Novel adaptation of video segmentation to 3D MRI: efficient zero-shot knee segmentation with SAM2 (Conference Presentation)
Lists
Only you see this
Abstract
Medical image segmentation models often suffer from poor generalizability and annotation costs. We propose an adaptation to Segment Anything Model 2 (SAM2) that leverages its video predictor and memory component to segment 3D medical images with no training and minimal prompting. When segmenting the tibia bone in knee MRI, the model achieves a median Dice similarity coefficient of 0.9643 with only one prompt per volume. We also present comparisons across different model sizes, prompt schemes, and the previous SAM1 model.
BibTeX
@misc{yu2024novel,
author = {A.S. Yu and Mohsen Hariri and Xuecen Zhang and Mingrui Yang and Vipin Chaudhary and Xiaojuan Li},
title = {{Novel adaptation of video segmentation to 3D MRI: efficient zero-shot knee segmentation with SAM2 (Conference Presentation)}},
year = {2024},
eprint = {2408.04762},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2408.04762},
}