Radiotherapy Planning Using an Improved Search Strategy in Particle Swarm Optimization.

Area of Science:

• Computational physics
• Medical physics
• Optimization algorithms

Background:

• Evolutionary stochastic global optimization algorithms are crucial for large-scale, nonconvex problems.
• Enhancing search efficiency and repeatability in these algorithms often requires customized solutions.
• Radiation therapy (RT) planning presents a complex, large-scale, nonconvex optimization challenge.

Purpose of the Study:

• To investigate a customized approach for enhancing evolutionary stochastic global optimization algorithms.
• To apply Particle Swarm Optimization (PSO) to a 4D radiation therapy inverse planning problem, incorporating respiratory motion.
• To compare a proposed virtual search method against conventional PSO and a Dynamically Penalized Likelihood (DPL) algorithm.

Main Methods:

• Utilized Particle Swarm Optimization (PSO) for a 4D lung cancer RT inverse planning problem, optimizing radiation fluence-weights across respiratory phases.
• Implemented three PSO variants: unconstrained, hard-constrained, and a novel virtual search approach.
• Compared the performance of the proposed virtual search PSO against a conventional Dynamically Penalized Likelihood (DPL) algorithm using five patient cases.

Main Results:

• The proposed virtual search technique demonstrated significant improvements in robustness to random initialization.
• Fewer iteration cycles were required for convergence with the virtual search approach across all tested cases.
• The Dynamically Penalized Likelihood (DPL) algorithm found the global optimum in only 2 out of 5 cases, requiring substantially more iterations.

Conclusions:

• The virtual search approach substantially enhances swarm search efficiency for PSO.
• This leads to improved optimization convergence rates and robustness in complex planning scenarios.
• The proposed method offers a potential improvement for time-sensitive optimization problems in RT planning.