A robust similarity measure for nonrigid image registration with outliers

Nonrigid image registration is a widely used technique in medical imaging. While most methods work very well on images without pathologies or artefacts, there is a high need for improved robustness on images from pathological subjects and acquisitions with artefacts such as intensity inhomogeneity. In this paper, we propose a novel similarity measure based on normalised gradients for nonrigid registration, which is robust on images with intensity inhomogeneities or pathologies. We provide both theoretical and experimental proof of the robustness and evaluate the approach on manually segmented and simulated pathological images. Compared to normalised mutual information and to an alternative similarity also based on normalised gradients, we obtain significant overlap improvements for images with intensity inhomogeneities. We further confirm improved robustness on images with simulated tumours.