Non-additive noise reduction in medical images using bilateral filtering and modular neural networks

Abstract

For the longest time, filters have been proposed to reduce speckle noise present in the medical images for improved visual quality and quantitative image analysis. Of the numerous image restoration and denoising techniques around, one of the most basic and productive is a neural network system-based noise suppression technique. Though noise suppression techniques perform well, they fail to preserve image edges. Consequently, denoising of medical images is a major area of research. This paper proposes a median filter-based multistage neural network for speckle noise suppression. The proposed method has two stages, median filter is used in the initial filtering stage. Statistical features extracted from its output are utilized to train the multistage neural network. The filtered images are again denoised using the multistage neural network, following which there is a performance evaluation using several assessment metrics. The proposed method has been evaluated on two different datasets namely breast ultrasound dataset and multispectral image dataset images. The simulation results show that the proposed method outperforms existing techniques in terms of the PSNR of 53.80%, SSIM of 98.54, EPI of 0.74, ENL of 96.62% and SSI of 0.22, respectively, for the breast ultrasound images and also, produces the PSNR of 67.38%, SSIM of 98%, EPI of 0.72, SSIM of 98.93% and the EPI of 0.32, respectively, for the multispectral images.