LeSAM: Adapt Segment Anything Model for Medical Lesion Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• The Segment Anything Model (SAM) excels in natural image segmentation but struggles with medical images due to domain differences.
• Accurate medical lesion segmentation is crucial for diagnosis and treatment planning but is challenging for existing models, especially for irregular or low-contrast lesions.

Purpose of the Study:

• To adapt SAM for improved medical lesion segmentation, addressing its limitations in handling medical image characteristics.
• To develop a specialized model, LeSAM, that integrates general segmentation capabilities with medical domain knowledge.

Main Methods:

• LeSAM employs an efficient adaptation module to learn medical-specific features, merging them with SAM's pre-trained knowledge.
• A modified, lightweight U-shaped network is used as a mask decoder for enhanced lesion boundary delineation and efficient training.
• The method was evaluated on diverse lesion segmentation tasks across CT, MRI, ultrasound, dermoscopic, and endoscopic imaging modalities.

Main Results:

• LeSAM outperformed state-of-the-art methods in 8 out of 12 lesion segmentation benchmarks.
• The model achieved competitive performance on the remaining 4 datasets, demonstrating broad applicability.
• Ablation studies confirmed the efficacy of the proposed adaptation modules and the modified decoder design.

Conclusions:

• LeSAM effectively bridges the domain gap, significantly improving SAM's performance for medical lesion segmentation.
• The proposed adaptation strategy and lightweight decoder offer a promising approach for specialized medical image analysis tasks.