Detection of locally stationary region for universal GMM and its application in denoising X-ray CT images

An adaptive Wiener filter (AWF) for denoising X-ray CT image has been proposed based on the universal Gaussian mixture distribution model (UNI-GMM). The universal model can be estimated by an assumption that the GMM is stationary. In the previous UNI-GMM-AWF method, a fixed observation block size of UNI-GMM has been adopted, assuming smaller block size makes the block more stationary, but the small block tend to suffer observation error due to image noise. Thus in the previous method, the observation region size was not small enough to satisfy the stationary assumption. Inversely the observation region size is not large enough for precise model detection and high denoising ability in stationary region. To overcome the problems, variable observation block sizes of the UNI-GMMs are adopted in this paper. Actually, in the new UNI-GMM-AWF method, two sizes of the UNI-GMMs are applied for each observation region and the most stationary UNI-GMM for each observation region is selected according to the normalized likelihood function, related to the Akaike's information criteria (AIC)[1]. Moreover, the new UNI-GMM which has a observation region with hole in its central region is applied to detect a small point shape structure like a small vessel or a bronchiole. Then the new UNI-GMM using observation region with hole is also selected for each observation block based on the AIC. Simulation results show that the proposed method performs better than median filter as a standard method in terms of the denoising and point like shadow preservation ability. Furthermore a simulation result shows that the new UNI-GMM-AWF is more flexible than the previous UNI-GMM-AWF method in terms of the applicability of fitting the stationary model.