Self-supervised denoising for high-dimensional magnetic resonance image

Xuanyu Tian, Jiangjie Wu, Guoyan Lao, Chenhe Du, Changhao Jiang, Yanbin Li, Jeff L. Zhang, Hongjiang Wei, Yuyao Zhang

Abstract: The acquisition in magnetic resonance imaging (MRI) presents the trade-offs between the signal-to-noise ratio (SNR), spatial resolution, and scanning time. Recently, self-supervised denoising methods without high SNR images for training are emerging as competitive alternatives in MRI denoising. However, self-supervised denoising methods for high-dimensional MRI data require further exploration, as the direct application of current methods is not efficient enough. In this work, we propose Noise2SR-M (N2SR-M), a self-supervised denoising method for high-dimensional MRI data. We discuss the signal correlation of voxels in high-dimensional MRI. Utilizing the correlations of signals, N2SR-M is trained on paired noisy data with different spatial sizes generated from individual noisy MRI data. The paired noisy data comprises sub-sampled noisy data in the spatial domain and the original noisy data. Meanwhile, N2SR-M performs joint denoising across various contrast dimensions. By leveraging contrast dimension constraints and the effectiveness of a super-resolution-based training strategy, N2SR-M achieves superior performance while preserving fine tissue structures in denoising MRI data. We extend comprehensive experiments involving simulated and real multi-echo gradient echo (mGRE), GRE phase, and diffusion weighted imaging data to evaluate the effectiveness of N2SR-M. The results show that N2SR-M successfully restores detailed image content and effectively avoids the generation of artifacts or overblurring. Furthermore, the N2SR-M denoised data induce a more accurate parameter mapping (i.e., R2∗, quantitative susceptibility mapping, and diffusion tensor imaging).