ESTRO 2025 - Abstract Book

S3770

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Conclusion: The integration of the XGBoost model with the SHAP strategy was able to assess lung lesions CR following SBRT, with the potential to assist clinicians in directing personalized SBRT strategies in an understandable manner.

Keywords: Radiomics, lung, SBRT

2047

Digital Poster Planning CT radiomic features for predicting loco-regional recurrence in head-and-neck cancer Ceilidh Welsh 1 , Rohan Mudkavi 2 , Tom Sheehan 2 , Sejal Karmarkar 2 , Annabel Follows 2 , Amy Bates 3 , Aviva Grisby 3 , Heather Keenan 4 , Marcos Martinez Del Pero 5 , Richard J Benson 3 , William Ince 3 , Raj Jena 3 , Gillian C. Barnett 3 1 Department of Oncology, University of Cambridge, Cambridge, United Kingdom. 2 School of Clinical Medicine, University of Cambridge, Cambridge, United Kingdom. 3 Department of Oncology, Cambridge University Hospitals NHS Trust, Cambridge, United Kingdom. 4 Department of Oncology, North West Anglia Foundation Trust, Peterborough, United Kingdom. 5 Department of Oncology, West Suffolk NHS Foundation Trust, Suffolk, United Kingdom Purpose/Objective: Locoregional recurrence (LRR) is a predominant relapse pattern in head-and-neck cancer (HNC) patients receiving radiotherapy. This study aimed to develop machine learning (ML)-based radiomics models to predict LRR across HNC subsites using radiotherapy planning CT scans. Material/Methods: Data from two Cambridge University Hospital datasets, the VoxTox study [1] (N=271) and a clinically audited cohort (N=118), were collected and combined (N=389). Clinical covariates were extracted from electronic health records, and univariate analysis was performed to investigate associations with recurrence outcomes. Patients with complete imaging and recurrence data, planning CT scans, and clinician-delineated primary and nodal CTVs were included. Planning CT scans underwent pre-processing involving normalisation, scaling, and discretisation. CTV volumes were resegmented to remove artifacts and bone or air HU intensity values. Radiomic features were extracted using PyRadiomics [2], with control samples (voxel shuffling and random intensity generation) to check features were not confounded by tumour volume. The dataset was divided into patients who had undergone primary surgery (N=123) and those who had not (N=204). Two logistic regression models were built using radiomic features from planning CT scans, with additional augmented CT images for recurrence cases to address class imbalance. Feature selection was performed using ANOVA F-tests (p < 0.05), and highly correlated features (r > 0.7) were removed to minimise redundancy and multicollinearity. Logistic regression models were trained using five-fold cross-validation (80:20 split), with performance evaluated through accuracy, ROC-AUC, and confusion matrices. Results: The primary surgery cohort model achieved 84% accuracy and a ROC-AUC of 0.89, and confusion matrix results: 112 true negatives, 22 false negatives for non-recurrence and 96 true positives, 19 false positives for recurrence. For the non-primary surgery cohort, the model achieved 79% accuracy and a ROC-AUC of 0.86, with 207 true negatives, 46 false negatives for non-recurrence and 142 true positives, 26 false positives for recurrence. For each model, a radiomic signature comprised of the top ten radiomic features was used for prediction. Top predictive features were derived from voxel intensity matrices, highlighting the importance of textural and zone-based patterns in predicting LRR. Negative controls showed reduced performance, indicating model features were not volume dominated. Furthermore, the addition of clinical covariates did not improve performance in either cohort