ESTRO 2025 - Abstract Book

S2531

Physics - Autosegmentation

ESTRO 2025

Conclusion: This study demonstrated the potential of AI models to automate cachexia monitoring during radiotherapy. By auto contouring 3D CBCT images and analyzing the dynamics of muscle and adipose tissue wasting, the approach may ultimately contribute to the improvement of patient management, detection of early onset of cachexia, and treatment outcomes.

Keywords: Lung Cancer, Cancer Cachexia, Deep Learning

References: 1. J. Zhang et al (2024), “Cancer cachexia as a predictor of adverse outcomes in patients with non-small cell lung cancer: A meta-analysis,” Clinical Nutrition, vol. 43, no. 7, pp. 1618–1625, Jul. 2024, doi: 10.1016/j.clnu.2024.05.025 2. Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211

3539

Digital Poster Automatic relabeling and QA analysis of OAR contours using auto-segmentation algorithms Aarohi Garg, Hannu Laaksonen, Elena Czeizler Treatment Planning Technology, Varian Medical Systems, a Siemens Healthineers Company, Helsinki, Finland Purpose/Objective: Variations in organ-at-risk (OAR) labeling and contouring techniques, create discrepancies that disrupt treatment planning and automated machine learning pipelines. Missing OAR structures are cause interruptions. We have devised a tool using autosegmentation to tackle these challenges. Material/Methods: This work uses the Head-Neck-PET-CT [3] dataset from The Cancer Imaging Archive (TCIA) [2], focusing on 65 patients and 16 OARs. TCIA is an open image archive service that supports cancer research. An in-house auto segmentation model generates new OAR contours [1], and the Sørensen–DICE index measures similarity between the generated and existing structures. We employ these for three use cases (workflow in Figure 1): Relabeling existing OARs: DICE score thresholds are set per organ, and the raw contour with the highest score above the threshold is matched to the generated organ. Existing OAR structures are then relabeled with a standardized label, with the original label If a patient’s structure isn’t mapped, e.g., because it was not contoured, the system can automatically add the automatically segmented contour. These contours are copied from the generated RS file and appended to the patient’s DICOM RS file, clearly flagged as autogenerated for transparency. Quality assurance analysis of OARs: We also calculate a DICE score using only common contour slices and show the difference between the two scores. A predefined threshold flags cases of partial contouring that may indicate errors or inconsistencies. The system also analyzes the box-plot distribution of the highest DICE scores for the mapped organs. A wider distribution in the box plot for an OAR suggests a significant inconsistency in contouring practices throughout the entire clinic dataset. stored for reference. Filling in missing OARs: