ESTRO 2025 - Abstract Book

S3843

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Conclusion: Although ML and DL survival models performed similarly, DL models showed better generalization capabilities and could extract relevant information directly from PET/CT images even without manual segmentations and without the added steps required for radiomics feature extraction. Important features and explainability heatmaps aligned with clinical expectations.

Keywords: survival analysis, radiomics, deep learning

References: 1. S. Pölsterl. 2020. scikit-survival: A Library for Time-to-Event Analysis Built on Top of scikit-learn, Journal of Machine Learning Research, 21:1–6. 2. M Tan, QV Le. 2020. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv:1905.11946v5 3. MF Gensheimer, B Narasimhan. 2019. A scalable discrete-time survival model for neural networks. PeerJ, 7:e6257 4. S Hooker, D Erhan, PJ Kindermans, B Kim. 2019. A benchmark for interpretability methods in deep learning. arXiv:1806.10758v3