ESTRO 2025 - Abstract Book

S3439

Physics - Machine learning models and clinical applications

ESTRO 2025

ROC analysis was performed to assess classification performance. The model was then evaluated on external plans delivered in two different institutions (Inst.2-3).

Results: The model MAE on the test set was 1.9%, with a mean PI width of 7.6% and coverage of 91.2%. On the external data, the model achieved a similar overall performance, with an MAE of 1.7%, mean PI width of 6.5% and coverage of 92.7%. However, in all datasets the performance drastically reduced for arcs with GPR below the action limit (GPR < 95%), where the MAE, mean PI width, and coverage were 7.3%, 12.5%, and 64% on the test set, and 8.3%, 6.9%, and 21.1% on external data, respectively (Figure 1). ROC analysis revealed an AUC of 0.83 and 0.66 on the test set and external data (Figure 2).

Conclusion: ML models can be used as support tools to reduce measurement-based PSQA workloads. Uncertainty quantification provides crucial information for implementing trustworthy virtual QA methods. Careful consideration