Abstract Background In Gamma Knife (GK) radiosurgery, effective treatment planning aims to achieve a desired dose distribution while minimizing the overall treatment duration. Clinically, plans are generated either manually through forward planning, or using inverse planning methods. To address the limitations presented by forward and inverse planning, a knowledge‐based planning (KBP) pipeline was recently developed for GK, using 3D dose predictions in combination with inverse optimization. However, this approach still relied on manually selected isocenter locations, which may limit plan quality. Purpose To develop a complete KBP pipeline for GK, combining dose prediction with the simultaneous optimization of isocenter locations and beam‐on‐times. Methods Data for 20 patients were obtained from a local care center. A previously trained deep learning model generated 3D dose predictions for each case. The primary treatment planning model was a mixed‐integer model (GK‐KBP‐OptIso) that combined isocenter selection with beam‐on‐time optimization. Plans were generated under several different maximum allowable isocenter limits and compared against (1) plans generated using an existing KBP pipeline with fixed, manually selected isocenters (GK‐KBP‐FixIso) and (2) historical clinical plans. Plan quality was assessed using coverage, selectivity, Paddick conformity, and gradient indices, along with overall treatment time. Results Under the same isocenter limits, GK‐KBP‐OptIso achieved comparable conformity to GK‐KBP‐FixIso ( vs ) while significantly improving dose falloff. It also matched clinical plan quality, though with slightly increased treatment times. However, when allowed to use variable number of isocenters, GK‐KBP‐OptIso produced plans with a higher average conformity () than clinical plans (). Conclusions A KBP pipeline that integrates isocenter selection yields plans that are equal or superior to those generated using KBP with manually selected isocenters or manual clinical planning. The presented approach holds the potential to streamline the effort required to generate high‐quality GK treatment plans.
Zhang et al. (Fri,) studied this question.