Segmentation of the lumbar spine with knowledge-based shape models

Abstract

A shape model for full automatic segmentation and recognition of lateral lumbar spine radiographs has been developed. The shape model is able to learn the shape variations from a training dataset by a principal component analysis of the shape information. Furthermore, specific image features at each contour point are added into models of gray value profiles. These models were computed from a training dataset consisting of 25 manually segmented lumbar spines. The application of the model containing both shape and image information is optimized on unknown images using a simulated annealing search first to acquire a coarse localization of the model. Further on, the shape points are iteratively moved towards image structures matching the gray value models. During optimization the shape information of the model assures that the segmented object boundary stays plausible. The shape model was tested on 65 unknown images achieving a mean segmentation accuracy of 88% measured from the percental cover of the resulting and manually drawn contours.