AutoPrompt-SAM3D: integrated generation and selection for SAM2-based 3D medical segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Segment Anything Model 2 (SAM2) for 3D medical image segmentation requires manual prompts, limiting its application.
• Existing prompt generation methods using auxiliary models have feature extraction bottlenecks and error propagation issues.
• Current approaches struggle with non-salient regions in complex 3D tumor datasets.

Purpose of the Study:

• To develop an automated and reliable prompt generation framework for 3D medical imaging.
• To enhance Segment Anything Model 2 (SAM2) for 3D segmentation without manual intervention.
• To improve the accuracy and efficiency of tumor localization in 3D medical datasets.

Main Methods:

• Proposed AutoPrompt-SAM3D with an Automatic Prompt Generator.
• Hierarchically integrated SAM2's tri-layer features within the generator.
• Implemented a supervised confidence frames filter for prompt selection and a full-sequence processing framework.

Main Results:

• AutoPrompt-SAM3D demonstrated superior 3D medical segmentation performance.
• The framework consistently outperformed or matched state-of-the-art prompt-based methods.
• Experiments were conducted on four public abdominal tumor datasets.

Conclusions:

• AutoPrompt-SAM3D removes the need for manual prompts in SAM2-based 3D segmentation.
• The framework enhances reliability and efficiency in tumor localization.
• Provides a practical tool for large-scale medical image analysis and clinical decision-making.