Introducing Learning Rate Adaptation CMA-ES into Rigid 2D/3D Registration for Robotic Navigation in Spine Surgery

TL;DR

This paper introduces a learning rate adaptation mechanism into CMA-ES for 2D/3D registration in robotic spine surgery, achieving faster and more accurate registration with reduced computational costs.

Contribution

The paper presents a novel framework that integrates learning rate adaptation into CMA-ES, enabling effective registration with smaller populations and lower computational expense.

Findings

Superiority in registration accuracy over baselines

Reduced runtime compared to traditional CMA-ES

Effective in complex surgical registration scenarios

Abstract

The covariance matrix adaptive evolution strategy (CMA-ES) has been widely used in the field of 2D/3D registration in recent years. This optimization method exhibits exceptional robustness and usability for complex surgical scenarios. However, due to the inherent ill-posed nature of the 2D/3D registration task and the presence of numerous local minima in the landscape of similarity measures. Evolution strategies often require a larger population size in each generation in each generation to ensure the stability of registration and the globality and effectiveness of search, which makes the entire process computationally expensive. In this paper, we build a 2D/3D registration framework based on a learning rate adaptation CMA-ES manner. The framework employs a fixed and small population size, leading to minimized runtime and optimal utilization of computing resources. We conduct experimental comparisons between the proposed framework and other intensity-based baselines using a substantial volume of synthetic data. The results suggests that our method demonstrates superiority in both registration accuracy and running time. Code is available at github.com/m1nhengChen/CMAES-reg.