ESTRO 2025 - Abstract Book

S3715

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

learning models, logistic regression models presented the highest performance. Models developed by the combination of radiomics and dosiomics (Logistic regression: AUC=0.921; CA=0.894; F1=0.891; Precision=0.891; and Recall=0.894) performed better when compared with radiomics-only models (AUC=0.897) and dosiomics-only models (AUC=0.843) (Table 1). ROC curves were plotted for all models (Figure 1).

Conclusion: Machine learning models utilizing combination of radiomic and dosiomic features from multiple OARs demonstrated promising performance in pre-treatment trismus prediction. It may pave the way in personalized RT treatment planning for NPC patients in facilitating prevention or reducing impact of trismus.

Keywords: Trismus, Nasopharyngeal carcinoma, Radiomics