S2458
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
liver function. Keywords: liver response, hypertrophy prediction
cancer. Material/Methods: We conducted a retrospective analysis of 207 patients treated with external beam radiotherapy for hepatocellular carcinoma (HCC), cholangiocarcinoma (CC), or liver metastases (Met). Inclusion required a single course of photon-based liver-directed radiotherapy and a contrast-enhanced CT scan ≥ 3 months post-treatment. Planning and follow-up CTs were imported into a treatment planning system, and liver segments were delineated using AI-assisted segmentation with radiologist review. Functional liver volumes were calculated by excluding tumors and lesions, and segments were categorized as left (2+3), central (1+4), or right (5–8). Dose mapping was performed using deformable image registration and converted to EQD2 ( α / β =3). Segmental response was classified as hypertrophy, atrophy, or stable based on empirically derived uncertainty thresholds. Statistical analyses included Wilcoxon tests and ROC analysis to identify dose-response relationships and thresholds. Predictive models were trained using nine different machine learning classifiers with nested cross- validation, incorporating dose-volume histograms and clinical features. Model performance was evaluated using AUROC, accuracy, sensitivity, and specificity on withheld test sets. Predictive models were then used to evaluate patient-specific selection of various treatment planning techniques. Results: Among 561 liver segments analyzed from 207 patients, 50% exhibited atrophy, 28% hypertrophy, and 22% remained stable. Segments with hypertrophy received significantly lower median doses (9Gy) compared to atrophic segments (22Gy, p < 0.05). ROC analysis identified 15 Gy as the optimal threshold for predicting hypertrophy (AUC = 0.71), with 44% of segments below this dose showing substantial hypertrophy. Dose- response varied by tumor type: CC segments received the highest doses (25Gy), followed by HCC (20Gy) and Mets (19Gy), with CC significantly different from the others. Hypertrophy was most frequent in left liver segments (40%) versus central (16%) and right (29%), reflecting lower median doses (10Gy vs. 26Gy and 19Gy). At low doses (0–15Gy), left segments had the highest regeneration rate (53%), while atrophy predominated in right segments (44%). Hill climb ensemble predictive models achieved AUROC scores of 0.80 (left) and 0.77 (right), enabling treatment personalization between IMRT and 3DCRT. Conclusion: This study establishes a clinically actionable dose threshold for liver regeneration, demonstrates spatial and histologic variability in radiation response, and introduces predictive modeling as a powerful tool for personalized treatment planning. Together, these advances enable optimization strategies to preserve
Digital Poster Highlight 3042
Assessing Transferability of Photon-derived NTCP Models in Predicting Radiation-Induced Optic Neuropathy After Proton Therapy juliette thariat 1 , jean-claude quintyn 2 , cyril moignier 1 , Ibrahim Chamseddine 3 , Harald Paganetti 3 , Thao- Nguyen Pham 1 1 radiotherapy, centre francois baclesse, caen, France. 2 ophthalmology, univeristy hospital, caen, France. 3 radiotherapy, MGH, Boston, USA Purpose/Objective: Radiation-induced optic neuropathy (RION) is a rare late toxicity after radiotherapy for cranial/head and neck tumors. As proton therapy becomes more widespread, assessing the validity of photon-derived normal tissue complication probability (NTCP) models for RION prediction is crucial. We aimed to externally validate photon normal tissue complication probability (NTCP) models to predict RION after proton therapy. Material/Methods: Photon-derived NTCP models (sigmoid, logistic and EUD-based) externally validated in a prospective cohort (PIOTox-INT 2019-2023) of patients undergoing scanned proton therapy ( ≥ 2-year follow-up and mild- or-no optic neuropathy at baseline from tumor compression, surgery or ocular comorbidities). Clinically-detected RION (6.7%) and visual-field-defined grade ≥ 2 RION (24%) as alternative endpoint for more sensitive detection of RION (grade 2+, more specific than visual acuity, less imbalanced output) served as endpoints. Photon-derived NTCP models were tested with fixed or re-estimated parameters. Proton-specific penalized regressions (logistic, probit, Poisson, log- logistic) integrating dosimetric + clinical variables were developed and cross-validated. In parallel, Bayesian network (BN) modeling was applied to uncover conditional dependencies among clinical, dosimetric, and biological variables, identifying the RION Markov blanket as compact predictor set. Impact of linear energy transfer (LET) on occurrence of RION was also conducted between patients with RION or without (case-control sub-study). Model performance was assessed using AUC, sensitivity, and specificity, as well as the Akaike (AIC) and Bayesian (BIC) information criteria for the Bayesian network (BN). Results: 6.7% of patients (N=105; 179 eyes) had severe RION and 24% grade 2+ RION by VF. Photon-models without or with parameter re-estimation yielded AUC 0.692– 0.719 or 0.677–0.706, respectively. Clinical-dosimetric proton models achieved AUC 0.865. Own clinical-
Made with FlippingBook - Share PDF online