Learning-based 2D/3D rigid registration using Jensen-Shannon divergence for image-guided surgery

Registration of 3D volumetric data to 2D X-ray images has many applications in image-guided surgery, varying from verification of patient position to working projection searching. In this work, we propose a learning-based method that incorporates the prior information on the expected joint intensity histogram for robust real-time 2D/3D registration. Jensen-Shannon divergence (JSD) is used to quantify the statistical (dis)similarity between the observed and expected joint histograms, and is shown to be superior to Kullback-Leibler divergence (KLD) in its symmetry, being theoretically upper-bounded, and well-defined with histogram non-continuity. A nonlinear histogram mapping technique is proposed to handle the intensity difference between the observed data and the training data so that the learned prior can be used for registration of a wide range of data subject to intensity variations. We applied the proposed method on synthetic, phantom and clinical data. Experimental results demonstrated that a combination of the prior knowledge and the low-level similarity measure between the images being registered led to a more robust and accurate registration in comparison with the cases where either of the two factors was used alone as the driving force for registration.