ESTRO 2025 - Abstract Book

S2479

Physics - Autosegmentation

ESTRO 2025

References: 1. Bates JE, Howell RM, Liu ; Qi, Yasui Y, Mulrooney DA, Sughosh Dhakal ;, et al. Therapy-Related Cardiac Risk in Childhood Cancer Survivors: An Analysis of the Childhood Cancer Survivor Study [Internet]. Vol. 37, J Clin Oncol. 2019. Available from: https://doi.org/10. 2. Feng M, Moran JM, Koelling T, Chughtai A, Chan JL, Freedman L, et al. Development and validation of a heart atlas to study cardiac exposure to radiation following treatment for breast cancer. Int J Radiat Oncol Biol Phys. 2011 Jan 1;79(1):10–8.

2555

Digital Poster Automated Tumor Segmentation Within Atelectasis Regions in CT Images Using nnUNet: Enhancing Diagnostic Accuracy and Radiotherapy Planning Minghua Li 1 , Qing Gu 2 , Jianfeng Ji 2 , Leonard Wee 1 , Andre Dekker 1 , Zhen Zhang 1,2 1 Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands. 2 Zhejiang Cancer Hospital, Institute of Basic Medicine and Cancer (IBMC), Chinese Academy of Sciences, Hangzhou, China Purpose/Objective: Accurate segmentation of atelectasis and tumors in CT images is critical for improving diagnostic accuracy and treatment planning. In clinical practice, delineating the boundaries between tumors and atelectasis on CT images presents significant challenges, complicating the precise implementation of radiotherapy. This difficulty can lead to increased exposure of healthy tissue, resulting in severe pneumonitis, or insufficient coverage of the tumor region, thereby elevating the risk of local recurrence. For developing countries, the economic burden of using PET/CT further limits accessibility. Therefore, it is essential to develop tools that can automatically segment atelectasis and tumors on CT scans, assisting clinicians in radiotherapy planning and improving clinical outcomes. The purpose of this study is to leverage the nnUNet deep learning framework to address these challenges, providing a reliable and automated solution for segmentation of tumors to improve clinical workflows. Material/Methods: The dataset consisted of CT images from 413 patients, with annotations for two segmentation tasks: a larger region encompassing atelectasis (including tumors) and a smaller region delineating tumors within the atelectasis. Of these cases, 313 were randomly selected for training, and 100 were reserved for testing. Given that atelectasis regions are relatively straightforward for clinicians to delineate compared to the task of accurately segmenting tumors within these regions, preprocessing involved enclosing the atelectasis regions in the smallest 3D bounding box. This targeted preprocessing step aimed to reduce complexity, allowing for a more precise segmentation of tumors embedded within atelectasis. The nnUNet framework was employed due to its robust, self-adaptive architecture for medical image segmentation. Model performance was evaluated using standard metrics, including Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), to compare predictions with expert manual annotations. Results: The nnUNet framework demonstrated high performance in tumor segmentation, achieving a mean Dice Similarity Coefficient of 0.836 for tumors on the test set. The Intersection over Union metrics further validated the model’s reliability, with a value of 0.745. Visual inspection confirmed the model's ability to delineate pathological regions accurately.