to leverage the high-quality generative capabilities of diffusion models while remaining strictly grounded in physical reality.

Experimental Design

The study evaluated MRF-DiPh using 2D brain MRF scans from eight healthy volunteers. The dataset was split, with 75% (six subjects) used for training and 25% (two subjects) for testing. To simulate high acceleration, the researchers retrospectively truncated the acquisition sequence from 1,000 time frames down to 200, representing a 5-fold (R=5) acceleration.

The performance of MRF-DiPh was benchmarked against several established methods:

  • SVDMRF: A standard baseline for MRF reconstruction.
  • MRF-ADMM: A traditional compressed sensing approach.
  • SCQ: A state-of-the-art CNN-based reconstruction method.

Results and Discussion

MRF-DiPh demonstrated superior performance across all quantitative metrics compared to the baseline models.

Quantitative Accuracy

The framework achieved significant improvements in tissue parameter mapping. Compared to the best-performing deep learning baselines, MRF-DiPh improved T1 mapping accuracy by approximately 2% and T2 mapping accuracy by 4.2%, as measured by Mean Average Percentage Error (MAPE). Furthermore, the Normalized Root Mean Square Error (NRMSE) for the time-series images was reduced by 8.6%.

Physical and Data Consistency

A key advantage of the physics-informed approach is its adherence to the acquired data. MRF-DiPh showed approximately 13% fewer errors in k-space consistency compared to purely data-driven diffusion models. This suggests that the inclusion of physical constraints acts as a "guardrail," preventing the diffusion model from generating plausible-looking but incorrect tissue values.

Training and Sampling Insights

The researchers observed that the physics-guided model reached high accuracy significantly earlier in the training process than its purely data-driven counterpart (MRF-IDDPM). This indicates that physical laws provide a structured "map" that facilitates faster convergence. Additionally, the study found that "conditioning" the diffusion model on the initial artifact-heavy images performed better than unconditional models, particularly in complex regions such as grey matter and cerebrospinal fluid (CSF).

Conclusion and Future Directions

MRF-DiPh successfully combines the generative strengths of diffusion models with the reliability of physical models. While the reconstruction time of approximately 44 seconds per slice is feasible for clinical research, the dictionary-matching step remains a computational bottleneck, accounting for 40-53% of the total reconstruction time.

Future work will focus on optimizing these matching algorithms and extending the framework from 2D slices to 3D volumes. Additionally, exploring different subsampling strategies beyond sequence truncation will be essential to fully realize the potential of physics-informed diffusion in diverse clinical MRI protocols.

Summary updated

BibTeX
@inproceedings{mayo2025physics,
  author = {Perla Mayo and Carolin M. Pirkl and Alin Achim and bjoern menze and Mohammad Golbabaee},
  title = {{Physics Informed Guided Diffusion for Accelerated Multi-parametric MRI Reconstruction}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2025},
  doi = {10.1007/978-3-032-05325-1\_41},
}

Loading related content...

Figure