Prompt Once, Segment Everything: Leveraging SAM 2 Potential for Infinite Medical Image Segmentation with a Single Prompt

Abstract

Semantic segmentation of medical images holds great potential for improving diagnostic and surgical procedures. Radiology specialists can benefit from automated segmentation tools that facilitate the identification and isolation of regions of interest in medical scans. However, to obtain accurate results, it is necessary to develop and train sophisticated Deep Learning models adapted to this specific task, a capability not universally accessible. SAM 2 is a foundation model designed for image and video segmentation tasks, built on the foundation of its predecessor, SAM. This article presents a novel approach that leverages SAM 2’s video segmentation capabilities to reduce the number of prompts needed to segment an entire volume of medical images. The study first compares the performance of SAM and SAM 2 in medical image segmentation. Evaluation metrics such as Jaccard Index and Dice Score are used to measure segmentation accuracy and quality. Then, our novel approach is presented. Statistical tests include the comparison of accuracy gains and computational efficiency, focusing on the trade-off between resource usage and segmentation time. The results show that SAM 2 achieves an average improvement of 1.76% in Jaccard Index and 1.49% in Dice Score compared to SAM, although with a tenfold increase in segmentation time. Our novel segmentation approach reduces by 99.95% the number of prompts needed to segment a volume of medical images. We demonstrate that it is possible to segment all slices of a volume and, even more, of an entire dataset, with a single prompt, achieving results comparable to those obtained by the most advanced models explicitly trained for this task. Our approach s…

José M. Conejero

Associate Professor

Associate Professor at the University of Extremadura. Research interests include model-driven development, data science, and machine learning.

Roberto Rodriguez-Echeverria

INTIA Director and Associate Professor

Associate Professor at the University of Extremadura. Software passionate, Deep learner, MTB rider and father of 2.