Summary:
UCLA researchers have developed a software algorithm for locally low-rank denoising that improves image quality and diagnostic reliability in multi-coil, multi-contrast MRI by correcting complex noise and preserving fine image details.
Background:
Magnetic resonance imaging (MRI) is a powerful, non-invasive tool widely used in both research and clinical diagnostics. To reduce scan time, modern MRI systems employ multi-coil acquisition schemes, where multiple receiver coils capture different parts of the signal simultaneously. These coil signals are then combined to reconstruct high-quality images from undersampled data, using advanced algorithms such as GRAPPA (Generalized Autocalibrating Partially Parallel Acquisition). While this multi-coil approach greatly accelerates imaging acquisition, it also changes how noise behaves in the data. In applications where image noise is already a limiting factor, such as low-field MRI or real-time cardiac imaging, the altered noise distribution can lead to degraded image quality and unreliable reconstructions.
Several denoising algorithms, including MP-PCA and NORDIC, have been developed to separate true signal from random noise using statistical modeling. However, these tools were originally designed for diffusion MRI, a specialized sequence with hundreds of image contrasts. When applied to other MRI applications with fewer image contrasts or varied coil setups, these methods often fail to perform optimally, resulting in over-smoothed images, residual noise, or poor generalization. Due to these limitations in the state of the art, there remains a critical need for a robust, generalizable denoising solution that is compatible with multi-coil, multi-contrast MRI across a wide range of clinical and research applications.
Innovation:
UCLA researchers have developed a locally low-rank image denoising software that significantly improves image quality for multi-coil, multi-contrast MRI datasets. The software is structured in two main stages:
1. Noise correction and normalization:
The preprocessing module directly handles raw imaging data to ensure that noise characteristics are consistent across all coils and contrasts used in GRAPPA reconstruction. This step effectively “normalizes” the noise to behave like Gaussian noise, the most predictable form, allowing more accurate and consistent denoising.
2. Adaptive denoising algorithms:
The core denoising module separates signal from noise during image reconstruction. Two locally low-rank techniques are implemented. Both methods leverage the intrinsic signal structure within the data but differ in mathematical formulation, providing flexibility for different imaging conditions. The algorithms jointly process the multi-coil, multi-contrast data and generate denoised, coil-combined images as output.
Compared to existing open-source tools such as MP-PCA and NORDIC, this UCLA software uniquely integrates multi-coil noise correction, signal decorrelation, and g-factor map generation, enabling robust performance in datasets with limited contrasts. It also extends denoising capability beyond diffusion MRI to broader use cases such as cardiac, abdominal, interventional, and low-field MRI, where noise has inherently limited image quality. By improving the signal-to-noise ratio (SNR) without blurring or loss of detail, this technology enhances diagnostic reliability, supports advanced quantitative imaging, and can be seamlessly integrated into commercial MRI reconstruction pipelines or post-processing platforms.
Potential Applications:
• Cardiac MRI – Enhances visualization of heart structure and motion in accelerated scans
• Abdominal and liver imaging – Improves image quality and SNR in motion-prone or deep-organ regions
• Interventional MRI – Provides clearer, real-time visualization during minimally invasive procedures
• Low-field MRI systems – Compensates for reduced SNR in portable or cost-efficient scanners
• Neuro and musculoskeletal MRI – Improves structural detail and quantitative measurement reliability
• Functional and emerging MRI applications – Benefits low-SNR modalities such as lung imaging and hyperpolarized MRI
Advantages:
• Noise reduction without added scan time – improves SNR using intrinsic image structure instead of repeated signal averaging
• Broad compatibility – applicable across diverse multi-coil, multi-contrast MRI sequences beyond diffusion imaging
• Integrated noise calibration – accurately models noise behavior with built-in signal decorrelation and g-factor correction for GRAPPA-based parallel imaging accelerated reconstruction
• Sharper, more reliable images – enhances diagnostic confidence by reducing noise without blurring fine details
• Seamless integration – easily incorporated into commercial MRI reconstruction software or research workflows
• Optimized for challenging conditions – maintains high performance in low-field, accelerated, or time-constrained imaging environments
State of Development:
The software has been validated on healthy volunteer MRI datasets. The implementation is available under an Academic Software License on GitHub: https://github.com/HoldenWuLab/LLR-image-denoising.
Related Publications:
Shih, Shu‐Fu, et al. "Improved liver fat and R2* quantification at 0.55 T using locally low‐rank denoising." Magnetic Resonance in Medicine 93.3 (2025): 1348-1364.
Reference:
UCLA Case No. 2025-056
Lead Inventors:
Shu-Fu Shih and Holden H. Wu from the Department of Radiological Sciences.