S2445
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Conclusion: Temporal lobe dose differences did not alter robust feature selection, confirming the association between high-dose and TRN risk. However, at equivalent model performance, delivered-dose models placed greater weight on V75Gy, altering absolute risk calibration and raising concerns regarding generalisability. Standardised dose computation and careful model recalibration are essential when applying NTCP models across cohorts and clinical settings. References: [1] C. Gillmann, A. J. Lomax, D. C. Weber, O. Jäkel, and C. P. Karger, “Dose–response curves for MRI-detected radiation-induced temporal lobe reactions in patients after proton and carbon ion therapy: Does the same RBE-weighted dose lead to the same biological effect?,” Radiotherapy and Oncology, doi: 10.1016/j.radonc.2018.01.018.[2] C. Schroeder et al., “NTCP Modeling for High-Grade Temporal Radionecroses in a Large Cohort of Patients Receiving Pencil Beam Scanning Proton Therapy for Skull Base and Head and Neck Tumors,” International Journal of Radiation Oncology, Biology, Physics, doi: 10.1016/j.[3] T. Dickscheid et al., siibra-python. (Jun. 04, 2025). Zenodo. doi: 10.5281/zenodo.15591482. Keywords: Proton therapy, NTCP, Radionecrosis Validation of an NTCP model of moderate-severe late side effects after whole breast radiotherapy, including the impact of advanced delivery techniques Monica Maria Vincenzi 1 , Laura Giannini 2 , Alessandro Cicchetti 3 , Roberta Castriconi 1 , Miriam Torrisi 2 , Paola Mangili 1 , Anna Chiara 2 , Chiara Lucrezia Deantoni 2 , Marcella Pasetti 2 , Roberta Tummineri 2 , Nadia Gisella Di Muzio 2 , Antonella del Vecchio 2 , Andrei Fodor 2 , Claudio Fiorino 1 1 Medical Physics Department, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Radiotherapy department, IRCCS San Raffaele Scientific Institute, Milan, Italy. 3 Data Science Unit, Fondazione IRCCS Istituto Nazionale dei Tumori di Milano, Milan, Italy Purpose/Objective: Skin and breast toxicities after whole breast irradiation (WBI) can significantly impact quality of life, particularly in patients with large breasts. The aim of this work is to validate a multivariable Normal Tissue Complication Probability (NTCP) model previously developed[1] to predict moderate-to-severe late side effects in patients receiving WBI. Material/Methods: Mini-Oral 2731 Current analysis dealt with data of 683 patients treated from 2017 to 2021 with WBI (40Gy in
differences on TRNs’ NTCP predictions. Material/Methods:
Accumulated delivered treatment doses (Dd) were reconstructed from fraction-specific machine log files and compared with planned dose (Dp) for 79 patients, including 19 (24%) with grade ≥ 2 TRN (CTCAE v5). Dose agreement was evaluated using local gamma analysis (3%, 3mm). Dose-volume histogram (DVH) metrics (VxGy, V%, DxCCand Dx%) were extracted from 158 atlas-segmented temporal lobes [3].Clinical and DVH features were screened for multicollinearity, standardised, and analysed using 5-fold cross- validationwith patient-level splitting. Stable features were identified from 2000 bootstrap iterations of LASSO regression based on their selection frequency and weight sign consistency. The best-performing models across folds were chosen by unbiased area- under-curve AUC, Brier score, and Akaike Information Criterion.
Results: Despite high plan-delivered dose agreement (median γ =99.8%; IQR: 99.5-100.0%; range, 90.9– 100.0%), median maximum TL point dose differences were 2.5 GyRBE (IQR: 2.0-3.0 GyRBE, range: 0.72-35.5 GyRBE). This resulted in a consistent set of selected features for either input, but with a shift in the dose- related feature weight. V75Gy showed the largest change in model weighting between plan- and delivered-dose fits (+40%), whereas the number of surgeries prior to treatment and age were minimally affected. Nevertheless, model performance remained comparable, with similar discrimination (AUCs 0.885 for plan vs 0.910 for delivered dose) and calibration (Brier scores: 0.117 vs 0.116).
Made with FlippingBook - Share PDF online