S2444
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
pancreatic cancer (LAPC). Material/Methods:
We retrospectively reviewed 42 patients with LAPC from two external centers (31 from Center 1, treated between 2010-2023, and 11 from Center 2 treated in 2016-2019).All patients underwent induction chemotherapy (regimen based on gemcitabine and abraxane) followed by external radiotherapy delivered with IMRT following different fraction schemes (BED10: median=60Gy, IQR=47-60). PET-based tumor volumes were segmented following a previously validated semi- automatic method and radiomic features (RFs) were extracted in compliance with IBSI guidelines. Then, RFs were harmonized using the ComBat method to match the training dataset. Four radiomic models previously trained[1] and internally validated[2] at San Raffaele Institute, Milano, were tested on the external cohort; each combined specific features and coefficients were derived from Cox proportional hazards analysis to compute prognostic indices, and used to stratify survival risk through ROC and Kaplan–Meier analyses. The models were then retrained to optimize risk discrimination. Results: According to the p-value of the model and the corresponding C-index, two models were validated (Table 1): the model incorporating Statistical- Percentile10 and Morphological-ComShift (Model1), and the one based on only Statistical-Percentile10 (Model2). Both confirmed moderate/high ability in stratifying the cohort in high/low-risk groups, as shown in Figure1 for Model 2. The inclusion of grading only marginally enhanced the performances of the models. Models’ performances remained robust after retraining (C-index~0.74-0.78). In contrast, models incorporating more complex features were not confirmed (Model 3 and 4), showing similar performances of Models 1 and 2 in the case of retraining.
Conclusion: External validation on a multicenter LAPC cohort confirmed that [18F]FDG-PET-based radiomic models based on one or two first-order RFs are reliable in predicting DRFS, showing robust generalizability across different clinical settings. These results highlight their potential clinical utility and support their integration into personalized decision-support systems. The study was granted by AIRC (IG25951). References: [1] Mori et al. Doi:10.1016/j.radonc.2020.07.003[2] Vincenzi et al. Doi:10.1016/j.radonc.2024.110700 Keywords: external validation, DRFS, radiomic models Normal tissue complication probability modeling of temporal lobe radionecrosis using delivered versus planned dose in skull base proton therapy. Bastien Golomer 1,2 , Damien Charles Weber 1,3 , Antony John Lomax 1,2 , Giovanni Fattori 1 1 Center for Proton Therapy, Paul Scherrer Institute, Villigen, Switzerland. 2 Department of Physics, ETHZ, Zürich, Switzerland. 3 Radiation Oncology Department, University Hospital Bern, Inselspital, Bern, Switzerland Poster Discussion 2710 Purpose/Objective: Temporal lobe radionecrosis (TRN) can occur after proton therapy for skull-base chordoma and chondrosarcoma treated with 74 and 70 GyRBE, respectively. This adverse event has been associated with temporal-lobe maximum dose and related dose metrics in several normal tissue complication probability (NTCP) models, making predictions sensitive to accurate characterization of local high- dose regions [1], [2]. As delivered dose may deviate from planned dose however, NTCP models derived from planned dose only may not accurately reflect real-world treatment conditions. This study quantifies the impact of planned-versus-delivered dose
Made with FlippingBook - Share PDF online