Automatic segmentation of vertebrae from digitized x-ray images
The segmentation of vertebrae in x-ray images is of prime importance in the assessment of abnormalities of the spine. Manual segmentation is prone to errors due to inter- and intra-subject variabilities due to the subjective judgement that is employed. The use of computer vision methods is, therefore, an attractive alternative to providing an automatic means for segmenting vertebrae. However, general-purpose algorithms present a number of shortcomings that limit their ability to locate and delineate precise vertebral shapes. Therefore, there is a need for a different approach. This work presents the development of an automatic segmentation methodology that employs a hierarchical approach to segmentation. The unique combination of the Generalized Hough Transform, Active Shape Models, and Deformable Models provides three levels of segmentation firom coarse to fine, respectively. Each algorithm has been customized to address the shortcomings of the other two, thus providing a robust framework. Generalized Hough Transform is used to estimate the pose of the spine within a target image. Then, the technique of Active Shape Models is used to find the boundaries of the vertebrae and to give a global approximation to their shape. Finally, the technique of Deformable Models is used to refine the shape of the vertebrae at key points of interest, such as anterior comers. Experimental results with a data set of 100 lateral views of cervical vertebrae and 100 lateral views of lumbar vertebrae have shown a success rate of 75% in finding boundaries of cervical vertebrae and 50% in lumbar vertebrae. The algorithm developed in this work represents a viable alternative to the currently available segmentation methods in which a unique combination of customized algorithms implements a hierarchical firamework.