3584 - AI Agent-Based Pre-Plan Auto-Setup for Online Adaptive Radiotherapy Treatment Planning System

Purpose/Objective(s): Adopting online adaptive radiotherapy (o-ART) in an established clinical environment often requires integrating new treatment planning system (TPS) into pre-existing contouring and planning workflows. Asking physicians or planners to transcribe directives across multiple platforms is tedious, prone to errors, and adds significant overhead. Moreover, the learning curve for new planners can be steep when they lack immediate access to experienced planners’ strategies. We developed an AI agent to automate and standardize pre-plan setup, addressing these challenges by minimizing manual transcription, reducing planning time, and incorporating best practices from experienced planners.

Materials/Methods: To validate our AI agent, we integrated it into an X-ray guided o-ART TPS, which requires an XML-based template as a starting point. Given a new patient, the agent automatically retrieves physician dose directives and natural language-based instructions from an in-house database. Since these directives are already well-structured, the agent invokes a tool to parse and convert them into XML-compatible elements. For physician instructions on target formulas, the agent leverages GPT-4o provided by HIPAA compliant Azure OpenAI service to analyze, extract, and standardize the formulas, then converts them into XML-compatible elements. Additionally, for commonly treated normal organs (e.g., lungs, kidneys), the agent automatically generates corresponding formulas as necessary, even if not explicitly mentioned. In a phase-one study, three experienced planners were asked to plan five cases for different treatment sites, where creating a general patient-agnostic template is particularly challenging. For each plan, planners recorded two metrics: the AI-assisted preparation time based our agent and the manual preparation time to complete the full setup, defined as starting from running the TPS to just before optimization.

Results: Depending on case complexity, the AI-assisted preparation time ranged from 44 to 72 minutes, while the full manual setup time ranged from 63 to 98 minutes. This translated to a 13.3% to 27.0% reduction in manual effort.

Conclusion: Our AI agent can efficiently streamline the pre-plan setup process by reducing manual effort and improving standardization. In the next phase, we will extend the system to automate generation of planning structures and optimization priorities, further ensuring consistency in treatment planning and expanding access to advanced radiotherapy techniques. This not only enhances efficiency but also shortens the learning curve for new planners by capturing strategies from experienced planners, making o-ART more accessible across clinical teams.