Utilizing salient region features for 3D multi-modality medical image registration

Abstract

Besides common solely intensity- or feature-based image registration methods, hybrid approaches make use of two or more image properties. A 3D representation of saliency can be used to automatically locate distinct region features within two 3D images and establish a robust and accurate hybrid registration method. The extracted features contain information about the underlying saliency, the scale of the regions and enclosed voxel intensities. Similar anatomical or functional content results in similar salient region features that can be used to estimate an image transform based on corresponding feature pairs. The refinement of this estimate results in a robust and sub-pixel accurate set of joint correspondences. An evaluation by a medical expert on various clinical data using a medical application demonstrates that the approach is robust to image overlap, artefacts and different fields of view.