PET Image Denoising Based on Non-local Low Rank Matrix Approximation

Abstract

Positron emission tomography (PET) is an effective molecular imaging method that allows non-invasive quantitative observation of physiological and biochemical changes in living organisms. PET image is reconstructed via sinogram, so the quality of the reconstructed PET image is often limited by various physical degradation factors. PET image denoising is vital because it can enhance both visual quality and quantitative metrics for better diagnostic decisions. However, because of the high contrast and low spatial resolution of the PET image, it is difficult to suppress the noise and preserve the fine details at the same time. To address this issue, this paper develops a novel PET image denoising scheme based on low rank matrix approximation (LRMA) that exploits both advantage of self-similarity and implied low rank of the images. Moreover, an enhanced back projection step is proposed for compensating the loss details to improve the denoising result. Experimental results have shown that the proposed approach outperforms related methods includng iNLM, ELME and BM3D in terms of both perceptual quality and quantitative metrics.