This study presents a fine-tuned adaptation of Meta’s Segment Anything Model (SAM) using Vision Transformer (ViT) backbones for image segmentation in robotic bariatric surgery. The research addresses the challenge of performing high-accuracy segmentation under limited data conditions, where existing convolutional models often struggle. The objective is to enhance intraoperative scene understanding by refining the SAM decoder specifically for adipose and muscular tissue segmentation. A small, annotated laparoscopic image dataset was expanded via augmentation to train the proposed model. Experimental results demonstrate that the fine-tuned SAM achieves a Dice coefficient of 86.5%, outperforming classical U-Net, ResUNet++, and Swin-Unet models. The method balances segmentation accuracy with decoder-level efficiency, indicating its potential for real-time surgical guidance in minimally invasive procedures.