A parallel K-SVD implementation for CT image denoising
Dominik Bartuschat, Anja Borsdorf, Harald Köstler, Ron Rubinstein, Markus Stürmer, Lehrstuhlbericht · 2009
In this work we present a new patch-based approach for the reduction of quantum noise in CT images. It utilizes two data sets gathered with information from the odd and even projections respectively that exhibit uncorrelated noise for estimating the local noise variance and performs edge-preserving noise reduction by means of the K-SVD algorithm. It is an efficient way for designing overcomplete dictionaries and finding sparse representations of signals from these dictionaries. For image denoising, the K-SVD algorithm is used for training an overcomplete dictionary that describes the image content effectively. K-SVD has been adapted to the non-gaussian noise in CT images. In order to achieve close to real-time performance we parallelized parts of the K-SVD algorithm and implemented them on the Cell Broadband Engine Architecture (CBEA), a heterogenous, multicore, distributed memory processor. We show denoising results on synthetic and real medical data sets.