Abstract

Automated image processing of medical ultrasound images offers great challenges for computer vision due to the impact of physical noise phenomena, e.g., the characteristical speckle noise. In this talk three different concepts are proposed to tackle these problems with the help of variational methods in the context of automated image segmentation. On the one hand, segmentation is formulated as a statistically motivated inverse problem based on Bayesian modeling. In contrast to this exact modeling of noise distributions, an alternative approach based on level set methods is elaborated subsequently, which performs segmentation based on the results of a discriminant analysis of medical ultrasound images. Motivated by the presence of structural artifacts in the data, e.g., shadowing effects, the latter two segmentation methods are extended by a shape prior based on Legendre moments in order to give additional support in terms of trained knowledge about expected shapes. The proposed methods are compared qualitatively and quantitatively on real patient data from echocardiography.