서강대학교

초록/요약

Effective speckle reduction in ultrasound B-mode imaging is important to enhance the image quality and improve the accuracy in image analysis. Various speckle reduction methods based on a speckle modeling have been extensively proposed for a couple of decades. However, a common limitation of the model-based methods is the requirement of an input-dependent parameter tuning. Thus, most of the methods suffer from over-smoothing, blurring and even generate images with artificial appearance owing to inappropriate selection of the model parameters. Convolutional neural network (CNN) is very promising approach in image processing since a complex and an inaccurate modeling are not required to achieve an objective. However, in ultrasound speckle reduction, in vivo ground-truth data is virtually unavailable as a real ultrasound image always contains a speckle pattern. In this paper, to overcome the limitations of accurate speckle modeling and in vivo data sample, speckle reduction and image enhancement network (SRIE-Net) based on CNN has been proposed using a set of computed tomography (CT) ground-truth and corresponding pseudo B-mode. The performance of SRIE-Net has been compared with oriented speckle reducing anisotropic diffusion (OSRAD), optimized Bayesian non-local mean (OBNLM) and anisotropic diffusion with memory based on speckle statistics (ADMSS) in simulation, phantom and in vivo studies. The results shown that the proposed SRIE-Net enhance the image quality by reducing ultrasound speckle and preserving details, while the classical methods suffer from blurring, over-smoothing and low-performance.