Automated spine and vertebrae detection in CT images using object-based image analysis

Although computer assistance has become common in medical practice, some of the most challenging tasks that remain unsolved are in the area of automatic detection and recognition. The human visual perception is in general far superior to computer vision algorithms. Object-based image analysis is a relatively new approach that aims to lift image analysis from a pixel-based processing to a semantic region-based processing of images. It allows effective integration of reasoning processes and contextual concepts into the recognition method. In this paper, we present an approach that applies object-based image analysis to the task of detecting the spine in computed tomography images. A spine detection would be of great benefit in several contexts, from the automatic labeling of vertebrae to the assessment of spinal pathologies. We show with our approach how region-based features, contextual information and domain knowledge, especially concerning the typical shape and structure of the spine and its components, can be used effectively in the analysis process. The results of our approach are promising with a detection rate for vertebral bodies of 96% and a precision of 99%. We also gain a good two-dimensional segmentation of the spine along the more central slices and a coarse three-dimensional segmentation.