No abstract

Segmentation of the carpal bones (CBs) especially for children above seven years old is a challenging task in computer vision mainly because of poor definitions of the bone contours and the occurrence of the partial overlapping of the bones. Although active contour methods are widely employed in image bone segmentation, they are sensitive to initialization and have limitation in segmenting overlapping objects. Thus, there is a need for a robust segmentation method for bone segmentation. This paper presents an automatic active boundary-based segmentation method, gradient inverse coefficient of variation, based on dynamic programming (DP-GICOV) method to segment carpal bones on radiographic images of children age 5 to 8 years old. A mapping procedure is designed based on a priori knowledge about the natural growth and the arrangement of carpal bones in human body. The accuracy of the DP-GICOV is compared qualitatively and quantitatively with the de-regularized level set (DRLS) and multi-scale gradient vector flow (MGVF) on a dataset of 20 images of carpal bones from University of Southern California. The presented method is capable to detect the bone boundaries fast and accurate. Results show that the DP-GICOV is highly accurate especially for overlapping bones, which is more than 85% in many cases, and it requires minimal user's intervention. This method has produced a promised result in overcoming both issues faced by active contours method; initialization and overlapping objects.