ESTRO 2025 - Abstract Book

S3393

Physics - Machine learning models and clinical applications

ESTRO 2025

Keywords: Triage, Artificial intelligence, Synthetic data

References: [1] Xie, L., Lin, K., Wang, S., Wang, F., & Zhou, J. (2018). Differentially private generative adversarial network. arXiv preprint arXiv:1802.06739. [2] Jordon, J., Yoon, J., & Van Der Schaar, M. (2018, September). PATE-GAN: Generating synthetic data with differential privacy guarantees. In International conference on learning representations.

2058

Digital Poster Robustness of predictive foundation model features in head and neck cancer

Katy L Scott 1 , Sejin Kim 1,2,3 , Jermiah J Joseph 1 , Mogtaba Alim 1,4,5 , Matthew Boccalon 1 , Mattea Welch 1,3 , Chris McIntosh 2,6,7 , Katrina Rey-McIntyre 7 , Shao Hui Huang 7,8 , Tirth Patel 6,7,3 , Tony Tadic 7,8,3 , Brian O'Sullivan 7 , Scott V Bratman 2,7,8 , Andrew J Hope 8,7 , Benjamin Haibe-Kains 1,2,3 1 Princess Margaret Research, Princess Margaret Cancer Centre, Toronto, Canada. 2 Department of Medical Biophysics, University of Toronto, Toronto, Canada. 3 Cancer Digital Intelligence, Princess Margaret Cancer Centre, Toronto, Canada. 4 Department of Cell & Systems Biology, University of Toronto, Toronto, Canada. 5 Department of Computer Science, University of Toronto, Toronto, Canada. 6 TECHNA Institute, University Health Network, Toronto, Canada. 7 Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, Canada. 8 Department of Radiation Oncology, University of Toronto, Toronto, Canada Purpose/Objective: In the past decade, image-based biomarker discovery has been dominated by radiomics, but rarely addresses common limitations: retrospective studies, limited external validation, and extremely variable processing procedure [1]. Further, independently predictive tumour volume measurements [2] are often included in these analyses. When volume is excluded, selected feature volume correlation and predictive performance of volume alone are rarely reported, raising questions about biomarker superiority over volume. Foundation models, trained on large datasets and adaptable to various tasks, represent a new frontier in feature extraction. The Foundation Model for Cancer Image Biomarkers (FMCIB) [3] was trained on 11,467 annotated CT lesions and evaluated on lesion anatomical site classification, lung nodule malignancy prediction, and lung cancer prognostication. In the present work, we evaluate FMCIB on the task of survival prediction in head and neck cancer (HNC) and explore deep learning feature robustness against volume invariant negative controls (NC). Material/Methods: Publicly available HNC datasets containing CT images, corresponding gross tumour volume (GTV) contours, and survival data were retrieved from The Cancer Imaging Archive. RADCURE [4] (n = 2994) was selected for model development. HEAD-NECK-RADIOMICS-HN1 [5] (n = 137) and HNSCC [6] (n = 441) were used for external validation. FMCIB features were extracted for the original CT and GTV, along with six NC images generated by our Radiomic Extraction and Analysis for DICOM Images to Refine Objective Quality Control pipeline (Figure 1A). Pearson’s correlation was calculated between original and NC features. Maximum relevance minimum redundancy (MRMR) feature selection and Cox Proportional Hazards models were employed in 5-fold cross-validation on RADCURE training data. Fitted models were applied to the RADCURE test and external validation data and evaluated using Harrel’s concordance index. A univariate model for tumour volume was compared. Results: Across 2,207 training, 713 testing, and 578 validation GTV contours from various HNC tumor sites, 4,096 features were extracted per image type. Analysis (Figures 1B, 1C) revealed similar correlations between original and GTV NC features which are absent in full and non-GTV NC features.