S2417
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
(CV) with systematic multicollinearity management. Material/Methods: We retrospectively analyzed 173 patient data sets with single brain metastases treated with single-fraction SRS between 2013 and 2024 at two centers (median [range] prescription dose: 20 [15-24] Gy to the 80% isodose line; median follow-up: 15 months). A total of 338 predictors were collected, comprising dosimetric, clinical and technical variables, such as DVH metrics, patient age and treatment planning system. The methodology follows a three-phase approach (Figure 1): Phase 0: Systematic multicollinearity elimination via pairwise Spearman correlation analysis (|r|>0.8), retaining the highest-AUC feature per group and high- performing singletons (AUC>0.52). Phase 1: Nested CV using a 5-fold stratified outer loop for unbiased evaluation and 10-fold stratified inner loop for model comparison and LASSO hyperparameter tuning. In each outer fold: (a) Z-score standardization; (b) Bootstrap-aggregated LASSO feature selection (1000 iterations, events-per-variable (EPV) ratio ≥ 10); (c) 10- fold inner CV model comparison (logistic regression) and selection via statistical criteria (elbow: Δ AUC>0.02, DeLong test, Akaike information criterion (AIC)), followed by final LASSO fit (CV-optimized hyperparameter) and held-out test evaluation. Phase 2: Final model fitting on entire cohort using parameters with ≥ 50% selection frequency, with bootstrap coefficient confidence intervals (CI) (1000 iterations).
Integrating radiomics and clinical variables with foundation model–derived representations from pre- treatment T2-weighted MRI improves recurrence prediction in ASCC. The foundation model approach performed comparably to established ML pipelines, suggesting that DL can capture clinically relevant imaging features. These findings highlight the potential of radiomics and foundation models for robust, multi-center recurrence prediction and individualised therapeutic planning in anal cancer management. References: [1] Andrea Vanzulli et al. “Radiomics to Predict Tumor Response to Combination Chemora- diotherapy in Squamous Cell Carcinoma of the Anal Canal: A Preliminary Investigation”. European Radiology Experimental 9.1 (Mar. 22, 2025), p. 35. issn: 2509- 9280. doi: 10.1186/s41747-025-00559-0.[2] Shanshan Tang et al. “Recurrence-Free Survival Prediction for Anal Squamous Cell Carcinoma after Chemoradiotherapy Using Planning CT-based Radiomics Model”. British Journal of Radiology 98.1166 (Nov. 2024), pp. 296–304. issn: 0007-1285. doi: 10.1093/bjr/tqae235.[3] Suraj Pai and Ibrahim Hadzic. AIM-Harvard/foundation-cancer-image-biomarker: v0.0.1. Version v0.0.1. Jan. 2024. doi: 10.5281/zenodo.10535536. Keywords: anal cancer, machine learning, radiomics Poster Discussion 777 Nested cross-validation with systematic multicollinearity management for predictive modeling of radionecrosis in stereotactic radiosurgery Karen Manger 1 , Maximilian Niyazi 2,3 , Raphael Bodensohn 2,3 , Daniela Thorwarth 1 1 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 2 Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 3 Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany Purpose/Objective: Radionecrosis (RN) remains a major complication following stereotactic radiosurgery (SRS) for brain metastases, often causing severe side effects. However, predicting RN remains challenging due to inconclusive predictive factors and the lack of universally accepted guidelines. While dosimetric studies have been conducted, robust and generalizable predictive models with stable feature selection and validated performance are still lacking. This study aims to develop and validate a prediction model for RN after SRS, using nested cross-validation
Results: Overall, 40 patients (23.1%) developed symptomatic grade 2 RN following SRS after a median time of 11 months. Multicollinearity management (93% reduction) mitigated non-informative correlations from redundant metrics. Nested CV consistently identified TB D45% [Gy] (dose delivered to 45% of tumor-brain (TB) volume; TB: healthy brain tissue plus GTV) as the sole robust predictor (100% selection frequency). Bootstrap-identified secondary candidates (including gradient index, margin size) failed to improve models significantly. TB D45% model performance on unseen data achieved an AUC of 0.70±0.11 (95% CI: 0.59-0.80) (Figure 2).
Made with FlippingBook - Share PDF online