matic bone segmentation in ultrasound images using local phase features and dynamic programming

We present a novel method for bone structure segmentation in two-dimensional (2D) ultrasound (US) images as a precursor to 3D bone surface reconstruction and registration. The main contributions of this paper are to develop a dynamic programming segmentation solution that: (a) eliminates the soft tissues above a bone structure by taking into consideration acoustic characteristics of the intensity profile along each US scan line, including the integrated backscattering (IBS) and acoustic shadows; and (b) combines the local energy, the local phase and local phase feature symmetry to highlight areas of the image that have a high probability of being bone structures. The automatic segmentation results were compared to manual segmentation ground truth carried out by clinical experts. The average Euclidean distance (ED) error between the two methods was less than 2 pixels (approximately 0.2mm). Our method significantly decreases the number of erroneous detections of the soft tissue compared to existing methods.