跳转至

利用幅度与复数测量结合学习先验提升动态磁共振成像重建性能

文章背景与核心概要

磁共振成像（MRI）重建通常依赖于欠采样的复数值 \(k\) 空间测量。虽然稀疏相位恢复表明纯幅度测量可以提供互补的信号恢复信息，但由于传统上需要额外的扫描时间，它们在 MRI 中的应用在很大程度上被忽视了。

本文引入了 \(\mathbb{C}+\text{Mag}\)，这是一个专为加速稳态动态 MRI 量身定制的、融合幅度信息的物理驱动深度学习重建框架。通过利用各帧之间 \(k\) 空间幅度的强大时间一致性，该方法集成了一个基于 ADMM（交替方向乘子法）展开的框架，并配备了新颖的幅度感知数据保真度公式。为了克服幅度约束固有的非可微性和非凸性，该方法采用了二次平滑优化和基于动量的更新。

在回顾性欠采样电影（cine）和相衬血流 MRI，以及前瞻性欠采样实时电影 MRI 采集上的实验验证表明，与传统的物理驱动深度学习方法相比，该方法表现出更优的伪影抑制能力、增强的解剖清晰度，并保留了相位信息。盲法专家读片评估进一步验证了这些发现。

Harnessing Magnitude-Only and Complex Measurements for Improved Dynamic MRI Reconstruction with Learned Priors

arXiv:2608.18036 [eess.IV]
Submitted: August 18, 2026
Authors: Mahdi Saberi, Yaşar Utku Alçalar, Merve Gülle, Chetan Shenoy, Mehmet Akçakaya
Primary Subject: Image and Video Processing (eess.IV)
Links: View PDF | HTML Version | DOI

Summary

Magnetic Resonance Imaging (MRI) reconstruction typically relies on undersampled complex-valued \(k\)-space measurements. While sparse phase retrieval suggests that magnitude-only measurements can offer complementary signal recovery information, they have been largely underutilized in MRI due to the traditional need for additional scan time.

This paper introduces \(\mathbb{C}+\text{Mag}\), a magnitude-informed, physics-driven deep learning reconstruction framework tailored for accelerated steady-state dynamic MRI. By leveraging the strong temporal consistency of \(k\)-space magnitudes across frames, the method integrates an ADMM-based unrolling framework equipped with a novel magnitude-aware data-fidelity formulation. To overcome the non-differentiability and non-convexity inherent to magnitude constraints, the approach utilizes quadratically smoothed optimization and momentum-based updates.

Experimental validations on retrospectively undersampled cine and phase-contrast flow MRI, alongside prospectively undersampled real-time cine MRI acquisitions, demonstrate superior artifact suppression, enhanced anatomical sharpness, and preserved phase information relative to conventional physics-driven deep learning methods. Blinded expert reader evaluations further validate these findings.

Metadata & Reference Information