Non-rigid Atlas-to-Image Registration by Minimization of Class-Conditional Image Entropy

We propose a new similarity measure for atlas-to-image matching in the context of atlas-driven intensity-based tissue classification of MR brain images. The new measure directly matches probabilistic tissue class labels to study image intensities, without need for an atlas MR template. Non-rigid warping of the atlas to the study image is achieved by free-form deformation using a viscous fluid regularizer such that mutual information (MI) between atlas class labels and study image intensities is maximized. The new registration measure is compared with the standard approach of maximization of MI between atlas and study images intensities. Our results show that the proposed registration scheme indeed improves segmentation quality, in the sense that the segmentations obtained using the atlas warped with the proposed non-rigid registration scheme better explain the study image data than the segmentations obtained with the atlas warped using standard intensity-based MI.