ESTRO 2025 - Abstract Book

S2438

Physics - Autosegmentation

ESTRO 2025

Results: The efficacy of the proposed approach was validated using the SegTHOR dataset, comprising 11,084 CT slices from 60 patients, and evaluated across multiple architectures, including 2D and 3D convolutional neural networks, transformer-based models, and hybrid systems. The integration of trainable morphological layers resulted in an average improvement of 0.66% in the Dice Similarity Coefficient (DSC) and a reduction of 1.09 mm in the 95 th percentile Hausdorff Distance (HD95), both statistically significant (p = 0.045). Qualitative analyses demonstrated enhanced delineation of organ contours, particularly in anatomically challenging regions, underscoring the effectiveness of the proposed refinement approach.