S2295
Physics - Machine learning and AI algorithms
ESTRO 2026
assessed via 1,000-sample bootstrap resampling on the internal holdout test set. Performance metrics included C-index, integrated Brier score (IBS), Kaplan- Meier stratification, modality ablation, SHAP feature attribution, and subgroup analyses. Architectural exploration compared DenseNet variants and ResNet backbones (18-101 layers). Results: The best ensemble model achieved a C-index of 0.735 (95% CI: 0.605-0.802) and an IBS of 0.168 on the internal holdout test. Multimodal fusion significantly outperformed clinical-only (0.675), radiomics-only (0.612), and imaging-only (0.562) models (Fig.1A). SHAP analysis identified Tobacco consumption and GTVn volume as top clinical predictors (Fig.1B). Risk stratification of HPV+ patients showed significant separation (log-rank p=0.0037 on test) (Fig.1D). Deeper PET/CT image encoders decreased internal holdout test performance (ResNet-18: 0.728; ResNet-101: 0.488) due to overfitting, with a hyperparameter-tuned DenseNet backbone proving optimal for this dataset size (Fig.2).
10.1016/j.ctro.2021.08.010[2] Nagtegaal, S. et al. Neuro-Oncology Advances 2 (Jan. 2020). doi: 10.1093/noajnl/vdaa060[3] Nagtegaal, S. et al. Clinical and Translational Radiation Oncology 26, 35–41 (Jan. 2021). doi: 10.1016/j.ctro.2020.11.005[4] Hovden, I. et al. Oct. 2023. doi: 10.25493/D948-J3P.[5] Liu, X. et al.
Sept. 2022. doi: 10.48550/arXiv.2209.03003. Keywords: Image generation, Glioma, MRI
Digital Poster Highlight 2371
GaMMA-Surv: Gated Multimodal Autoencoding Joint Fusion of PET/CT Imaging and Clinical Data for Multi-Centric Head and Neck Cancer Survival Prediction Yujing Zou 1,2 , Sébastien Quetin 1,2 , Juan Duran 1,2 , Shirin Abbasinejad Enger 1,2 1 Medical Physics Unit, Department of Oncology, Faculty of Medicine, McGill University, Montreal, Canada. 2 Lady Davis Institute for Medical Research, Jewish General Hospital, Montreal, Canada Purpose/Objective: Accurate relapse-free survival (RFS) prediction in human papillomavirus-positive (HPV+) oropharyngeal squamous cell carcinoma (OPSCC) remains challenging, as recent de-escalation trials have failed to demonstrate non-inferiority. To address this, we developed GaMMA-Surv, a gated multimodal variational autoencoder (VAE) framework that jointly fuses pre-treatment PET/CT imaging and clinical data for robust, clinically interpretable OPSCC risk stratification. Material/Methods: GaMMA-Surv was trained on 542 OPSCC patients from seven cancer centres in the HECKTOR 2025 dataset [1] using 5-fold cross-validation and evaluated on an internal holdout test set (n=136, 27 events). Primary (GTVp) and nodal (GTVn) tumour volumes were automatically segmented by a trained 3D nnUNet to define regions of interest. PET/CT volumes were cropped to 176 × 144 × 176 voxels (1 mm ³ spacing with 5 mm tumour margins). A 3D DenseNet encoder extracted deep PET/CT representations [2], while a tabular transformer encoded clinical variables (demographics, stage, HPV status) and nnUNet-derived radiomics. The VAE fusion module concatenated imaging, clinical, and radiomic embeddings into a unified multimodal representation. Attention-based gating compressed the input into a shared 192- dimensional latent space for use by a DeepHit survival head. The model was trained end-to-end with time- aware contrastive learning to enhance temporal discrimination. Predictions were ensemble-averaged across the five folds, weighted by their validation concordance index (C-index). Model uncertainty was
Made with FlippingBook - Share PDF online