Phase mutual information as a similarity measure for registration

This article describes a new method for non-rigid alignment of multimodal images. Multimodal image registration is most often accomplished by modelling, in some sense, an intensity mapping between the images. Here, the alternative strategy of modelling a relationship between local image phase is introduced. This method is intrinsically image feature based, and searches for relationships between feature appearances, rather than tissue class intensity. This enables registration of modalities for which image intensity is not a simple function of tissue class, for example ultrasound. It is also demonstrated that this method performs comparably to an intensity method even when the images are related by a simple intensity transform, but that the phase method is significantly more robust to image artifacts which corrupt the ideal intensity mapping.