ESTRO 2025 - Abstract Book

S2451

Physics - Autosegmentation

ESTRO 2025

Results:

Inference times for nnUNet, B-spline, and TransMorph were 144, 100, and 1.2 seconds, respectively. Average DSCs for each model are summarized in Table 1. TM MSE+DICE achieved the highest mean DSC of 0.90/0.79 and 0.84/0.77 for OARs/GTV in the C1 and C2 stage 1-2 dataset, respectively. On out of distribution data, TM MSE+DICE achieved mean DSCs of 0.87/0.79, 0.83/0.78, and 0.87/0.75 for the three stage III cohorts, respectively. Statistical tests showed that TM MSE+DICE and TM MSE were significantly better (p<0.05) than other methods on the C1/C2 stage 1-2 datasets while no significant difference between them was found.