Main Session
Sep
30
SS 49 - Smart Planning and Adaptation
369 - A Dual-Agent Multimodal Large Language Model Architecture for Fully Automated IMRT Planning in Small Cell Lung Cancer
Presenter(s)
Shuoyang Wei, PhD - Mayo Clinic Arizona, Phonix, AZ
S. Wei1, S. Yan1, Y. Liang2, J. Yang1, X. Meng1, W. Li1, B. Yang1, J. Qiu1, and W. Liu3; 1Department of Radiotherapy, Peking Union Medical College Hospital, Beijing, China, 2Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China, 3Department of Radiation Oncology, Mayo Clinic, Phoenix, AZ
Purpose/Objective(s):
Intensity-Modulated Radiation Therapy (IMRT) planning for lung cancer remains difficult to fully automate due to the combined need for beam geometry design and iterative inverse optimization. We propose a reasoning-driven dual-agent multimodal large language model (MLLM) architecture that reproduces the physicist decision workflow and enables automated IMRT planning for small cell lung cancer (SCLC).Materials/Methods:
Forty SCLC patients treated with IMRT (45 Gy in 30 fractions, BID) were retrospectively included (training n=10, testing n=30). Two independent multimodal agents based on Qwen3.5-397B-A17B were constructed. For beam geometry design, a rule-guided decision policy describing conditional template selection principles was provided. Rather than directly executing fixed rules, the beam agent analyzed CT images with target and Organs at Risk (OAR) contours, evaluated spatial proximity and volumetric relationships, and selected beam templates through anatomy-aware reasoning. The optimization agent operated as a closed-loop controller: after each optimization cycle, it ingested Dose-Volume Histogram (DVH) statistics and constraint deviations, analyzed target-OAR trade-offs, and reformulated objective functions until convergence. The separation of geometric reasoning and optimization control enabled independent yet coordinated decision modules. In-context learning was applied using the training cohort. MLLM-generated plans were compared with expert manual plans using DVH metrics, conformity index (CI), homogeneity index (HI), monitor units (MU), and delivery verification via gamma analysis (2 mm/2%).Results:
The proposed framework successfully completed fully automated IMRT planning for all testing cases. Beam configurations were identical to expert plans in 40% of patients and clinically comparable in the remainder. Plan optimization converged in a mean of 4 iterations with total runtime under 8 minutes per case. Target coverage, CI, HI, and MU were statistically comparable to manual plans. PTV Dmax was modestly reduced (p<0.05), and bilateral lung V5 decreased by 1.1% (p<0.05), indicating improved hot spot control and low-dose lung sparing. Spinal cord D0.1cc increased slightly but remained within tolerance. All plans passed delivery verification with a ? passing rate >99.5%.Conclusion:
A reasoning-driven dual-agent MLLM architecture can integrate beam geometry selection and inverse optimization within a unified modular framework, achieving fully automated IMRT planning with dosimetric quality comparable to, and in selected metrics improved over, expert planning. By decoupling geometry reasoning from inverse control, this approach establishes a foundation-model-driven paradigm for intelligent radiotherapy planning that may generalize to other geometry-dependent treatment sites beyond lung IMRT.