ESTRO 2025 - Abstract Book

S2532

Physics - Autosegmentation

ESTRO 2025

Figure 1: Workflow. The tool provides abstraction and usability by containerizing the solution. The code uses patient wise parallel processing to significantly enhance efficiency and reduce processing time when working with large datasets. Results: Figure 2 illustrates the original and updated OAR labels where original French labels are standardized to template English labels with 91.82% accuracy. The box plot highlights QA distribution, with Trachea showing the widest spread, indicating inconsistent contouring.

Figure 2: Results - Left: Example of original and updated OAR labels; Right: Box plot DICE score distribution for each OAR.

Conclusion: We implemented a tool that uses auto-segmentation to address common problems while dealing with DICOM patient data, such as inconsistent labels, missing structures, and inadequate quality of contours. This makes it especially useful for large-scale data preprocessing workflows in clinical environments.

Keywords: OARs, relabeling, QA