ESTRO 2025 - Abstract Book

S2557

Physics - Autosegmentation

ESTRO 2025

remaining four session images per patient and included geometric accuracy (Dice score (DSC)), average surface distance (ASD)) and dosimetric analysis using a previously-implemented treatment plan, measuring D0.1cm 3 differences. Results: Each training required only 5h, with results generated in 30sec. Figure1 and Table1 demonstrate that the proposed CondPSeg approach outperformed registration outputs, achieving a DSC and ASD of 0.93±0.01 and 1.86±0.48mm, compared to DSC and ASD of 0.76±0.04 and 6.20±2.67mm. CondPSeg demonstrated less inaccuracy across all OARs, except patient one's stomach contours ASD. Dosimetric analysis of D0.1cm³ revealed that CondPSeg contours deviated by 2.51±3.82% from the ground-truth, while Prop-ROIs showed a larger deviation of 5.70±6.26%, when normalized to prescribed dose (40Gy). These indicate that CondPSeg requires fewer modifications, providing clinicians with more time to focus on accurately reviewing OARs within the PTV+2cm margin.