KEY POINTS
- This study developed an automated whole-brain radiotherapy planning workflow combining a deep learning hyperparameter prediction model with a large language model-based conversation module for plan refinement.
- The hyperparameter prediction model was trained on 55 whole-brain radiotherapy cases and evaluated on 15 independent cases.
- 14 of 15 deep learning-generated plans were clinically acceptable: 11 were acceptable as generated, 3 required minor edits, and 1 was rejected because of excessive hotspots and relatively high lens dose.
- After normalization to identical clinical target volume D95%, there were no statistically significant differences between automated and clinical plans for clinical target volume D1% or D99%, or mean dose to the eyes and lenses.
- The automated workflow required less than 1 minute of computation and about 7 minutes end-to-end with a single mouse click, compared with about 15 minutes for manual planning.
CLINICAL TAKEAWAY
This workflow shows that artificial intelligence-assisted whole-brain radiotherapy planning can produce clinically acceptable field-in-field plans with reduced planning time and a structured route for physician feedback. The large language model component is interesting because it translates natural-language feedback into constrained planning actions rather than acting as an autonomous planner. But this remains a small retrospective technical study in a relatively simple planning site, so it is workflow-relevant, not proof of clinical superiority.