Study On Self-similarity Denoising For Medical Tomographic Image

As one of important medical imaging technology, computed tomography method plays an important role in assisting doctors in the diagnosis and treat-ment. Due to unavoidable noise in the image acquisition process, image is not clear, which influences physician’s interpretation or further analysis of intelligent machine. Researchers propose a lot of noise removal methods, for example, the spatial domain adaptive filter, the transform domain filter, the statistics and stochastic analysis, the partial differential equations, the sparse representations. Most methods study a key issue of natural image denoising: feature-preserving image filtering.For computer tomography imaging medical details and features are of great significance, which are different from natural images. Using the self-similarity and the redundancy of medical image, the self-similarity based non-local means method has a significant impact on the feature-preserving image denoising field. However, this method is not ideal for medial tomographic image with non-stationary noise, though it works well for natural image with stationary noise. Using the self-similarity and the distribution of the non-stationary noise, we propose two improved algorithms based on the classic non-local means method, and obtain better image filtering results in the case of the non-stationary noise.1. The probabilistic patch-based self-similarity denoising. The classic non-local algorithm uses the Gaussian weighted Euclidean distance to measure the similarity of overlapping blocks. However, for the medical to-mographic image with dense textures and the uneven noise distribution, the measurement with the Gaussian weighted Euclidean distance is not very ac-curate. According to the probability of the non-stationary noise, we measure the similarity between image blocks with their probability, which better esti-mates the self-similarity in the case of non-stationary noise, and obtain better filtering results.2. The self-similarity denoising based on probabilistic patch and noise analysis. Considering the weakness of the probability patch-based self-similarity denoising method in the case of high noise level, we present the selection of the smoothing parameter adjusted by local mean with the analysis of image feature and the non-stationary noise. At the same time, we propose a reasonable preselection strategy considering the statistical property of image patch and the distribution of the patch similarity; Finally, employing above two boosting strategies and the two-step prediction-correction strategy, we obtain more robust denoising results.The method based on the probabilistic patch-based and the noise analy-sis deepens and enriches the self-similarity method in the application of the filtering of medical tomographic image, which has important theoretical and clinical values.