ESTRO 2025 - Abstract Book

S4361

RTT - Treatment planning, OAR and target definitions

ESTRO 2025

used in different modalities. In this work the evaluations of SAM and SAM2 for organ delineation were conducted using the open-source Slicer3D tool.

Material/Methods: The interactive segmentation module of Slicer3D was extended to support SAM models with standard prompts, such as bounding box and points. The adapted SAM2 processes volumetric images as videos, utilizing a two-pass system starting from a user-defined contour and iterating through axial slices. Using manual contours as prompts for SAM2 was very effective, improving segmentation accuracy in other slices. This pipeline was identified as the most efficient for volumetric images.

Figure 1. Ground-truth (green), SAM-assisted (red) and SAM2-assisted (blue) organ contours We analyzed three patient scans (T2-weighted MR) contoured for seven pelvic organs using three methods: manual, SAM, and SAM2 with and without correction from four observers (Fig1). The contours were compared to ground truth (created from manual contours by consensus) using DICE Similarity Score (DSC), and the mean DSC was computed for each organ. Time to create the contour was also measured for each method and organ.