FASTR: Using Local Structure Tensors as a Similarity Metric

We describe a novel structural image descriptor for image registration called the Fractionally Anisotropic Structural Tensor Representation (FASTR), calculated from the local structural tensor (LST). The metric has several characteristics that are advantageous for multi-modality registration, such as not depending on absolute voxel intensities, and being insensitive to slowly varying in- tensity inhomogeneities across the image. This latter property is very useful, since many imaging modalities suffer from such artefacts. Registration accuracy is tested on both computed tomography (CT) to cone-beam CT (CBCT) rigid registration, and CT to magnetic resonance (MR) rigid registration. The performance is compared with Mutual Information (MI) metric and the Self Similarity Context (SSC) descriptor. The results show that, for images with significant intensity inhomogeneity, FASTR produced more accurate results than MI, and faster results than SSC. The results suggest FASTR gives similar benefits in images with intensity inhomogeneity, but at a fraction of the computation and memory demand.