KEY POINTS
- Investigators retrospectively tested an LLM-based planning agent in 20 head and neck cancer patients treated with nine-field 6 MV IMRT. Prescription was 70 Gy to the boost PTV and 44 Gy to the primary PTV, using 2 Gy fractions.
- The agent interacted directly with Eclipse through ESAPI, repeatedly extracting DVH metrics and optimization losses and then modifying inverse-planning constraints. It was not fine-tuned on previous treatment plans and did not use patient-specific reference plans during generation.
- Clinical objectives covered 10 structures: two PTVs plus the bilateral parotids, oral cavity, larynx, pharynx, spinal cord with 5 mm margin, brainstem, and mandible. Physician-specified patient-specific priorities remained part of the workflow.
- Compared with clinically approved plans, LLM plans reduced maximum dose from 108.8% to 106.5% of prescription (p < 0.05).
- Boost-target conformity improved from a mean conformity index of 1.39 to 1.18 (p < 0.05). Primary-target conformity was similar at 1.88 versus 1.82 (p = 0.47).
- OAR sparing was broadly comparable and in several structures numerically better. For example, mean left-parotid D50 was approximately 22.7 Gy clinically versus 19.2 Gy with the selected LLM configuration, while pharyngeal D50 was 47.5 versus 39.9 Gy.
- Removing optimization-specific prior knowledge from the prompt materially worsened OAR sparing. The result therefore demonstrates reference-free planning rather than knowledge-free planning: substantial clinical and TPS-specific scaffolding remained necessary.
- The study used a single institution, one TPS, one beam arrangement, and only 20 retrospective cases. There was no prospective planner-time comparison, independent physician preference study, external-site validation, or clinical delivery of the generated plans.
CLINICAL TAKEAWAY
An off-the-shelf LLM can act as an iterative optimization agent inside a commercial TPS without learning from a library of prior plans. The concept is technically important, but human-defined objectives, domain knowledge, independent review, and prospective safety validation remain indispensable.