Towards LLM-Based Generalized Automated Treatment Planning for Locally Advanced Rectal Cancer
Abstract
Purpose
The escalating global burden of cancer necessitates reliable and broadly generalizable automated treatment planning (ATP) systems. Current data-driven ATP approaches often exhibit insufficient generalization when applied to diverse clinical prescriptions. To address this limitation, we developed a novel large language model (LLM)-driven ATP framework for locally advanced rectal cancer. Our approach aims to enhance clinical adaptability through explicit, stepwise reasoning and achieve prescription-agnostic planning.
Methods
Departing from data-intensive paradigms, our framework utilizes few-shot transfer learning to build a prescription-agnostic ATP system. LLM integration enables three core capabilities: (1) distilling general planning knowledge from a minimal set of example plans; (2) generating plan initializations adaptable to varying prescriptions via an LLM-based adapter module; and (3) providing explicit, step-by-step rationales for inverse planning decisions. To evaluate the generalizability of the framework, we tested it on a cohort of 45 LARC patients exhibiting substantial variation in target volume definitions and clinical prescriptions. An independent, blinded expert evaluation was conducted on this validation set.
Results
The proposed ATP framework successfully generated clinically acceptable plans for all 45 patients, with a median of 6 optimization iterations. The LLM-based knowledge extractor inferred robust planning strategies from only three cases. Across diverse prescriptions, the ATP consistently produced dose distributions that were comparable or superior to manual plans. Notably, the LLM provided a transparent, step-by-step reasoning trajectory that justified each planning decision, substantially enhancing explainability. Blinded expert assessment rated the automated plans as clinically superior or equivalent to manual plans in 86.7%–88.9% of cases.
Conclusion
Our LLM-driven ATP framework provides a generalizable solution for generating high-quality treatment plans. It establishes a foundation for reducing clinical workload, improving planning efficiency, and advancing the development of generalized, explainable ATP systems.