Geometry-driven-diffusion filtering of magnetic resonance images using model-based conductance

The paper deals with the problems of staircase artifacts and low-contrast boundary smoothing in filtering (magnetic resonance MR) brain tomograms that is based on geometry-driven diffusion (GDD). A novel method of the model-based GDD filtering of MR brain tomograms is proposed to tackle these problems. It is based on a local adaptation of the conductance that is defined for each diffusion iteration within the variable limits. The local adaptation uses a neighborhood inhomogeneity measure, pixel dissimilarity, while gradient histograms of MR brain template regions are used as the variable limits for the conductance. A methodology is developed for implementing the template image selected from an MR brain atlas to the model-based GDD filtering. The proposed method is tested on an MR brain phantom. The methodology developed is exemplified on the real MR brain tomogram with the corresponding template selected from the Brainweb. The performance of the developed algorithms is evaluated quantitatively and visually.