ESTRO 2025 - Abstract Book

S4008

Radiobiology - Tumour radiobiology

ESTRO 2025

and representativeness of texture features through the fuzzy statistical approach to characterize a comprehensive image phenotype profile for better prognosis prediction.

Material/Methods: A cohort of 1,321 patients with oropharyngeal cancer from the public RADCURE dataset was employed to develop a survival model. In this study, the conventional region of interest (ROI) was substituted with a fuzzy region [2] to mitigate the impact of ROI boundary uncertainty. Subsequently, the statistical values of matrices were computed using the principle of fuzzy inference system (FIS) [3] to enhance the representational capacity of the texture features (Figure 1). Ablation comparative experiments were conducted to validate the fuzzy statistical approach in terms of texture feature reliability and survival model performance via the ICC and C-index metrics. ICC values were computed on the perturbed dataset under 10 contour randomizations. The comparative results of C-index were obtained from models evaluated using five-fold cross-validation with 20 repetitions.

Results: While the number of features with ICC values exceeding 0.6 remained consistent at 102±1 across the four methods, the ICC distribution of extracted texture features indicated that employing fuzzy statistics resulted in a greater number of features with higher ICC values (illustrated in Figure 2A). Notably, the proposed fuzzy statistical approach exhibited significant advantages with 43 features achieving ICC values within the (0.85, 0.95] interval, compared to 25, 34, and 33 features in other methods. Regarding survival model performance, although the individual