Denoising-Methods For Preserving Edge Of PET Images

Malignant tumor has become one of the leading causes of death. Computer Tomography (CT), Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET) images provide tools for tumor diagnosis, staging, radiation therapy plan and curative effect evaluation. PET images can reflect the human's metabolism, physiological and biochemical process. So PET is a key imaging tool for tumor. However there are a lot of noises in PET images which needs to be removed. Based on PET image and its noise characteristics, we proposed a multi-directional geometric nonlinear diffusion filtering (MD-GNLDF) and a non-local geometric nonlinear diffusion filtering (NL-GNLDF) for denoising of PET images and preserving edge in the PET images as follows:(1)Geometric nonlinear diffusion filtering (GNLDF) only considers the edge of the images in the horizontal direction and vertical direction. In fact, there are a lot of edges with different directions in images. So, we proposed a multi-directional geometric nonlinear diffusion filter (MD-GNLDF). Based on the directional gradient magnitudes of the 8-neighborhood pixels, the pixel first is judged as edge point or noise point. If it is a noise pixel, it is diffused via GNLDF in the four directions, such as horizontal, vertical, main and deputy diagonal directions. If it is an edge pixel, it is diffused via GNLDF only along the edge direction in the points. We found the MD-GNLDF can improve peak signal-to-noise ratio (PSNR) 1.56 dB, reduce root mean square error (RMSE) 0.001, and increase image enhancement factor (IEF) 2.37 more than GNLDF.(2)Even if GNLDF can get good result for denoising of CT imges, it may result block effect for denoising of PET images. So we propose a novel non-local geometric nonlinear diffusion filter (NL-GNLDF) for denoising of PET images. First, we calculate the coefficients of a geometric nonlinear diffusion filter (GNLDF). Then, we replace the coefficients with its non-local means, a non-local averaging of all coefficients. Finally, we use the resulting NL-GNLDF to denoising of PET images. We found that the NL-GNLDF can improve the PSNR and structural similarity of PET images, and visually enhance PET images by comparing it with GNLDF, Non-local Mean Filter and PURE-LET. The NL-GNLDF is an efficient filter for PET image noise reduction. In fact, we found the NL-GNLDF can improve PSNR about 1.2 dB, reduce RMSE 0.00064, improve SSIM 0.0051 and increase IEF 2.7 more than PURE-LET.