Automated three‐dimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures

TLDR

Voxel‑intensity similarity measures show promise for fully automated MR‑PET brain image registration without requiring structure extraction. This study presents a multiresolution optimization method and compares five voxel‑intensity similarity measures, including cross‑correlation, PET intensity variation minimization, intensity distribution moments, entropy, and mutual information. The method evaluates registration recovery on ten clinical MR‑PET pairs, assessing sensitivity to initial misregistration, final precision, and visual accuracy, with success defined as a visually acceptable solution to a trained observer. Mutual information proved most robust, registering 98.8 % of 900 trial misregistrations, achieving sub‑millimetre translation and sub‑degree rotation precision in high‑resolution searches, and completing on a typical workstation in 4–8 minutes.

Abstract

Approaches using measures of voxel intensity similarity are showing promise in fully automating magnetic resonance (MR) and positron emission tomography (PET) image registration in the head, without requiring extraction and identification of corresponding structures. In this paper a method of multiresolution optimization of these measures is described and five alternative measures are compared: cross correlation, minimization of corresponding PET intensity variation, moments of the distribution of values in the intensity feature space, entropy of the intensity feature space and mutual information. Their ability to recover registration is examined for ten clinically acquired image pairs with respect to the size of initial misregistration, the precision of the final result, and the accuracy assessed by visual inspection. The mutual information measure proved the most robust to initial starting estimate, successfully registering 98.8% of 900 trial misregistrations. Success is defined as providing a visually acceptable solution to a trained observer. A high resolution search ( mm step size) of 30 trial misregistrations showed that optimization using the mutual information measure provided solutions with 0.13 mm, 0.11 mm and 0.17 mm standard deviations in the three Cartesian axes of the translation vector and 0.2°, 0.3° and 0.2° standard deviations for rotations about the three axes. The algorithm takes between 4 and 8 minutes to run on a typical workstation, including visual inspection of the result.