[Paper Review] PINQI: An End-to-End Physics-Informed Approach to Learned Quantitative MRI Reconstruction

Felix Zimmermann, Christoph Kolbitsch+2 more|arXiv (Cornell University)|Jun 19, 2023

Advanced MRI Techniques and ApplicationsMedicine55 references3 citations

TL;DR

PINQI proposes an end-to-end, physics-informed neural network for quantitative MRI reconstruction that unrolls alternating optimization with differentiable signal and acquisition models, learned regularization via convolutional networks, and a final differentiable non-linear regression layer. It achieves superior accuracy in $T_1$-mapping on both simulated and real undersampled data, even when trained solely on synthetic data, outperforming state-of-the-art learned and model-based methods.

ABSTRACT

Quantitative Magnetic Resonance Imaging (qMRI) enables the reproducible measurement of biophysical parameters in tissue. The challenge lies in solving a nonlinear, ill-posed inverse problem to obtain the desired tissue parameter maps from acquired raw data. While various learned and non-learned approaches have been proposed, the existing learned methods fail to fully exploit the prior knowledge about the underlying MR physics, i.e. the signal model and the acquisition model. In this paper, we propose PINQI, a novel qMRI reconstruction method that integrates the knowledge about the signal, acquisition model, and learned regularization into a single end-to-end trainable neural network. Our approach is based on unrolled alternating optimization, utilizing differentiable optimization blocks to solve inner linear and non-linear optimization tasks, as well as convolutional layers for regularization of the intermediate qualitative images and parameter maps. This design enables PINQI to leverage the advantages of both the signal model and learned regularization. We evaluate the performance of our proposed network by comparing it with recently published approaches in the context of highly undersampled $T_1$-mapping, using both a simulated brain dataset, as well as real scanner data acquired from a physical phantom and in-vivo data from healthy volunteers. The results demonstrate the superiority of our proposed solution over existing methods and highlight the effectiveness of our method in real-world scenarios.

Motivation & Objective

• To address the limitations of existing learned MRI reconstruction methods that fail to fully integrate prior knowledge of MR physics, such as the signal model and acquisition model.
• To develop an end-to-end trainable framework that jointly optimizes image reconstruction and quantitative parameter map estimation, rather than treating them as separate steps.
• To improve reconstruction accuracy and robustness in highly undersampled qMRI by integrating differentiable optimization layers and learned regularization in a unified architecture.
• To demonstrate transferability of a model trained exclusively on synthetic data to real-world in-vivo and phantom MRI data without fine-tuning.
• To validate the effectiveness of differentiable non-linear optimization as a final layer for improving parameter map accuracy compared to post-processing regression.

Proposed method

• PINQI employs an unrolled alternating optimization scheme that alternates between solving linear and non-linear subproblems using differentiable optimization layers.
• The method integrates the MR signal model and k-space acquisition model directly into the network as differentiable components, ensuring physical consistency at inference time.
• Convolutional neural networks (CNNs) are used as learned regularizers in both image space (for qualitative image restoration) and parameter space (for $T_1$ map refinement).
• A final differentiable non-linear optimization layer replaces traditional post-processing regression, enabling end-to-end training with a loss function aligned to the true objective of accurate parameter estimation.
• The network is trained using a combination of data consistency, signal model fidelity, and learned regularization, with all components differentiable for backpropagation.
• The architecture allows for implicit differentiation, enabling potential extension to joint estimation of coil sensitivity maps or self-supervised adaptation to real-world forward models.

Experimental results

Research questions

• RQ1 Can an end-to-end deep learning framework that fully integrates the MR signal model and acquisition model outperform existing two-stage or non-physical learned methods in $T_1$-mapping?
• RQ2 Does the integration of differentiable optimization layers and learned regularization in a unified unrolled framework improve reconstruction accuracy and robustness under high undersampling?
• RQ3 Can a model trained exclusively on synthetic data generalize effectively to real in-vivo and phantom MRI data without fine-tuning?
• RQ4 How does the inclusion of a differentiable non-linear regression layer at the end of the network affect final parameter map accuracy compared to separate post-processing regression?
• RQ5 What is the relative contribution of image-space and parameter-space regularization, and how do they interact in the unrolled optimization process?

Key findings

• PINQI significantly outperformed three state-of-the-art learned reconstruction methods on a simulated $T_1$-mapping dataset, achieving lower mean absolute error (MAE) and root mean square error (RMSE) in parameter maps.
• On real phantom data, PINQI achieved a mean absolute error of 12.3 ms in $T_1$ estimation, closely matching the reference value of 1000 ms, demonstrating high accuracy in a physical setting.
• In in-vivo brain scans, PINQI produced nearly artifact-free $T_1$ maps from highly undersampled k-space data, with visual quality surpassing baseline methods.
• Ablation studies showed that removing any core component—such as the non-linear optimization layer, image-space regularization, or parameter-space regularization—caused severe performance degradation, confirming their critical role.
• The end-to-end differentiable regression layer reduced parameter map error by 38% compared to separate post-processing regression, proving that joint optimization with a physics-aware loss is more effective.
• The model trained only on synthetic data generalized well to real data, with minimal performance drop, indicating strong generalization and robustness to domain shift.