Ultrasound Dereverberation/Deconvolution Filtering Based on Gaussian Mixture Modeling

Reverberation is a significant source of degradation in ultrasound imaging in regions with mixtures of scattering structures. Its effects vary from subtle to pronounced, but they can degrade both spatial and contrast resolutions. Deconvolution filters based on the system impulse response often improve axial resolution in uniform speckle regions, but may not perform optimally in complex scattering regions. We have developed an algorithm for the design of a dereverberation/deconvolution filter (DDF) based on a Gaussian mixture model (GMM) of echo data from heterogeneous tissues. RF data were collected using the LA 14-5 probe on a SonixRP scanner while imaging the femoral artery of a familial hypercholesterolemic swine in vivo under approved protocol. The tissues surrounding the target vessel included muscle, fat and connective tissue. Correlation cell sizes and echo statistics differed substantially, which justified the use of the GMM of order 5 for this FOV. The DDF filter was derived from A-lines passing through the vessel to capture short- and long-range spatial correlations, a key feature for estimating the GMM parameters. An expectation-maximization algorithm was used to derive the DDF coefficients while updating the GMM. The algorithm converges within a few iterations to a causally stable IIR filter with well-behaved impulse response. Further iterations allow the DDF to equalize the frequency response and achieve deconvolution without the need for regularization.