Automated detection and segmentation of cylindrical fragments from calibrated C-arm images for long bone fracture reduction

Long bone fracture belongs to one of the most common injuries encountered in clinical routine trauma surgery. Automated identification, pose and size estimation, and contour extraction of diaphyseal bone fragments can greatly improve the usability of a computer-assisted, fluoroscopy-based navigation system for long bone fracture reduction. In this paper, a two-step solution is proposed. In the first step, the pose and size of a diaphyseal fragment are estimated through a three-dimensional (3D) morphable object-based fitting process using a parametric cylinder model. This fitting process is optimally solved by a hybrid optimization technique coupling a random sample consensus (RANSAC) paradigm and an iterative closest point (ICP) matching procedure. Monte Carlo simulation was used to determine the parameters for the RANSAC paradigm. The results of the fragment detection step are then fed to the second step, where a region information based active contour model is used to extract the fragment contours. We designed and conducted experiments to quantify the accuracy and robustness of the proposed approach. Our experimental results conducted on images of a plastic bone as well as on those of patients demonstrate a promising accuracy and robustness of the proposed approach.