Using Segment Anything Model 2 for Zero-Shot 3D Segmentation of Abdominal Organs in Computed Tomography Scans to...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Medical image segmentation is vital for diagnosis and treatment planning but often requires manual effort and specialized data.
• The Segment Anything Model 2 (SAM 2) offers potential for automated 3D medical image segmentation without domain-specific training.
• SAM 2's effectiveness in abdominal CT imaging is currently unexplored.

Purpose of the Study:

• To evaluate the zero-shot performance of SAM 2 for 3D segmentation of abdominal organs in CT scans.
• To investigate the impact of prompt settings on SAM 2's segmentation accuracy.

Main Methods:

• Retrospective study using a subset of the TotalSegmentator CT dataset from eight institutions.
• Assessed SAM 2's segmentation of eight abdominal organs initiated from three z-coordinate levels.
• Measured performance using Dice Similarity Coefficient (DSC) and analyzed the effect of negative prompts.

Main Results:

• Larger organs like the liver (DSC 0.821) and spleen (DSC 0.891) showed high zero-shot segmentation performance.
• Smaller organs, including the gallbladder (DSC 0.531) and pancreas (DSC 0.361), had lower performance.
• Initial slice selection and negative prompts significantly influenced results; removing negative prompts decreased DSC for six organs.

Conclusions:

• SAM 2 demonstrates promising zero-shot segmentation for certain abdominal organs in CT scans, particularly larger ones.
• Segmentation accuracy is significantly influenced by negative prompts and initial slice selection.
• Optimizing prompt engineering and slice selection is crucial for maximizing SAM 2's utility in medical imaging.