S2479
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Digital Poster Highlight 4000 Deep Learning scaling law in radiotherapy: outcome prediction performance vs dataset size Miguel Garrett Fernandes 1 , Ying Zhang 1 , Zhuoyan Shen 1 , Dirk De Ruysscher 2 , Richard Canters 2 , Andrew Hope 3 , Barbara Stam 4 , Johan Bussink 5 , René Monshouwer 5 , Maria A Hawkins 1,6 , Charles-Antoine Collins Fekete 1 1 Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom. 2 Department of Radiation Oncology (MAASTRO), GROW School for Oncology and Reproduction, Maastricht University Medical Centre, Maastricht, Netherlands. 3 Department of Radiation Oncology, University of Toronto and Radiation Medicine Program, Princess Margaret Hospital, University Health Network, Toronto, Canada. 4 Department of Radiation Oncology, Netherlands Cancer Institute, Amsterdam, Netherlands. 5 Department of Radiation Oncology, Radboud University Medical Centre, Research Institute for Medical Innovation, Nijmegen, Netherlands. 6 Radiotherapy Physics, University College London Hospitals NHS Foundation Trust, London, United Kingdom Purpose/Objective: Deep Learning (DL) performance in computer vision improves with increased dataset size according to a power scaling law [1]. In radiotherapy, such models could predict patient outcome or identify toxicity patterns; however, datasets are typically limited to a few hundred patients, restricting DL applicability. In this study, we use a large radiotherapy dataset to empirically derive the power scaling law for a typical Convolutional Neural Network (CNN) predicting Overall Survival (OS), enabling us to estimate the dataset size required for a target level of performance. Material/Methods: Imaging, clinical, and OS data was extracted for 2203 Stage III NSCLC patients from 4 different institutions (RUMC, NKI, PMCC, MAASTRO). A residual CNN (128K parameters) was trained to predict OS using clinical features (Age, Sex, GTV, Dose Regimen) and aligned CT+Dose inputs – 160x100x35 voxels in Left-Right (LR), Dorsal-Ventral (DV), and Craniocaudal (CC) directions, respectively – cropped around the lungs with 5 mm LR/DV margins and 10 mm CC margin. Cox loss was used to leverage censored survival times. The model was trained using 10-fold cross-validation. Concordance index (Cindex) was used to assess model performance. The same training procedure was repeated for increasing dataset sizes (200/500/1000/2000), drawn at random from the full training set. Validation sets were kept equal across different size runs (n=203). A power law
rCBV (AUC = 0.53). External validation confirmed these findings. In Cohort 2, mean AUCs were 0.68 for CTH, 0.60 for OEF, and 0.55 for rCBV, while in Cohort 3, they reached 0.72, 0.72, and 0.50, respectively. Full results are summarised in Table 1.
Conclusion: Advanced perfusion biomarkers (CTH and OEF) consistently outperformed rCBV in predicting relapse- prone voxels and demonstrated stable performance after validation in two independent cohorts, supporting their robustness in a multicentre HGG
population. References:
1. Henry, T. et al. (2021). Brain Tumor Segmentation with Self-ensembled. Lecture Notes in Computer Science(), vol 12658. Springer, Cham. https://doi.org/10.1007/978-3-030-72084-1_302.Chang PD, et al. Predicting Glioblastoma Recurrence by Early Changes in the Apparent Diffusion Coefficient Value and Signal Intensity on FLAIR Images. American Journal of Roentgenology. 2017;208(1):57-65. doi:10.2214/AJR.16.162343. Sim Y, et al. Clinical, qualitative imaging biomarkers, and tumor oxygenation imaging biomarkers for differentiation of midline-located IDH wild-type glioblastomas and H3 K27-altered diffuse midline gliomas in adults. European Journal of Radiology. 2024;173:111384. doi:10.1016/j.ejrad.2024.111384 Keywords: perfusion biomarkers, HGG relapse- prediction
Made with FlippingBook - Share PDF online