Medical SAM adapter: Adapting segment anything model for medical image segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• The Segment Anything Model (SAM) demonstrates broad applicability in image segmentation.
• SAM exhibits limitations in medical image segmentation due to a lack of domain-specific knowledge.
• Adapting foundation models like SAM for specialized tasks is crucial for advancing AI in medicine.

Purpose of the Study:

• To develop an effective method for adapting SAM to medical image segmentation tasks.
• To enhance SAM's performance in segmenting diverse medical imaging modalities.
• To propose lightweight adaptation techniques that preserve SAM's general capabilities.

Main Methods:

• Introduction of the Medical SAM Adapter (Med-SA), a novel adaptation technique for SAM.
• Utilizing a light adaptation strategy, avoiding full model fine-tuning.
• Development of Space-Depth Transpose (SD-Trans) for 3D medical image adaptation.
• Implementation of Hyper-Prompting Adapter (HyP-Adpt) for prompt-conditioned adaptation.

Main Results:

• Med-SA demonstrated superior performance across 17 diverse medical image segmentation tasks.
• The adaptation process required updating only 2% of SAM's parameters (13 million).
• The proposed methods successfully adapted 2D SAM for 3D medical image analysis.

Conclusions:

• Med-SA offers an efficient and effective solution for enhancing SAM in medical image segmentation.
• Lightweight adaptation techniques can successfully integrate domain-specific knowledge into large foundation models.
• The approach shows significant potential for improving AI-driven diagnostics and analysis in healthcare.