Non-linear diffusion for interactive multi-scale watershed segmentation

Abstract

Multi-scale watersheds have proven useful for interactive segmentation of 3D medical images. For simpler segmentation tasks, the speed up compared to manual segmentation is more than one order of magnitude. Even where the image evidence does not very strongly support the task, the interactive watershed segmentation provides a speed up of a factor two. In this paper we evaluate a broad class of non-linear diffusion schemes for the purpose of interactively segmenting gray and white matter of the brain in T1-weighted MR images. Through a new scheme GAN, we show that diffusion similar to the nonlinear Perona- Malik scheme is superior to the other evaluated diffusion schemes. This provides a speed up factor of two compared to the linear diffusion scheme.