ESTRO 2024 - Abstract Book

S2998

Physics - Autosegmentation

ESTRO 2024

Figure 1: Mean DSC values of ViT-B with example segmentations.

Conclusion:

The experiments showed that box+points prompting strategy provides the best accuracy regardless of the anatomy. Furthermore, SMO performs better as multi-mask is designed for ambiguous prompts. Refinement step could improve the result for box prompt. Given ideal (ground-truth based) prompts, FMs are superior to the supervised model and suitable for annotating medical images. We expect similar performance from the human operator who follows the same prompting strategy, but this shall be confirmed by future study.

Keywords: SAM, zeroshot, OAR

References: