Mid-level Shape Priors for Supervised Contouring

Abstract. This paper introduces shape priors that benefit 2-dimensional, interactive contouring, which is still a common practice in lesion segmentation due to problems of boundary ambiguity and variable shape. The new Statistical Shape Models (SSMs) capture global shape information without requiring correspondence. We use time series representations of 2-dimensional boundaries and adapt Langevin and Gaussian Process dynamical models to capture global shape properties based on boundary fluctuations. Experiments reveal the benefits of the shape priors in supervised segmentation frameworks.