Summary:
UCLA researchers in the Department of Radiological Sciences have developed a novel physics-guided, self-supervised deep learning framework for multi-dimensional magnetic resonance imaging that delivers superior image quality and preserves spatiotemporal fidelity at high acceleration rates.
Background:
Magnetic resonance imaging (MRI) is a versatile, non-invasive medical imaging tool that uses a magnetic field and computer-generated radio waves to create detailed images of organs and tissues in the human body, playing a critical role in diagnostic medicine. Multi-dimensional MRI builds on these capabilities by enabling the acquisition and analysis of data across multiple dimensions, offering a richer characterization of biological structure and function in real time. Despite its high clinical value and growing research interest, broader adoption of multi-dimensional MRI remains limited by long acquisition times which can disrupt clinical workflow, reduce patient tolerance, and constrain wider use. While existing acceleration strategies reduce scan time, higher acceleration rates come at the expense of image quality and spatiotemporal fidelity, hindering the ability to obtain reliable, information-rich images needed for advanced clinical and research applications. There is therefore a clear unmet need for improved MRI methods that enable faster multi-dimensional imaging, while maintaining high image quality and reliable spatiotemporal resolution.
Innovation:
To address these limitations, UCLA researchers have developed a physics-guided, self-supervised deep learning framework for reconstructing highly accelerated multi-dimensional MRI data with high image quality and spatiotemporal fidelity. By combining data-driven reconstruction with physics-based constraints, the technology is designed to preserve spatial detail and temporal consistency even at high acceleration rates, where existing methods struggle. This innovation enables faster imaging without sacrificing the quality of information needed for advanced clinical and research applications. The self-supervised nature of the framework makes it especially well-suited for multi-dimensional MRI, where fully sampled ground-truth data are often difficult to obtain. Ultimately, this novel innovation is adaptable across a wide range of multi-dimensional MRI applications and enables faster imaging while maintaining high image quality and spatiotemporal fidelity, making it a valuable tool for improving diagnostic imaging and easing clinical workflow.
Potential Applications:
- Cardiovascular MRI applications – Enables highly accelerated multi-dimensional cardiovascular imaging, including 4D MUSIC and 4D flow MRI, with preserved image quality and spatiotemporal fidelity.
- Quantitative and contrast-enhanced MRI – Supports advanced MRI applications such as proton density fat fraction (PDFF) mapping, T1 mapping, T2 mapping, T2*/R2* imaging, and
contrast-enhanced multi-dimensional imaging.
- Software integration for existing clinical MRI scanners – Can be deployed as a reconstruction add-on to enhance image quality and support faster multi-dimensional MRI workflows.
- Broad multi-dimensional MRI applications – Can be adapted for a wide range of advanced
multi-dimensional MRI applications while enabling faster acquisition without compromising image quality or spatiotemporal fidelity.
Advantages:
- Enables highly accelerated multi-dimensional MRI reconstruction
- Maintains high image quality even at high acceleration rates
- Preserves spatiotemporal fidelity in dynamic MRI dataset
- Uses self-supervised learning from under sampled MRI data
- Eliminates the need for fully sampled ground-truth training datasets
- Can be adapted across a broad range of multi-dimensional MRI applications
- Supports faster MRI acquisition without sacrificing important diagnostic information
State of Development:
The inventors have developed a physics-guided, self-supervised deep learning reconstruction framework for highly accelerated multi-dimensional MRI and demonstrated its performance in representative cardiovascular MRI applications, including 4D MUSIC and 4D flow MRI. The technology has been trained and evaluated on prospectively acquired under sampled datasets and has shown improved image quality, reduced artifacts, and preserved spatiotemporal fidelity at high acceleration rates.
Related Papers:
- Finn JP, Shih SF, Nguyen KL, Bedayat A, Yoshida T, Jin N, Han F, Zhong X. Four-dimensional, ferumoxytol-enhanced MUSIC (multi-phase steady-state imaging with contrast) in a single breath-hold: Technical feasibility in structural heart disease. Proc Intl Soc Mag Reson Med 2025;33:402. Honolulu, Hawaii.
Reference:
UCLA Case No. 2026-164
Lead Inventors:
Xiaodong Zhong, Department of Radiological Sciences; John Finn, Department of Radiological Sciences