Automated Radiation Target Contouring and Treatment Planning Via RATA: A Multi-Center RCT

Radiotherapy, a cornerstone of comprehensive cancer treatment, relies heavily on clinician expertise for target volume contouring and treatment planning. Consequently, clinical practice is often constrained by procedural bottlenecks, inter-observer variability, and prolonged iterative feedback cycles. Although current artificial intelligence (AI) models have demonstrated progress in automated contouring, key limitations persist, including insufficient integration of multimodal clinical text, a lack of closed-loop dosimetric feedback, and limited generalizability across complex cancer types.

This study aims to develop and validate RATA (Radiotherapy Treatment Multi-Agent Assistant), a novel large model system powered by a collaborative multi-agent architecture. Driven by Large Language Model (LLM) decision engines, RATA processes pathological imaging, electronic health records (EHR), and multimodal medical data to construct a end-to-end, fully automated closed-loop pipeline encompassing "target contouring, plan design, dosimetric evaluation, and reflective refinement." Innovatively breaking through traditional unidirectional radiotherapy workflows, the system establishes a dosimetric feedback-driven adaptive iteration mechanism, leveraging dose distribution outcomes to guide multi-agent dynamic fine-tuning of target volumes, thereby mimicking the collaborative clinical logic between senior radiation oncologists and medical physicists.

For validation, this study adopts a multi-center, three-stage clinical validation framework-comprising multi-center retrospective validation, human-AI collaborative efficacy assessment, and a prospective randomized controlled trial (RCT)-to comprehensively evaluate the system's performance in improving contouring precision, mitigating experience-dependent disparities across clinician seniority, and reducing total workflow processing time. Implementation of the RATA platform holds potential to eliminate redundant operational steps, enabling a minute-level "scan-to-treat" response. Furthermore, its natural language interaction interface optimizes human-AI collaboration and substantially reduces clinical workload. Overall, this study not only establishes a standardized, intelligent assistance paradigm for precision radiation oncology but also provides essential evidence-based support and technological foundations for the clinical translation of multimodal AI in complex cancer management.