An interactive virtual dosimetrist for adjusting radiotherapy dose quality.

Background and Purpose: Effective education in radiotherapy plan quality review requires exposure to treatment plans of varying quality and opportunities for plan improvements. However, the current clinic-based paradigm does not fully support these needs. Trainees often review limited numbers of diverse treatment plans and are not fully exposed to the planning decisions essential for developing strong evaluation skills. Materials and Methods: To address these challenges, we developed "Virtual Dosimetrist" models that generate and modify radiotherapy dose distributions through language conditioning, similar to communicating with a planner during plan review. This enables creation of diverse suboptimal examples and real-time, language-guided refinement. Results: The models modify dose distributions following language directives, allowing the generation of lower-quality examples and subsequent improvement. Dose predictions were produced quickly and without reliance on a planning system, enabling interactive adjustment that mirrors the clinical review-edit cycle. For the 11 patients in the test set, good control over the extent and magnitude of dose changes across 17 OOIs, together with robustness to variations in the language directives, was observed. When comparing to dose originating from a TPS, median RMSE was 52 cGy across all dose-level comparisons and 98% of predictions met the prespecified equivalence criterion of RMSE Conclusions: This work is the first to combine deep learning-based dose prediction with natural language processing specifically designed to support the language-conditioned generation of suboptimal dose distributions and their improvement. The framework could be used as a foundation for educational applications, such as replicating plan quality review in an interactive environment.