S2461
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
clinician-defined segmentation provided a 3D bounding box surrounding the primary tumor for each patient.As a baseline clinical model, we trained a logistic regression classifier to predict any recurrence at three years post-treatment using patient age, tumor stage, histological subtype and lymph node status as covariates. In parallel, we trained a 3D convolutional neural network (CNN) for the same prediction task using only the PET SUV within the 3D tumor bounding boxes as input (Figure 1). We then produced an ensemble clinical + PET model by combining the risk score prediction from the 3D CNN with the clinical features and training a new logistic regression classifier.
Conclusion: We developed a practical model for predicting individual risk of RN after SRS. Its high sensitivity (88%) provides a valuable screening tool for identifying high- risk patients, enabling closer monitoring, earlier symptom management, and candid discussions about treatment-related risks. While external validation is necessary, this model establishes a foundation for clinically relevant risk stratification to improve patient safety and outcomes following SRS. Keywords: radiation necrosis, prognostic model Digital Poster 3300 Deep learning with pretreatment fluorodeoxyglucose positron emission tomography for recurrence prediction in locally advanced cervical cancer. Lev Paciorkowski, Matthew Inkman, Michael Waters, Thomas Mazur, Julie Schwarz, Jin Zhang Department of Radiation Oncology, Washington University School of Medicine, St. Louis, USA Purpose/Objective: Roughly one-third of patients diagnosed with locally advanced cervical cancer (LACC) show resistance to standard-of-care concurrent chemoradiotherapy (CCRT), leading to poor outcomes1. The discovery of biomarkers from non-invasive pretreatment imaging capable of predicting CCRT failure could facilitate the identification of high-risk patients who may benefit from alternative therapeutic options, such as immunotherapy2 or neoadjuvant chemotherapy followed by surgery3. Material/Methods: We analyzed data from 273 LACC patients who underwent pretreatment full-body fluorodeoxyglucose positron emission tomography (FDG-PET). PET images were normalized to standardized uptake values (SUV) and uniformly resampled to a 4x4x4 mm3 voxel size. A
Figure 1: Architecture of 3D CNN trained on FDG-PET images.We trained both the clinical and ensemble models using 5x4 nested cross-validation for hyperparameter tuning on the same 273-patient cohort. For each model, “high risk” patients were designated as those having a predicted risk score above 0.33; patients below this threshold were considered “low risk”. We subsequently generated cross-validated Kaplan-Meier curves as described in Simon et al. (2011)4 and used the Wilcoxon test to evaluate both models. Results: Within three years post-treatment, 106 (39%) patients had persistent or recurrent disease. The clinical model achieved marginally significant separation between high and low-risk patients (Wilcoxon p = 0.059; Figure 2A). In contrast, the ensemble clinical + PET model demonstrated a strong ability to separate high and low-risk patients (Wilcoxon p = 0.00032; Figure 2B).
Made with FlippingBook - Share PDF online