An Image Registration Approach Based on 3D Geometric Projection Similarity of the Human Head

Abstract

This study aims to develop a novel brain image registration algorithm, the geometry projection similarity (GPS) algorithm, which utilizes the anatomical feature of human head and simplifies the objective function for optimization and search for registration parameters in a step-by-step fashion, largely reducing the solution space and improving the efficiency. The algorithm was tested and compared with existing algorithms in simulation environment and tested on computed tomography/magnetic resonance imaging (CT/MRI) clinical datasets pre- and post-craniotomy surgery. Compared with the maximal mutual information algorithm, the GPS algorithm showed significantly smaller errors (1.2 ± 1.1 vs. 4.5 ± 3.5, 1.0 ± 1.0 vs. 3.5 ± 2.5, 0.8 ± 1.1 vs. 2.8 ± 2.4 pixels, at cluster size of 100, 150 and 200, respectively) and shorter computational durations. Compared with the iterative closest point algorithm, the GPS algorithm showed a greater robustness to the perturbed initial mismatch between registering images (registration failure rate 0 vs. 29 out of 200 trials), while keeping a satisfactorily low registration error (0.71 ± 0.83 pixels). Overall results demonstrated that registration accuracy, efficiency, and robustness of this new algorithm surpass those of existing methods. Pilot test on CT/MRI clinical datasets also showed a satisfactory outcome, demonstrating great potentials for future applications of registration of brain images across modalities.