ESTRO 2025 - Abstract Book

S4362

RTT - Treatment planning, OAR and target definitions

ESTRO 2025

Results:

Table 1. Average accuracy (DSC, %) and time (min/case) for contours by Manual, SAM, and SAM2 (+C indicates correction; N = successful experiments). Table 1 shows the accuracy and time for each method and organ. In terms of DSC, SAM (85%) and SAM2 (85.6%) were inferior to manual contouring (89.9%), but after correction, both were competitive (SAM+C 90.0%, SAM2+C 88.8%). With small sample size (N), the prostate challenged SAM, and the seminal-vesicles challenged both SAM and SAM2 due to the characterics of these organs such as heterogeneous texture and blurred boundaries. Total time needed for contouring (including SAM inference) was significantly lower for SAM2 (13.1) and somewhat lower for SAM (15.7) compared to manual (17.4). The observer's experience was the most significant factor influencing contouring time, with substantial variability observed across annotators. Conclusion: In conclusion, our study demonstrates SAM's potential to enhance interactive segmentation in medical imaging. Our findings revealed SAM2 significantly reduced contouring time for select organs, including the pelvis body, rectum, and bladder. SAM2 consistently outperforms SAM, even with corrections, maintaining contour quality. These findings highlight the critical need for observer training and standardization to improve segmentation quality and consistency across imaging applications.

Keywords: OAR delineation, Segment Anything Model, MR image

References: Koos, Krisztian, et al. "202: Evaluation of foundation models for semi-automated organ delineation in MR images." Radiotherapy and Oncology 194 (2024): S2995-S2999