SAM² for abdomen: One-shot and no finetuning

The study introduces SAM<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, a novel method for fully automatic abdominal organ segmentation using the Segment Anything Model 2 (SAM2) and Self-Supervised Anatomical Embedding (Emb-SAM). Unlike traditional approaches, our model achieves superior segmentation performance without the need for finetuning or large labeled datasets, relying on a single labeled image. SAM2 leverages its memory bank for mask propagation, originally designed for video segmentation, which we adapt for 3D medical imaging by treating CT slices as sequential frames. Emb-SAM generates precise pseudo-labels by matching anatomical points across images using self-supervised learning. The proposed method effectively segments abdominal organs such as the liver, kidneys, spleen, and aorta, demonstrating superior consistency compared to baseline models on the BTCV dataset. Experimental results show that our model achieves competitive performance, significantly reducing computational costs and manual intervention, thus offering a promising solution for automated medical image segmentation.