MedicoSAM: Robust Improvement of SAM for Medical Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for clinical practice and research.
• Current deep learning models require task-specific training and labeled data, which is costly.
• Vision foundation models like SAM offer potential for universal medical image segmentation.

Purpose of the Study:

• To investigate optimal finetuning strategies for improving SAM in medical image analysis.
• To evaluate the performance of finetuned SAM on diverse segmentation tasks.
• To introduce MedicoSAM, an enhanced model for medical image segmentation.

Main Methods:

• Compared various finetuning strategies on a large, diverse medical imaging dataset.
• Evaluated finetuned models on interactive and automatic semantic segmentation tasks.
• Assessed compatibility with existing data annotation tools.

Main Results:

• Finetuning strategies significantly improve SAM's performance on medical images.
• Interactive segmentation shows pronounced performance gains.
• Semantic segmentation benefits, though advantages over traditional methods are inconsistent.
• MedicoSAM demonstrates practical value and compatibility with annotation tools.

Conclusions:

• Strategic finetuning is key to unlocking SAM's potential for medical image segmentation.
• MedicoSAM represents a significant advancement, offering practical benefits for clinical and research applications.
• The developed model facilitates efficient and accurate medical image analysis.