S2303
Physics - Machine learning and AI algorithms
ESTRO 2026
Variational autoencoder for regression: Application to brain aging analysis. MICCAI (pp. 823-831). Keywords: hierarchical, variational inference, regression
pipeline. The first predicted the toxicity at a given timepoint, the other estimated the parameters of the survival curve. The remaining two models used an additional variational autoencoder to help the CNN learn anatomically meaningful features. All models used the ELBO (Evidence Lower BOund) loss function.
Digital Poster 3079
causal discovery of radiotherapy-induced cardiotoxicity and survival in lung cancer Bowen Jiang 1 , Miren Summers 1 , Jason Hartford 2 , Kathryn Banfill 3 , Eliana Vasquez Osorio 4 , Alan Mcwilliam 4 1 School of Medical Science, University of Manchester, Manchester, United Kingdom. 2 Department of computer Science, University of Manchester, Manchester, United Kingdom. 3 Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom. 4 Cancer Science, University of Manchester, Manchester, United Kingdom Purpose/Objective: Radical radiotherapy is used as curative treatment for approximately 40% of patients with lung cancer. Higher doses to sub-structures of the heart have been associated with poorer overall survival; however, the causal relationships between cardiac dose and overall survival remain unclear. Traditional regression analyses are not designed to disentangle complex causal relationships among clinical risk factors. Therefore, Bayesian networks (causal discovery) and directed acyclic graphs (DAGs) are emerging as powerful tools for exploring underlying causal structures. This study aims to investigate the underlying causal relationships between clinical, dosimetric, and demographic variables and overall survival. We compare two major classes of causal discovery approaches—constraint-based methods (PC algorithm) and score-based methods (DAGSLAM)— against a clinician-derived expert consensus DAG used as a reference benchmark (Summers et al., 2025). Material/Methods: A total of 1,051 lung cancer patients treated at a single academic centre between 2010 and 2013 were included. Two causal discovery approaches were applied. The first, a constraint-based approach, infers causal structures by testing conditional independence relationships between variables. The second, a score- based approach, formulates causal discovery as an optimisation problem, aiming to identify the network structure that best fits the data according to a defined scoring metric. Two evaluation metrics for DAGs were applied: precision and recall. Precision measures the proportion of predicted edges that are correct, while recall measures the proportion of true edges successfully recovered.
Results: A simple population fit to equation 1 yielded a model with an RMSE of 23.7, for randomly generated data this would be expected to be 28.9 (Figure 2). The CNN- MLP predicting individual timepoints had an RMSE=26.1. The Hierarchical-CNN-MLP has and RMSE of 0.19. The VAE regularised models, Hierarchical-VAE (Variational-Autoencoder) and Hierarchical-VAER, both had an RMSE of 0.16-0.17. Conclusion: Using a hierarchical model did perform better than both the naïve model and directly predicting the MDADI. The additional complexity of the VAE based models appeared to add a small additional benefit. VAEs however, can generate pseudo-samples matching a given toxicity profile. References: [1] Chen, A. Y., et al. (2001). The development and validation of a dysphagia-specific quality-of-life questionnaire for patients with head and neck cancer. Arch Otolaryngol Head Neck Surg, 127(7), 870-876.[2] Mehanna, H., et al. (2019). Radiotherapy plus cisplatin or cetuximab in low-risk human papillomavirus- positive oropharyngeal cancer (De-ESCALaTE HPV). Lancet, 393(10166), 51-60.[3] de Vette, S. P. M., et al. (2025). Deep learning NTCP model for late dysphagia after radiotherapy for head and neck cancer patients based on 3D dose, CT and segmentations. Radiother and oncol, 213, 111169.[4] Zhao, Q., et al. (2019).
Made with FlippingBook - Share PDF online