Generalized clustering-based image registration for multi-modality images

We propose a generalized and intuitive framework for multimodality image registration, based on clustering in the intensity mapping plot (IMP, also known as feature space or joint histogram). Established methods such as Woods', Hill's moment-based, joint-entropy and mutual-information methods can be represented as special cases of this generalization. To register two images, the IMP, which is a 2D scatter plot, is constructed by plotting an (x,y) point for every corresponding voxel pair in the images, x = voxel intensity from image I, y = voxel intensity from image 2. Clusters are formed when the images are registered. The degree of clustering is assessed by assigning a clustering measure (local density of the IMP) to every point in the IMP, and summing the local density of all points. We investigated three issues: 1. the choice of monotonically increasing functions for evaluating the local density, 2. the effects of normalization, and 3. the effects of background elimination. Choices for these three issues allow customization of our generalized framework to previously established methods. By systematic analysis of these choices, we propose new modifications to improve registration accuracy and computational efficiency.