ESTRO 2025 - Abstract Book

S2518

Physics - Autosegmentation

ESTRO 2025

Version 2024B showed the highest sDSC values, with improvements in the femoral heads, bladder, and prostate (femoral heads v11B/2024B 0.58/0.76; bladder: 0.55/0.77; prostate: 0.33/0.45). These improvements align with the 2024B release notes, which cite an expanded training dataset and model refinement for male pelvic structures. Earlier advancements were also noted for the prostate between versions 12A and 2024A. For head-and-neck structures, most sDSC mean values remained consistent, but there were improvements in the spinal cord (11B/2024B 0.56/0.69) and brainstem (0.50/0.56), while performance declined for the lens (0.91/0.87) and optic nerves (0.76/0.67). No quantitative improvements were observed for breast or lung structures, however improved consistency in contouring the base of the heart could be observed in the spatial plots (Figure 2). Figure 2 shows that version 2024B has greater consistency with the gold standard, as indicated by the colour changes across structures.

Conclusion: The findings underscore the importance of validating autocontouring software after updates. While some changes were documented in the release notes, others were not, emphasizing the need for thorough checks, as per RCR guidelines. Additional modifications were observed beyond those listed, demonstrating that release notes alone may not capture all changes. This highlights the need for comprehensive validation to maintain the accuracy and reliability of autocontouring tools.

Keywords: Monitoring, upgrade, autocontouring

References: [1] NICE. (2023, September 27). Artificial intelligence technologies to aid contouring for radiotherapy treatment planning: early value assessment. [2] Doolan, P. J., et al. (2023). A clinical evaluation of five AI contouring systems for radiotherapy. Frontiers in Oncology [3] Royal College of Radiologists. (2024, June 4). Auto-contouring in Radiotherapy: Guidance for clinicians (open consultation) [4] Mackay, K., et al. (2023). Review of metrics used to assess Auto-Contouring Systems in radiotherapy [5] RaySearch. (2023).RayStation 2024A: Deep Learning Segmentation Model Data Sheet