Comparison of a Self-Supervised Denoising Technique for Diffusion-weighted Images in Brain MR with Other Methods

Purpose:
One of the challenges in DW imaging is the presence of noise, which reduces SNR. The presence of noise can definitely cause incorrect quantitative calculations, and especially in DT imaging where we need accurate calculations of the connections of brain white fiber structures, low SNR will be very troublesome. In this study, several noise reduction methods have been implemented on Diffusion-Weighted MR images of brain and compared with each other.
Materials and Methods:
DW images of 10 patients were initially used in a self-supervised method to denoise. This method, which uses an entire volume to learn a full-rank local linear eliminator for that volume, can separate structure from noise without requiring an explicit model for each, using q-space oversampled DWI data. Then three conventional methods of Non- Local Means and PCA with experimental thresholding and PCA by Marcenko-Pastur method are performed on all the images. In order to compare the success rate of denoise, the SNR value in the corpus callosum region was measured on all images.
Results:
According to the calculation of SNR in each direction, the average of all directions for each method was calculated in all images. The average SNR for self-supervised, non-local means, PCA with experimental thresholding, and PCA by Marcenko-Pastur method were 24.8, 23.4, 23.7, and 23.2, respectively.
Conclusion:
It seems that the self-supervised method is more successful than other methods in denoise DW images. Unlike other methods that assume certain properties in the signal structure, this method does not make such an assumption about the signal, instead, it uses the fact that the noise in the 3D volumes of the DWI signal originates from random fluctuations in the acquired signal, which is closer to the reality of the noise event.