ESTRO 2025 - Abstract Book

S2493

Physics - Autosegmentation

ESTRO 2025

This study evaluates the performance of an automated segmentation tool, TotalSegmentator [1], on a diverse set of radiotherapy planning CT images. The aim is to validate its capability to support high-throughput research while maintaining accuracy and reliability. Material/Methods: The validation cohort of 2,500 radiotherapy planning CT images from 2,300 patients with heterogeneous disease sites, treated since 2022, was extracted from the Aria system using pyesapi-based script. To avoid partially segmented organs, manual screening was performed for all extracted masks. The most recent version of TotalSegmentator (2.4.0) was used for validation. Segmentation was performed with a predefined list of structures in the ‘roi_subset_robust’ parameter to ensure computational efficiency and robustness. The validation included structures overlapping between the TotalSegmentator default list of 117 structures and the list of ROIs used in our department, resulting in 21 structures (Table 1). Dice score, Hausdorff distance, and relative volume difference were used to evaluate segmentation quality. Results: The analysis (Table 1) shows that for 67% of the evaluated ROIs, the median Dice score exceeds 0.85, while thyroid (0.83), heart (0.78), trachea (0.74), prostate (0.72), and vena cava inf. (0.71) are showing lower scores; carotid arteries were not detected by TotalSegmentator. Manual analysis revealed that thyroid, while generally well-segmented, sometimes lacks a connection between two lobes (Fig. 1A). Discrepancy in heart segmentation (Fig. 1B) arises from differing guidelines: our department follows Standardized Myocardial Segmentation and Nomenclature for Tomographic Imaging of the Heart, which differs from TotalSegmentator standards. Complete segmentations of the vena cava inf., as well as carotid arteries, were present only once in the validation dataset, rendering it insufficient for reliable conclusions. The median Hausdorff distance was below 1.5cm for 71% of the structures, exceptions are spleen (2.0cm), stomach (2.5cm), heart (3.2cm), and vena cava inf. (3.5cm). The analysis of the high (>10cm) Hausdorff distances revealed the rare presence of falsely predicted voxels distant from the target ROI (Fig. 1C). Relative volume error showed that TotalSegmentator tends to underestimate structures’ sizes, compared to validation dataset.