A region-based algorithm for automatic bone segmentation in volumetric CT

Abstract

In Computed Tomography (CT), bone segmentation is considered an important step to extract bone parameters, which are frequently useful for computer-aided diagnosis, surgery and treatment of many diseases such as osteoporosis. Consequently, the development of accurate and reliable segmentation techniques is essential, since it often provides a great impact on quantitative image analysis and diagnosis outcome. This chapter presents an automated multistep approach for bone segmentation in volumetric CT datasets. It starts with a three-dimensional (3D) watershed operation on an image gradient magnitude. The outcome of the watershed algorithm is an over-partioning image of many 3D regions that can be merged, yielding a meaningful image partitioning. In order to reduce the number of regions, a merging procedure was performed that merges neighbouring regions presenting a mean intensity distribution difference of ±15%. Finally, once all bones have been distinguished in high contrast, the final 3D bone segmentation was achieved by selecting all regions with bone fragments, using the information retrieved by a threshold mask. The bones contours were accurately defined according to the watershed regions outlines instead of considering the thresholding segmentation result. This new method was tested to segment the rib cage on 185 CT images, acquired at the São João Hospital of Porto (Portugal) and evaluated using the dice similarity coefficient as a statistical validation metric, leading to a coefficient mean score of 0.89. This could represent a step forward towards accurate and automatic quantitative analysis in clinical environments and decreasing time-consumption, user dependence and subjectivity.