Segmenting Carotid Plaques with Different SAM Prompts

Zhiwei Zhang, Yimeng Dou, Robert J Dempsey, Carol C. Mitchell, Tomy Varghese

Keywords: Ultrasound Imaging, foundation model, deep learning, image segmentation

TL;DR: This paper experienced on different methods to generate prompt for ultrasound plaque segmentation using foundation models and reported the results.

Abstract:

Large foundation models have been built to address B-mode ultrasound segmentation for regions with homogeneous morphology. However, most of the foundation models still rely on human input to provide interactive prompting during inference. We explored the prompt type and generation when utilizing the ultrasound foundation model, UltraSAM for the plaque segmentation problem and compared the prompt performance with the ground truth baselines. The results showed that automatic generated prompt variants can reach similar performance with ground truth prompting in the bounding box segmentation of ultrasound plaque segmentation.

Submission Number: 67