Locally adaptive conductance in geometry-driven-diffusion filtering of magnetic resonance tomograms

A novel methodology for locally adapting the exponential conductance in geometry-driven diffusion (GDD) is proposed which employs pixel dissimilarity measures. Two alternative approaches are developed; both are based on a transient interval, within which the relaxation parameter K is selected. In the first case, the limits of the interval are derived from global quantiles of the intensity gradients; in the second case, they are derived from the optimal variable parameter K_0pt, calculated from a specific cost function. This function is designed using intensity gradient histograms of region interiors and boundaries in an appropriate image template of an MR brain tomogram. As a local measure, the mean direction dissimilarity has been used. Computer experiments with the locally adaptive geometry-driven diffusion filtering of an MR-head phantom have been performed and quantitatively evaluated. They include, as a reference, two other GDD filtering methods