ESTRO 2024 - Abstract Book

S3034

Physics - Autosegmentation

ESTRO 2024

1. Kong FM (Spring), Ritter T, Quint DJ, et al. Consideration of Dose Limits for Organs at Risk of Thoracic Radiotherapy: Atlas for Lung, Proximal Bronchial Tree, Esophagus, Spinal Cord, Ribs, and Brachial Plexus. Int J Radiat Oncol. 2011;81(5):1442-1457. doi:10.1016/j.ijrobp.2010.07.1977 2. Wang W, Matuszak MM, Hu C, et al. Central Airway Toxicity After High Dose Radiation: A Combined Analysis of Prospective Clinical Trials for Non-Small Cell Lung Cancer. Int J Radiat Oncol. 2020;108(3):587-596. doi:10.1016/j.ijrobp.2020.05.026

966

Digital Poster

Uncertainty in deep learning-based automatic segmentations of head and neck cancer tumors

Bao Ngoc Huynh 1 , Aurora Rosvoll Groendahl 2 , Oliver Tomic 1 , Ingerid Skjei Knudtsen 3 , Frank Hoebers 4 , Wouter van Elmpt 4 , Einar Dale 5 , Eirik Malinen 6,7 , Cecilia Marie Futsaether 1 1 Norwegian University of Life Sciences, Faculty of Science and Technology, Ås, Norway. 2 Vestre Viken Hospital Trust, Section of Oncology, Drammen, Norway. 3 Norwegian University of Science and Technology, Department of Circulation and Medical Imaging, Trondheim, Norway. 4 GROW School for Oncology and Reproduction, Department of Radiation Oncology (MAASTRO), Maastricht, Netherlands. 5 Oslo University Hospital, Department of Oncology, Oslo, Norway. 6 Oslo University Hospital, Department of Medical Physics, Oslo, Norway. 7 University of Oslo, Department of Physics, Oslo, Norway

Purpose/Objective:

Although deep learning (DL)-based auto-segmentation approaches can assist human experts with potentially faster and more consistent gross tumor volumes (GTV) delineations, there are cases where these models struggle to contour the GTV correctly. This study explores how quantification and visualization of the uncertainty of DL automatic contours can be utilized as a quality control step for auto-segmentation of head and neck cancer (HNC) GTVs.

Material/Methods:

To estimate and visualize DL model uncertainty, a suitable DL model for HNC GTV auto-segmentation must first be developed. Therefore, 3D U-net models (Ronneberger et al., 2015; Milletari et al., 2016) were optimized using 90 different combinations of hyper-parameters such as architecture complexity levels, types of image preprocessing and augmentation methods. Input data were baseline 18F-FDG PET/CT scans from HNC patients. Manual GTV contours were used as the ground truth. The Dice similarity coefficient (DSC) was used to quantify the auto-segmentation performance. Patients from two different centers (N=197 and N=113) were included, in which all patients from the second center served as the external test set. All models were trained and validated with 157 patients from the first center. Then the model with the highest validation DSC was evaluated on 40 patients in the test set of the first center (internal test set) and 113 patients in the external test set from the second center. This final model was then subjected to uncertainty analyses.