Foveated Nonlocal Means Despeckle Filtering for Ultrasound Imaging: Imaging Perspective

Abstract

Ultrasound speckle noise degrades imaging contrast and hides anatomical details; thus causing inaccuracy in clinical diagnosis. Although speckle reduction methods such as classical nonlocal means (NLM), optimized Bayesian nonlocal means (OBNLM), and speckle reducing anisotropic diffusion (SRAD) filters have been proposed for years, they still suffer two major problems – insufficient preservation of characteristic details such as calcifications and inordinate blurring making image appearance artificial. To solve the two problems, we propose a novel foveated nonlocal means despeckle filtering technique, inspired by the human visual system. Conventional NLM filters despeckle via searching for analogous patches at different areas within the image and then estimating the impulse response by the degrees of similarity appraised by a windowed Euler distance between the target and searching patches. In our technique, foveated self-similarity is used instead of conventional self-similarity. The foveated self-similarity is based on a new patch operator mimicking human retina properties, sharpening patch pixels in the center and blurring them near the periphery. Moreover, throughout the literature, the tuning of the search window and patch sizes and other parameters are not consistent; nonetheless, in this study, they are tuned universally from imaging perspective, i.e., according to the size of point spread function which allows the adaption to different imaging systems and settings. Simulations and clinical data (not shown here) were used to verify our proposed method. The performance of our proposed method is also compared with the classical despeckle filters. The results demonstrate that the proposed technique can remove speckles forcefully while more effectively retaining structural edge details, textures, and point-like structures. Quantitative measures such as contrast-to-noise ratio, edge preservation index and contrast measure are also presented.