The paper introduces CellSAM, a universal foundation model designed to automate cell segmentation across diverse biological imaging data. While previous deep learning models were often limited to specific domains, this new framework generalizes across different cell morphologies and imaging modalities, including fluorescence and brightfield microscopy. The system integrates a specialized object detector called CellFinder, which automatically identifies cells and provides bounding box prompts to the Segment Anything Model (SAM). This innovative approach achieves human-level accuracy and enables high-performance segmentation in both zero-shot and few-shot learning scenarios. Furthermore, the authors demonstrate that CellSAM can be seamlessly integrated into complex workflows like spatial transcriptomics and live-cell imaging to accelerate biological discovery.

References:

• Marks M, Israel U, Dilip R, et al. CellSAM: a foundation model for cell segmentation[J]. Nature Methods, 2025: 1-9.