ESTRO 2025 - Abstract Book

S3841

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Conclusion: An AI-assisted qCT analysis identified structural changes following SF SABR to peripheral lung tumors, including airway stenosis, vascular pruning/dilation and parenchymal alterations. This novel approach may allow for changes associated with different SABR schedules to be correlated with patient symptoms and long-term pulmonary function.

Keywords: Quantitative CT, Pulmonary vasculature, Lung SABR

References: [1]

Tekatli H, Bohoudi O, Hardcastle N, Palacios MA, Schneiders FL, Bruynzeel AME, et al. Artificial intelligence assisted quantitative CT analysis of airway changes following SABR for central lung tumors. Radiother Oncol 2024;198. https://doi.org/10.1016/j.radonc.2024.110376. [2] Carmo DS, Ribeiro J, Comellas AP, Reinhardt JM, Gerard SE, Rittner L, et al. MEDPSeg: End-to-end segmentation of pulmonary structures and lesions in computed tomography 2023.

3728

Poster Discussion Machine and deep learning time-to-event models with explainability for head and neck cancer outcome predictions Bao Ngoc Huynh 1 , Min Jeong Cheon 1 , Torjus Strandenes Moen 1 , Aurora Rosvoll Groendahl 2 , Ingerid Skjei Knudtsen 3 , Kristian Hovde Liland 1 , Frank Hoebers 4 , Wouter van Elmpt 4 , Einar Dale 5 , Eirik Malinen 6,7 , Oliver Tomic 1 , Cecilia Marie Futsaether 1 1 Faculty of Science and Technology, Norwegian University of Life Sciences, Ås, Norway. 2 Section of Oncology, Vestre Viken Hospital Trust, Drammen, Norway. 3 Dept. of Circulation and Medical Imaging, Norwegian University of Science and Technology, Trondheim, Norway. 4 Dept. of Radiation Oncology (MAASTRO), GROW School for Oncology and Reproduction, Maastricht, Netherlands. 5 Dept. of Oncology, Oslo University Hospital, Oslo, Norway. 6 Dept. of Radiation Biology, Oslo University Hospital, Oslo, Norway. 7 Dept. of Physics, University of Oslo, Oslo, Norway Purpose/Objective: Time-to-event modelling in radiotherapy may provide more detailed understanding of timing and risk of events, allowing for better predictions of treatment outcomes. We compared time-to-event predictions using machine (ML) and deep learning (DL) survival models for the clinical endpoints overall survival (OS) and disease-free survival (DFS) for patients with head and neck squamous cell carcinoma (HNSCC) receiving radiotherapy. Material/Methods: HNSCC patients with pre-treatment FDG PET/CT from two clinics were included. Patients from Oslo University Hospital (OUS, n=139) were used for model training and internal testing. Maastro Clinic Maastricht (n=99) patients were used for external testing to evaluate cross-institutional model generalizability. The Cox Proportional Hazards model (CoxPH) [1] and the ML models Random Survival Forest (RSF) [1] and Component-wise Gradient Boosting (CGB) [1] were trained using tabular data consisting of clinical data and radiomics features extracted from the gross primary tumor volume (GTVp) within the PET/CT images. DL time-to-event models used a 3D EfficientNet B1 architecture [2], with a negative log-likelihood loss function [3] to handle discrete time intervals and censoring. DL models were trained directly on CT images adjusted with a narrow soft-tissue window and/or PET images, with or without inclusion of GTV segmentations (GTVp and/or nodal volume GTVn) to assess modality importance for survival prediction. Variance of Gradients (VarGrad) explainability [4] was used to generate heatmaps illustrating important image regions for DL model predictions. Model performances were evaluated using Harrel’s Concordance Index (C-index).