Multimodal 3D rigid image registration based on expectation maximization

Image registration is an important task in medical imaging, capable of finding displacement fields to align two images of the same anatomic structure under different conditions (e.g. acquisition time and body position). Specifically, multimodal image registration is the process of aligning two or more images of the same scene using different image acquisition techniques. In fact, most of the current image registration approaches are based on Mutual Information (MI) as a similarity metric for image comparison; however, the cost function used in MI methods is difficult to optimize due to complex relationships between variables and pixels intensities. This work presents an Expectation Maximization (EM) 3D multimodal rigid registration approach, which introduces a low computational cost alternative with a linear optimization strategy and an intuitive relation among the free variables. Our approach was validated against a state-of-the-art MI-based technique with synthetic T1 MRI brain volumes. The EM 3D achieved a global average DICE index of 96.68 % with a computational time of 22.72 seconds, whereas the MI methodology reported 96.11 % and 35.13 seconds, respectively.