医学图像配准的多策略优化方法库及LLM动态调度研究

Research on Multi-Strategy Optimization Method Library and LLM Dynamic Scheduling for Medical Image Registration

摘要: 医学图像配准中，不同优化方法在不同样本和优化阶段的表现并不一致，单一固定算法往往难以兼顾精度与稳定性。针对CT-MRI多模态配准中的9参数受限仿射变换问题，文章构建了包含8种方法的候选库，并在三组等预算配置下比较其整体精度、阶段收敛特征和跨样本稳定性。结果表明，各方法之间存在明显互补性：CRO-ZZ-QLS整体表现最好，Powell收敛较快但容易停留在局部最优，CRO-ZZ-QLS与CRO-SL在跨样本上更稳定。在此基础上，进一步提出基于LLM的动态调度方法，使其在等评估预算下从候选算法与种群配置中进行离散选择。实验结果显示，LLM调度的平均NMI为0.2388，较最佳静态方法提高0.72%，且14个样本中有9个取得最优结果。说明将全局探索与后期局部精修结合起来，有助于进一步提升配准性能。

Abstract: In multimodal medical image registration, different optimization methods often behave differently across samples and optimization stages, so a single fixed algorithm is usually insufficient. For the 9-parameter constrained affine transformation problem in CT-MRI registration, this paper builds a candidate library of eight methods and compares their overall accuracy, stage-wise convergence, and cross-sample stability under three equal-budget settings. The results show clear complementarity among the methods: CRO-ZZ-QLS achieves the best overall performance, Powell converges quickly but is more likely to get trapped in local optima, and CRO-ZZ-QLS, together with CRO-SL, shows better stability across samples. Based on these observations, an LLM-based dynamic scheduling method is further introduced to select algorithms and population settings under the same evaluation budget. Experimental results show that the proposed scheduler reaches a mean NMI of 0.2388, improving the best static method by 0.72% and obtaining the best result on 9 of 14 samples. These results suggest that combining global exploration with later local refinement is beneficial for improving registration performance.