3D Diffusion tensor magnetic resonance images denoising based on sparse representation

Abstract

Diffusion tensor magnetic resonance imaging (DT-MRI) is widely used to characterize white matter health and brain disease. However, the DT-MRI is very sensitive to noise. This paper proposes a sparse representation based denoising method for 3D diffusion weighted images (DWI) in DT-MRI. As consecutive 2D images in DWI volume have similar content and structure, we can process a fixed number of adjacent images from DWI volume simultaneously. The proposed method first learned a dictionary from the selected 2d diffusion weighted images according to the K-SVD learning algorithm. Then the clean images are obtained by gradually approximating the underlying images using the bases selected from the learned dictionary based on sparse representation. At last, the tensor images are estimated from the diffusion weighted images. The experiments on both synthetic and real DT-MRI images show that the proposed method performs better than classical techniques by preserving image contrast and structures.