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

Abstract

Objective: Denoising low-field MR images is often essential to obtain image quality that is adequate for clinical diagnosis while keeping scan time patient-friendly. The recently introduced zero-shot self-supervised approach shows great promise, requiring no prior data collection for training, which is particularly challenging at low-field. Here, this scan-specific denoising approach is adapted to low-field MR data and optimized to accelerate the training process.

Material and method: We extended the zero-shot noise-as-clean method by modifying the training process to achieve faster training times. The proposed method was compared to BM4D and the recent zero-shot noise2noise methods. Denoising performance was first evaluated quantitatively on high-field data where high SNR images are available, then assessed qualitatively on prospective low-field data (0.1 T). Ultimately, we studied the denoising performance with respect to training on portions of the original data matrix as a potential strategy for further training acceleration.

Results: The proposed method achieved high denoising performance across different SNR levels within a few seconds on a GPU for typical low-field data dimensions. Additionally, training on portion of the data showed potential for further training acceleration.

Discussion: In the context of low-field MRI, this denoising method shows great potential, as it could be integrated into acquisition workflows relatively seamlessly to improve image quality. Code: https://github.com/reinaayde7/zs-nac.git.