A Study on Performace of Levenberg-Marquardt and CMA-ES Optimization Methods for Atlas-based 2D/3D Reconstruction

Abstract

Abstract: In this study, we compare the performance of our previously proposed deformable 2D/3D registration approach based on the Levenberg-Marquardt optimization with methods exploiting Covariance Matrix Adaptation (CMA) and Covariance Matrix Self Adaptation (CMSA) evolution strategies. The aim of the registration is to reconstruct a patient-specifc 3D bone model from a small set of plain 2D X-ray images what is achieved by ftting a deformable bone atlas onto the X-ray images. The comparison of diferent optimization methods is focused on both the robustness and the speed. The results were obtained using a large-scale data set of synthetic X-ray images. We show that our method is several times faster in comparison with the approaches based on evolution strategies while the robustness of the reconstruction is preserved. To speed-up the reconstruction process, certain parts of the registration pipeline are accelerated using graphics hardware. The median error of our proposed method was 1.12mm and the median reconstruction time was 7.2s. The median time reached by the CMA-ES and CMSA-ES methods was 48.5 s and 138.5 s respectively.