Evaluation of computed tomography to ultrasound 2D image registration for atrial fibrillation treatment

This work aims to evaluate intensity-based similarity measures used in rigid or elastic registration techniques to align preoperative computed tomography (CT) and transesophageal ultrasound (US) images of the left atrium and the pulmonary veins, in order to guide high intensity focused ultrasound (HIFU) ablation therapy in patients with atrial fibrillation. The evaluation of the registration techniques has four steps: (i) design of anatomical 2D models including the left atrium and the pulmonary veins from the cryosections of the Visible Human Project, (ii) application of a known rigid/nonrigid transformation in order to produce a gold standard, (iii) creation of a pair of synthetic US and CT images, (iv) estimation of the transformation by rigid/elastic registration of the simulated images and comparison with the gold standard. The performance of eight similarity measures was evaluated for rigid registration: mutual information (MI), normalized mutual information (NMI), entropy correlation coefficient (ECC), joint entropy (H), point similarity measure based on MI (PSMI), energy of histogram (E), correlation ratio (CoR) and Woods criterion (WC). The evaluation protocol estimated the accuracy and the global optimization tendency around the gold standard. For the elastic registration the level of accuracy obtained using MI and NMI was measured.