Evaluating effectiveness of clustering algorithms in multiple target stereotactic radiosurgery

Abstract

Objective. Single-isocenter-multiple-target technique for stereotactic radiosurgery (SRS) can reduce treatment duration but risks compromised dose coverage due to potential rotational errors. Clustering targets into two groups can reduce isocenter-target distances, mitigating the impact of rotational uncertainty. However, a comprehensive evaluation of clustering algorithms for SRS is absent. This study addresses this gap by introducing the SRS Target Clustering Framework (Framework), a comprehensive tool that utilizes commonly used clustering algorithms to generate efficient cluster configurations. Approach. The Framework incorporates four distinct optimization objectives based on two key metrics: the isocenter-target distance and the ratio of this distance to the target radius. Agglomerative and weighted agglomerative clustering are employed for minimax and weighted minimax objectives, respectively. K-means and weighted k-means are utilized for sum-of-squares and weighted sum-of-squares objectives. We applied the Framework to 126 SRS plans, comparing results to ground truth solutions obtained through a brute force algorithm. Main results. For the minimax objective, the average maximum isocenter-target distance from agglomerative clustering (4.8 cm) was slightly higher than the ground truth (4.6 cm). Similarly, the weighted agglomerative clustering achieved an average maximum ratio of 15.1 compared to the ground truth of 14.6. Notably, both k-means and weighted k-means clustering showed close agreement (within a precision of 0.1) with the ground truth for average root-mean-square target-isocenter distance and ratio (3.6 cm and 11.1, respectively). Significance. These results demonstrate the Framework’s effectiveness in generating clusters for SRS targets. The proposed approach has the potential to become a valuable tool in SRS treatment planning. Furthermore, this study is the first to investigate clustering algorithms for both minimizing maximum and sum-of-squares uncertainty in SRS.