ESTRO 2025 - Abstract Book

S3842

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Results: RSF models achieved the highest C-indices on the internal test set (Table 1a) but generalized poorly to the external test set. CoxPH and CGB models had lower internal test set performances but generalized better for DFS prediction. The most important features for OS and DFS prediction were HPV status, TNM stage, GTVp sphericity, pack years and PET texture. For OS, the DL model based on PET/CT images and GTVp segmentations had the highest C-index (Table 1b). For DFS, models based solely on PET images without segmentations performed best. DL models generalized better. VarGrad explainability (Figure 1a) highlighted the GTV, even when models were not given segmentations, and found PET more important than CT (Figure 1b).