Bayesian Optimization-Based Automatic Inverse Treatment Planning for Gamma Knife Radiosurgery for Patients with Multiple Brain Metastases
Abstract
Purpose
Gamma Knife (GK) radiosurgery is a specialized radiosurgery for brain metastases (BMs). GK plan quality depends strongly on target size and shape, making conventional plan metrics difficult to compare across patients. Dosimetric quality is also coupled with beam-on time (BOT). Thereby, no standardized definition of optimal plan metrics or planning preferences exist for GK. Clinical workflows rely on repeated manual priority adjustment among objectives to achieve physician-satisfactory plans. We propose an automated priority selection framework using Bayesian optimization (BO) to reduce human intervention.
Methods
We employed a linear programming-based inverse planning model, with four priorities controlling target underdose, overdose to inner and outer shells, and BOT. Priority tuning was formulated as a black-box optimization problem over these four priorities. BO employed a Gaussian process surrogate with expected improvement acquisition function to efficiently explore a five-order-of-magnitude search space per priority. Intermediate plan quality was scored using piecewise-linear functions, relative to either a reference plan quality obtained from similar historical cases or a user-specified planning goal. The conformity index at 50% prescription dose (CI50) was used instead of gradient index to allow direct comparison of plans with varying conformity.
Results
A retrospective study was conducted on 46 patients with 128 BMs, with mean lesion volume of 0.28±0.52 cc. All BO-generated plans met the clinical requirement of >99% coverage. Compared to clinical plans, BO improved all plan metrics in 11 BMs and improved all three dosimetric metrics at the cost of longer BOT in 61 BMs. On average, BO improved selectivity (0.65 vs. 0.52) and R50 (8.48 vs. 10.99) with comparable coverage (99.84% vs. 99.67%), at a modest BOT increase (20.25 vs. 17.81 min).
Conclusion
BO enables automated GK priority weight tuning and produces clinically acceptable plans. It reduces the need for manual trial-and-error adjustments and readily adapts to diverse planning preferences.