Year: 2025
Type: article
Abstract: PURPOSE: Accurate quantification of body fat distribution is essential in preventive medicine due to its public health outcomes. Specifically, visceral fat area measured from abdominal computed tomography (CT) serves as a key diagnostic criterion for obesity-related metabolic disorders. This study evaluated the efficacy of zero-shot segmentation capabilities of the Segment Anything Model (SAM) and its medical variant (MedSAM) for adipose tissue delineation on abdominal CT images. METHODS: Segmentation of the subcutaneous and visceral fat compartments was performed using point prompts on abdominal images. For a comprehensive evaluation, we compared the foundation models with a supervised U-Net architecture trained specifically for this task. The analysis utilized the publicly available annotated abdominal adipose tissue CT image dataset, which contained expert-annotated ground truth labels for adipose tissue compartments. RESULTS: The average Dice scores for visceral fat segmentation were 0.59 for SAM, 0.29 for MedSAM, and 0.93 for the U-Net. These findings indicate that while SAM achieved moderate performance, MedSAM exhibited limited effectiveness for this task. Both foundation models showed substantial performance gaps compared to the task-specific supervised approach. CONCLUSION: The findings revealed limitations in the zero-shot segmentation capabilities of SAM and MedSAM for specialized medical imaging tasks, particularly for adipose tissue quantification in abdominal CT. Although foundation models offer theoretical advantages in generalizability and deployment efficiency, their performance in specialized medical applications requires significant enhancement through task-specific fine-tuning, and optimized prompt engineering strategies. Further studies are required to enhance the clinical utility of these models for reliable and accessible fat segmentation.
Author: Sho Maruyama
Institution: Gunma Prefectural College of Health Sciences
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4414318383