ESTRO 2025 - Abstract Book

S2511

Physics - Autosegmentation

ESTRO 2025

Results: As in Figure 2, the use of PG significantly enhanced segmentation performance when imaging characteristics were different. Compared to the original DynUnet, the mean improvements were 5.0% in DSC, 14.3% in HD, 25.6% in ASSD, and 41.1% in RVE. Mean±STD of coverage, selectivity, gradient index, and D95%/99% of proposed segmentation with clinical plans were 0.98±0.05, 0.62±0.19, 2.91±0.19, and 13.5±1.7/12.3±2.0 Gy, respectively, indicating clinically acceptable results. The lowest mean overlapping ratio across all datasets were 99.8±1.4%. The average inference time was significantly reduced, decreasing from 43.6 to 0.41 seconds on a GPU and from 1401 to 14.5 seconds in a CPU environment.