SAM-Geo3D: A Geometrical Method to Extend SAM to 3D

Jikai Zhang, Zafer Yildiz, Hanxue Gu, Haoyu Dong, Maciej A Mazurowski

Keywords: Segmentation, Segment Anything Model, 3D SAM

Abstract: The advent of Segment Anything Model (SAM) offers a promising approach for image segmentation tasks. However, SAM produces less satisfying results in the segmentation of 3D medical images, such as MRI, as compared to 2D natural images. To eliminate this deficiency and avoid onerous deep learning training, we proposed SAM-Geo3D, a geometrical method that extends SAM into the 3D manner. Given a few prompt points on a target component, SAM-Geo3D segments the component through all slices in 3D. We validated SAM-Geo3D on five knee MRI volumes. Results showed that SAM-Geo3D outperforms SAM when using the same, limited number of input prompt points.

Submission Number: 156