Despeckling of Medical Ultrasound Images using Sparse Representation

Recently there has been a growing interest in the sparse representation of signals. Particularly, many new multiscale transforms have been proposed in this direction. Instead of using fixed transforms such as wavelets, curvelets etc., an alternative way is to train a dictionary from the image itself. This paper presents a novel despeckling scheme for medical ultrasound images using such a sparse and redundant representation. It is shown that the proposed algorithm can be used effectively for removal of multiplicative speckle noise by introducing a simple preprocessing stage before an adaptive dictionary is learned from the image patches (called atoms) for sparse representation. This learning process, called the K-SVD, is efficiently performed using an Orthogonal Matching Pursuit (OMP) and a Singular Value Decomposition (SVD). Results are evaluated both on US images and artificially speckled photographic images.