Radiosurgery is a non-invasive cancer treatment method in which a tumor is irradiated from multiple directions using narrow beams of ionizing radiation. The radiation that eliminates the cancerous tissue also damages healthy tissue and organs surrounding the tumor, creating a tradeoff between adequate tumor coverage and sparing of healthy structures. Treatment planning systems have been developed to address this tradeoff by formulating it as an optimization problem. RayStation, developed by RaySearch
Laboratories AB, is one such system.

This master’s thesis extends RayStation with a prototype that enables treatment planning for gyroscopic radiosurgery platforms, which is currently not supported. To achieve this, a prototype for the ZAP-X platform, which is one such system developed by ZAP Surgical, is implemented. Then, using this prototype, both existing and novel optimization methods for producing ZAP-X treatment plans are implemented and evaluated.

Two published methods are surveyed and implemented: the sphere packing method and the clustering method. Both determine isocenter positions and collimator sizes through heuristic algorithms, after which only the irradiation times are optimized. Two novel contributions are then developed. The first extends the optimization to include the isocenter positions and collimator sizes as continuous variables within a Sequential Quadratic Programming framework. The second introduces filtering, a technique that removes low contribution control points during optimization to reduce plan complexity while preserving plan quality.

The methods are evaluated on three planning target volumes with different characteristics. Continuously optimizing the isocenter positions leads to substantially better plan quality than the existing methods across all cases. Filtering removes up to 90% of the control points without degrading plan quality, which has the potential to significantly reduce delivery times. These improvements come at a computational cost: continuous isocenter optimization increases computation times by roughly a factor of eight, and continuous collimator size optimization proves too expensive for clinically realistic plan sizes.