367 - Plan Any Tumor (PAT): A Generalizable, Site-Agnostic Autonomous Radiotherapy Planning System across Multiple Cancer Sites

Purpose/Objective(s): Manual radiotherapy treatment planning remains labor-intensive and subject to variability in efficiency and plan quality. While recent artificial intelligence (AI) approaches have shown promise in automating planning, most existing methods are developed for specific tumor sites and lack generalizability. Driven by the goal of mimicking the decision-making and adaptability of a skilled human dosimetrist capable of planning across multiple disease sites, we developed the Plan Any Tumor (PAT) model, a tumor site-agnostic, autonomous planning framework that leverages a GPT-5–empowered virtual planner to operate the commercial RayStation treatment planning system (TPS) in a dosimetrist-like manner, enabling consistent and efficient automated planning across diverse cancer sites.

Materials/Methods: PAT was implemented as a standalone Python program that interacts with TPS through its scripting interface and an internally hosted GPT-5 platform approved for institutional clinical use. Workflow was built for five representative sites: prostate, gynecologic, lung, pancreas, and head-and-neck. For each case, PAT extracted physician-prescribed planning objectives and conduct initial preparation steps, such as generating optimization structures and defining planning parameters. The optimization engine was invoked to generate an initial plan. PAT summarizes plan quality and generates input to GPT-5 model to iteratively adjust optimization objective weighting factors in a closed loop to emulate the decision process of a human dosimetrist, continuing until clinical acceptability was achieved. The framework was evaluated on ten independent test cases spanning the five tumor sites (two per site). Plan quality was assessed through dosimetric comparison with clinically approved plans, and total planning time was recorded.

Results: PAT successfully generated clinically acceptable plans for all test cases, demonstrating robust generalizability without site-specific customization. Dosimetric performance was comparable to clinically approved plans, with mean differences in target metrics of 0.38% for PTV D_max, 0.60% for D₉₅, and 1.12% for V₁₀₀. For organs at risk across different sites, mean differences were 0.84% for D_max and 2.67% for D_mean. All plans met institutional clinical criteria, and no statistically significant differences were observed in dose-volume metrics relative to clinical plans (p=0.26). Average end-to-end planning time was ~2 hours per case.

Conclusion: By emulating the workflow and decision-making of a skilled human dosimetrist through an agentic, closed-loop optimization process, the PAT framework represents a step toward fully autonomous treatment planning by introducing a site-agnostic, AI-driven virtual planner and adapting across multiple disease sites. PAT demonstrates the potential to streamline clinical operations, reduce planning workload and variability, and improve scalability of high-quality treatment planning.