S1619
Physics - Autosegmentation
ESTRO 2026
and US. Medically trained professionals annotated the ROIs following international guidelines2,3,4,5 and cross-reviewed them. Finally, an independent Radiation Oncologist reviewed and approved the contours. A commercial CE-marked AI–based segmentation solution (Contour+TM) was used to generate LN contours. The automatic and manual contours were compared using Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and surface DSC with 2mm tolerance (sDSC2).A subset of 11 radiotherapy CT scans for each gender was used for qualitative evaluation by 3 independent Radiation Oncologists from different institutions. Each evaluator reviewed all scans and scored the AI contours from 1 (unusable) to 5 (use-as-is). Results: The models achieved high accuracy for all LNs, with the male ROIs scoring slightly higher metrics. Corresponding median HD95 values situated between 4.6 and 6.0 mm for female and 3.1 and 3.8 for the male LNs. The results are presented in Table 1.Table 1. Quantitative evaluation
efficient and reliable tool to streamline contouring workflows in prostate cancer radiotherapy. References: 1. Yorke, A. A., et al. (2019). Pelvic Reference Data (Version 1) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.2019.WOSKQ5OO2. Small W Jr et al., Int J Radiat Oncol Biol Phys. 2008;71(2):428- 434. doi:10.1016/j.ijrobp.2007.09.0423. Lawton CA et al., Int J Radiat Oncol Biol Phys. 2009;74(2):383-387. doi:10.1016/j.ijrobp.2008.08.0024. Harris VA et al., Int J Radiat Oncol Biol Phys. 2015;92(4):874-883. doi:10.1016/j.ijrobp.2015.03.0215. Hall WA et al., Int J Radiat Oncol Biol Phys. 2021;109(1):174-185. doi:10.1016/j.ijrobp.2020.08.034 Keywords: deep-learning, prostate, cervical cancer Morphometric Outlier Detection for Automated Quality Assurance of Deep Learning-Segmented Thoracic Structures Amal Joseph Varghese 1 , Barbara Stam 1 , Jan-Jakob Sonke 1 , Leonard Wee 2 , Zeno Gouw 1 , Tomas Janssen 1 1 Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 2 Department of Radiation Oncology (Maastro), GROW School for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands Digital Poster 5104
Purpose/Objective: To conduct dose-effect analyses consistent
Qualitative evaluation shows high satisfaction among evaluators, with an overall average score of 4.62 across all investigated ROIs. More than 90% of the grades were 4 or 5 for all ROIs, requiring no or minor changes. For the lymph node ROIs more than 90% of the grades suggested they are suitable for clinical use without any or with only minor edits. Results for each ROI are presented in Table 2.Table 2. Qualitative evaluation
segmentations of the organs of interest (OoI) are beneficial. Deep learning-based segmentation might provide this; however, inaccuracies will still be present. When analysing a large cohort of patients, manual verification of all structures is not feasible, therefore there is a need of contour quality assurance. We hypothesize that morphometric outlier detection can identify segmentation errors from DL bases segmentation algorithms. The aim of this study is to develop a morphometric outlier detection framework for quality assurance of thoracic structure segmentations in large-scale lung cancer datasets. Material/Methods: Total-Segmentator[1] and Platipy[2] models provided segmentations of 33 thoracic structures on 1441 NSCLC retrospective planning CTs (development: 1000; test: 441). Post-processing removed overlap with the Gross Tumour Volume (GTV) and disconnected sub- volumes <5% of main structure volume. Seventeen 3D shape and bounding-box features were extracted using PyRadiomics (v3.0.1). After excluding features with Pearson correlation >0.75 (prioritizing simpler features), only the non-collinear features that were in the feature sets of at least 20 structures were selected. Morphometric outliers were identified using
Conclusion: The AI-based auto-contouring software demonstrated high accuracy and clinical acceptability for both male and female pelvic elective nodal irradiation. Quantitative evaluation showed strong agreement between automated and manual contours across all lymph node ROIs. Qualitative assessments by independent radiation oncologists confirmed that the majority of contours required no or only minor edits, indicating the model’s suitability for clinical implementation. Although more real-world data is needed, the results support the use of the device as an
Made with FlippingBook - Share PDF online