Feature-Preserving Smoothing of Diffusion Weighted Images Using Nonstationarity Adaptive Filtering
Yan-Li Zhang, Wan-Yu Liu, Isabelle E. Magnin, and Yue-Min Zhu
Volume: 60, Issue: 6, Page: 1693-1701
Investigating the microstructure of tissues is essential for understanding the cause of their diseases, which is in turn beneficial to the diagnosis and the treatment in clinics. For instance, many cardiovascular diseases, such as myocardial infarction known as heart attack, actually result from the death of heart muscle. Owing to its ability of measuring the diffusion of water molecules, diffusion tensor magnetic resonance imaging (DT-MRI or DTI) offers a promising technique to get insights into the fiber architecture of in vivo tissues.
However, the performance and the potentiality of DTI are hampered by the presence of high-level noises in diffusion weighted (DW) images. In certain situations, these images are corrupted so severely that their features, such as edges or details, can be buried in the speckled mosaic-like patterns aroused by noise, and subsequent DTI analyses are vitiated by uncertainty and errors. Reducing high-level noise in DW images while preserving desirable features still remains a persistent challenge. In this paper, a new feature-preserving smoothing method, called the nonstationarity adaptive filtering (NAF), is proposed. The basic idea of the method is to replace a pixel’s intensity by the averaged intensity of its homogeneous neighborhood, but with the particularity of appropriately determining the homogeneous neighborhood even in highly noisy cases. Experimental results on both synthetic and real human DW images showed that the NAF makes DW images gain better visual appearance, regularizes diffusion tensor fields effectively while restoring the anisotropy of tensors, allows for more accurate measurements of diffusion properties, and ensures more reliable fiber reconstruction.