AI-Assisted Dose and Dose Rate Optimization for Deliverable Proton Conformal Flash Radiotherapy
Abstract
Purpose
Proton conformal FLASH radiotherapy must simultaneously satisfy clinical dosimetric goals and deliver ultra-high-dose-rates under machine constraints. This highly coupled optimization problem is currently addressed through manual trial-and-error, limiting plan optimality and efficiency, and hindering clinical translation. This study aims to develop an AI-assisted Dose And dose-Rate opTimization (DART) framework to autonomously generate FLASH plans simultaneously meeting the dose and dose-rate criteria.
Methods
Fifteen head-and-neck reirradiation patients (40 Gy in 5 fractions) were retrospectively selected under IRB approval. Proton conformal FLASH planning was performed in RayStation-FLASH treatment planning system configured for an IBA ConformalFLASH delivery system. The DART framework integrated a Deep Reinforcement Learning (DRL)-based auto-planning agent and post-planning spot sequence optimization. The planning agent interfaced with RayStation-FLASH to autonomously guide inverse optimization. It received plan information, including dose-volume histograms and dose-rate-volume histograms, and output continuous parameter updates governing dose and dose-rate delivery. Training was conducted using a policy-based DRL approach guided by a composite reward function that jointly evaluated dosimetric and dose-rate performance. Following plan generation, raster scanning sequences were optimized using a genetic algorithm to maximize dose-rate. Plan integrity was evaluated via log-file-based dose and dose-rate reconstruction.
Results
The trained DRL agent autonomously operated RayStation-FLASH to generate clinically acceptable FLASH plans with clinically acceptable dose quality, achieving organs-at-risk sparing and reduced hotspots comparable to manual planning while substantially increasing high dose-rate coverage (>40GyRBE/s). Across all organs-at-risk, the V40GyRBE/s, V60GyRBE/s, V80GyRBE/s, and V100GyRBE/s improved by 12.40%, 9.40%, 8.47%, and 5.10% on average. The largest dose-rate gains were observed in the parotids, submandibular glands, and larynx, with V40GyRBE/s increases of 19.30%, 17.23%, and 10.21%. Post-planning spot sequence optimization further enhanced dose-rates without altering dose quality.
Conclusion
We developed and validated the AI-driven DART for proton conformal FLASH head-and-neck reirradiation, generating high-quality and deliverable plans with optimized dose and dose-rate.