Dictionary learning-based denoising algorithm with expected patch log likelihood in diffusion-weighted magnetic resonance image

Abstract

Diffusion-weighted imaging (DWI) is one of the most sensitive techniques to noise among magnetic resonance imaging (MRI) techniques. As the b-value used to acquire the DWI image increases, an image in which the difference in diffusion is emphasized can be obtained. However, DWI images with increased b-values inevitably have a major drawback in that noise is amplified. Thus, in this study, a dictionary learning (DL)-based denoising algorithm was modeled and applied to DWI images. The designed algorithm was modeled as a DL-based algorithm using the expected patch log likelihood. The DWI images were obtained by adjusting the b-value from 400 to 400 intervals. When the proposed DL-based denoising algorithm was applied to DWI, we confirmed that the contrast-to-noise ratio and coefficient of variation were improved by approximately 4.26 and 5.22 times, respectively, compared with noisy images. In conclusion, we expect that the proposed DL-based denoising algorithm will be highly efficient in acquiring DWI images using a high b-value, which is useful for observing acute cerebral infarction and microvascular disease.