CellSAM: a foundation model for cell segmentation.

Area of Science:

• Computational Biology
• Biotechnology
• Machine Learning

Background:

• Accurate cell segmentation is crucial for analyzing cellular imaging data.
• Existing deep learning models often lack generalizability across different cell types and imaging modalities.
• There is a need for universal cell segmentation models that can scale and adapt to diverse datasets.

Purpose of the Study:

• To develop CellSAM, a universal deep learning model for generalized cell segmentation.
• To enable accurate cell identification across diverse cellular imaging data.
• To provide a scalable and adaptable solution for bioimage analysis workflows.

Main Methods:

• Leveraged the Segment Anything Model (SAM) architecture.
• Developed a prompt engineering approach for mask generation.
• Trained an object detector, CellFinder, to automatically prompt SAM for cell segmentation.

Main Results:

• CellSAM achieves human-level performance in segmenting mammalian cells, yeast, and bacteria.
• Demonstrated strong zero-shot generalization capabilities across various imaging modalities.
• Showcased improved performance with few-shot learning and applicability in diverse bioimage analysis workflows.

Conclusions:

• CellSAM offers a universal and highly effective solution for cell segmentation.
• The model generalizes well across different cell types, species, and imaging techniques.
• CellSAM represents a significant advancement in automated bioimage analysis.