Image gradient based shape prior for the segmentation of not that spherical structures

A popular method for the segmentation of somewhat spherical structures (e.g. certain types of tumors, lymph nodes, lung nodules) from 3D medical images is sending out radial rays from a central point and determining the most likely radius for each ray, resulting in a closed surface. Besides satisfying some image based criteria, a regularization term or shape prior typically ensures a smooth contour by preferring similar radii of neighboring rays. In this paper we show that the structures it is often applied to, are in fact not that spherical. We propose an alternate shape prior depending on the gradient direction, preferring smooth structures that are not necessarily spherical. We quantitatively evaluate the proposed shape prior with the traditionally used shape prior on a set of 49 lymph nodes from clinical images. A dice similarity coefficient improvement of 4% has been observed (0.80 vs. 0.77), yielding in segmentation accuracy close to manual segmentation (DSC of 0.83).