Fast zero-shot deep learning-based denoising method for low-field MR images.

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging (MRI)
• Image Processing

Background:

• Low-field MRI is crucial for accessible diagnostics but often suffers from image noise, impacting clinical utility.
• Traditional denoising methods require extensive training data, which is challenging to acquire for low-field MRI.
• Zero-shot self-supervised learning offers a promising alternative by eliminating the need for prior training data.

Purpose of the Study:

• To adapt and optimize a zero-shot denoising approach for low-field MRI.
• To accelerate the training process of scan-specific denoising methods.
• To improve the image quality of low-field MRI for enhanced clinical diagnosis.

Main Methods:

• Extended the zero-shot noise-as-clean (ZS-NAC) method with modifications for faster training.
• Compared the proposed method against BM4D and zero-shot noise2noise techniques.
• Evaluated denoising performance quantitatively on high-field data and qualitatively on prospective low-field (0.1 T) data.
• Investigated training on partial data matrices for further acceleration.

Main Results:

• The optimized ZS-NAC method achieved high denoising performance across various signal-to-noise ratio (SNR) levels.
• Rapid processing times (seconds) were observed on GPU for typical low-field MRI data dimensions.
• Training on data subsets demonstrated potential for significant training acceleration.

Conclusions:

• The developed denoising method shows substantial potential for seamless integration into low-field MRI acquisition workflows.
• This approach can effectively improve image quality, aiding clinical diagnosis.
• The method's efficiency and adaptability make it suitable for real-world clinical applications.