Towards an Ultra-Fast GPU-Based Multi-Criteria Optimization Algorithm for HDR Brachytherapy

Abstract

Most inverse planning optimization algorithms used in clinic such as Inverse Planning Simulated Annealing (IPSA) are designed to quickly optimize a single treatment plan based on a population-based template objective function called class solution. Since the weight values of each individual objective are patient specific, the user may have to manually change the weight values of the class solution and make several attempts to obtain a plan that encapsulates all the goals set forth by the physicians. This iterative process can be time consuming and can be dependent on the optimization skills of the user. This study intends to facilitate this process by proposing a graphics processing units (GPU)-based L-BFGS or gL-BFGS optimizer (Limited-memory Broyden Fletcher Goldfarb Shanno) which can calculate multiple plans in parallel for multi-criteria optimization (MCO) for high-dose-rate (HDR) brachytherapy. Thus a patient-specific plan pool of Pareto optimal plans can be generated in real-time for each patient from which the user can chose the most suitable option depending on the clinical situation.