Unifying Encoding of Spatial Information in Mutual Information for Nonrigid Registration

As encoding spatial information into mutual information (MI) can improve the nonrigid registration against bias fields where the conventional MI is challenged, we propose to unify this encoding into the computation of the joint probability distribution function (PDF). The PDF is computed based on local volumes while the global intensity information is also incorporated to maintain the global intensity class linkage. We demonstrate this computation method can unify the PDF computation in regional MI, conditional MI, and the conventional MI. We then derive two categories of methods and apply them to different registration tasks. The experimental results demonstrate that both categories can significantly improve the registration.