3D/2D Image Registration Using Weighted Histogram of Gradient Directions

Three dimensional (3D) to two dimensional (2D) image registration is crucial in many medical applications such as image-guided evaluation of musculoskeletal disorders. One of the key problems is to estimate the 3D CT-reconstructed bone model positions (translation and rotation) which maximize the similarity between the digitally reconstructed radiographs (DRRs) and the 2D fluoroscopic images using a registration method. This problem is computational-intensive due to a large search space and the complicated DRR generation process. Also, finding a similarity measure which converges to the global optimum instead of local optima adds to the challenge. To circumvent these issues, most existing registration methods need a manual initialization, which requires user interaction and is prone to human error. In this paper, we introduce a novel feature-based registration method using the weighted histogram of gradient directions of images. This method simplifies the computation by searching the parameter space (rotation and translation) sequentially rather than simultaneously. In our numeric simulation experiments, the proposed registration algorithm was able to achieve sub-millimeter and sub-degree accuracies. Moreover, our method is robust to the initial guess. It can tolerate up to 90'rotation offset from the global optimal solution, which minimizes the need for human interaction to initialize the algorithm.