ESTRO 2025 - Abstract Book

S3387

Physics - Machine learning models and clinical applications

ESTRO 2025

1930

Digital Poster Estimating uncertainty for AI-based dose modelling in MRI-guided RT using ensemble networks, mean variance estimation and Monte Carlo dropout Tabea Eberhardt 1 , Moritz Schneider 1 , Christian F. Baumgartner 2,3 , Paul Fischer 2,3 , Maximilian Niyazi 4,5 , Daniela Thorwarth 1,2,5 1 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 2 Cluster of Excellence "Machine Learning", Eberhard Karls University Tübingen, Tübingen, Germany. 3 Faculty of Health Sciences and Medicine, University of Lucerne, Lucerne, Switzerland. 4 Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 5 German Cancer Consortium (DKTK), partner site Tübingen, and German Cancer Research Center (DKFZ), Heidelberg, Germany Purpose/Objective: Monte Carlo simulations provide accurate dose calculations in radiotherapy but are computationally demanding, limiting their use in real-time applications. Neural networks promise both, accuracy and speed in dose modelling, though some faulty cases were reported. Identifying these challenging situations upfront is crucial. This work aims to implement uncertainty estimation to neural network-based dose modelling using deep ensemble networks (DE), mean variance estimation (MVE), and Monte Carlo dropout (MCD) and explore its relationship with observed errors. Material/Methods: All 3D U-Nets in this study were trained on 3,956 segments from 75 clinical 1.5T MR-linac treatment plans to model radiation dose and tested on 646 segments from 10 prostate cancer treatment plans. The DE consisted of 10 identical 3D U-Nets initialized from different seeds. The dose output was the arithmetic mean of the individual model outputs, uncertainty was extracted from their standard deviation. Four models using MCD were trained with either standard or 3D-spatial dropout [1] applied to all (lv1-4) or all but the first level (lv2-4) of the 3D U-Net with a dropout rate of 0.1. Ten predictions were made with every model, the arithmetic mean was used as the dose output, the standard deviation as the uncertainty estimate. For MVE, a second output modelling the uncertainty was added to the 3D U-Net [2]. Three variants were trained, starting from one of the DE models: one with fixed weights in all but the uncertainty output branch (MVE1), one with fixed weights in the split-off dose output branch (MVE2), and one without fixed weights (MVE3). The quality of uncertainty estimation was assessed using reliability diagrams, expected normalized calibration error (ENCE), and confidence curves. Results: Except for MCD approaches with dropout in all layers, the uncertainty estimations overall matched the prediction error (figure1).