Lightweight open-source fine-tuning of SAM2 enables domain-specific microscopy segmentation

Area of Science:

• Computational Biology
• Bioimage Analysis
• Machine Learning for Microscopy

Background:

• Accurate quantitation of biological structures in microscopy images is essential for analysis.
• Automated segmentation of cellular and tissue structures is challenging due to image variations.
• Existing foundation models like SAM require significant computational resources and large datasets for adaptation.

Purpose of the Study:

• To introduce a lightweight, open-source Google Colab pipeline for efficient fine-tuning of Segment Anything Model 2 (SAM2).
• To enable domain-specific adaptations of SAM2 for robust biological image segmentation without specialized hardware or extensive training data.
• To lower computational and data barriers for advanced image segmentation in microscopy.

Main Methods:

• Developed a lightweight, open-source Google Colab pipeline for fine-tuning SAM2.
• Employed mask-decoder fine-tuning coupled with biologically informed post-processing.
• No additional architectural layers or specialized hardware were required.

Main Results:

• Achieved robust segmentation across diverse imaging modalities, including hippocampal and single-cell images.
• Demonstrated substantial accuracy gains compared to basic SAM2.
• Matched the performance of leading segmentation tools with significantly lower computational cost.

Conclusions:

• The developed framework provides a scalable and accessible paradigm for domain-specific SAM2 adaptation in microscopy.
• This approach significantly lowers computational and data requirements for advanced image segmentation.
• Enables researchers to achieve high segmentation accuracy with minimal input and resources.