Interobserver Study on HDR Prostate—Human-Selected Versus AI-Selected Plans

Purpose The development of the multicriteria optimization algorithm in high-dose-rate brachytherapy brings a shift in the objective. Instead of generating an acceptable plan, the objective is to choose the best plans out of thousands. To tackle this new objective, an artificial intelligence (AI) approach was developed. The AI approach of selecting a plan from the multicriteria optimization pool needs to be validated in a retrospective study. Methods and Materials The AI approach uses new criteria, named spatial criteria, on top of the more commonly used dose-volume histogram criteria. With these criteria, AI models can be tuned to one's liking. In this study, 2 different sets of criteria are used to train models, which are then used on 20 previously treated cases. The plans chosen by the AI are evaluated in addition to the clinically selected plan in a blind study. A 5-point Likert scale is used to evaluate each plan with 4 radiation oncologists (RO) and 2 medical physicists. Results No statistical difference is observed between the AI approach and the clinical plans (for the ROs); statistical differences are observed for the medical physicists. The AI-selected plans are preferred or equivalent more than half of the time by the ROs. The AI approach is fast (on average, 12 seconds) to rank the plan of a single plan pool. Conclusions The study confirms our hypothesis that the plans chosen by AI can be at least as good as the plans chosen by the medical staff. The use of AI models allows for fast selection of plans.