Knowledge based automated planning for non-coplanar VMAT stereotactic radiosurgery in ocular malignancies

Background and purpose: Stereotactic radiosurgery/radiotherapy (SRS/SRT) has emerged as a less invasive alternative to enucleation in the management of ocular malignancies. SRS/SRT planning is time-consuming, complex, and the plan quality depends on the experience of the treatment planner. We demonstrate the use of RapidPlan modeling in conjunction with HyperArc for the treatment of ocular diseases to improve planning efficiency, consistency, and quality. Materials and methods: A HyperArc-based RapidPlan (HARP) model was iteratively trained on 80 patient datasets that simulated ocular malignancies. Twenty additional patient datasets were reserved for model testing and comparison with manual plans. Each testing dataset was both manually planned and replanned with the final RapidPlan model. Target volumes were defined in the HyperArc module as the GTV and PTV. The PTV was generated by a 2 mm static expansion of the GTV. 25 Gy in 1 fraction was prescribed, and all plans were normalized such that PTVD_95% = 25 Gy. Treatment planning was done using Eclipse V16 with the Acuros XB dose engine on a TrueBeam LINAC with Millennium 120 MLCs (5 mm width). All plans underwent EPID-based patient-specific QA and an independent Monte Carlo (MC) second-check. Results: Model-based plans demonstrated similar degrees of conformality, gradient, and target coverage compared to manual planning. OAR sparing showed statistically significant improvements with model-based plans. Optic nerve Dmax (D0.03 cc) decreased to 4.56 Gy (manual: 7.18 Gy), and lens Dmax decreased to 7.64 Gy (manual: 9.36 Gy). Lacrimal gland mean dose decreased from 11.61 to 6.47 Gy. Modulation factor, monitor units, and beam-on times also all decreased. Patient-specific QA results showed improvements compared to manually generated plans. MC second check results also improved, increasing from an average of 97.81%–98.74%. Plan optimization times decreased significantly from approximately 120 min for manual planning to 15 min on average for model-based plans. Conclusion: RapidPlan-generated ocular SRS plans showed either comparable performance or improvements in all plan metrics measured, compared to manual planning. Planning times were significantly reduced while maintaining or improving plan quality. Clinical implementation of this HARP model is ongoing at our clinic.