The use of a multiobjective evolutionary algorithm to increase flexibility in the search for better IMRT plans

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Intensity-modulated radiation therapy (IMRT) optimization aims to maximize tumor coverage while minimizing dose to organs at risk (OARs).
• Traditional IMRT optimization methods may be limited by a constrained search space, potentially affecting the quality of achievable treatment plans.
• Multiobjective optimization techniques offer a way to explore a wider range of trade-offs between competing objectives.

Purpose of the Study:

• To evaluate the impact of a more flexible and comprehensive multiobjective search on IMRT plan optimization performance.
• To investigate how expanding the search space influences the quality of IMRT plans generated by a multiobjective evolutionary algorithm (MOEA).

Main Methods:

• A multiobjective evolutionary algorithm (MOEA) was employed to explore expanded search spaces for IMRT optimization.
• Three strategies were used: simultaneous optimization of weights and dose parameters, voxel-specific penalty functions, and a heuristic voxel-specific improvement (VSI) algorithm.
• Plans were evaluated using independent, clinically relevant criteria, with dominated plans eliminated to approach the Pareto front.

Main Results:

• Simultaneous optimization of weights and dose parameters yielded superior IMRT plans compared to fixed parameters, especially with overlapping structures.
• Voxel-specific penalty functions improved results in simpler cases but not complex ones, with benefits increasing as structure overlap decreased.
• The VSI algorithm consistently improved plan quality across all tested cases, including complex prostate and head and neck scenarios.

Conclusions:

• Increased flexibility in the search space, through varying objective function parameters or removing uniform penalty constraints, enhances IMRT optimization performance.
• Strategies like optimizing weights/doses and employing voxel-specific penalties generated plans that dominated conventionally Pareto optimal sets.
• While large search spaces can cause convergence issues in MOEAs for complex cases, the VSI algorithm remains effective.