Microscopy Cell Segmentation: Review and Benchmarking of Task-Specific and Foundation Models

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Cell segmentation is crucial for biomedical imaging analysis, including single-cell studies and pathology.
• Classical deep learning models (U-Net, StarDist, HoVer-Net) show limitations in generalizing across different microscopy modalities due to domain-specific training.
• Foundation models, exemplified by the Segment Anything Model (SAM), offer a promising shift towards universal and adaptable segmentation.

Purpose of the Study:

• To review advancements in microscopy cell segmentation, covering both traditional and foundation model-based techniques.
• To experimentally compare the performance of four representative cell segmentation models: YOLO-SAM, CellSAM, Cellpose-SAM, and StarDist.
• To evaluate model performance on diverse microscopy data (fluorescence and brightfield) with varied cell types and shapes.

Main Methods:

• Comprehensive literature review of cell segmentation methods in microscopy.
• Experimental benchmarking of YOLO-SAM, CellSAM, Cellpose-SAM, and StarDist.
• Testing models on fluorescence and brightfield microscopy datasets encompassing diverse cell populations and morphologies.

Main Results:

• Foundation models demonstrate significant promise for cross-domain generalization in cell segmentation.
• The study identified trade-offs between accuracy, robustness, and adaptability among the evaluated models.
• Results highlight the potential of foundation models to overcome limitations of traditional, domain-specific approaches.

Conclusions:

• Foundation models represent a paradigm shift in microscopy cell segmentation, enhancing adaptability and generalization.
• The experimental comparison provides practical insights into model selection for researchers.
• Future work should focus on developing robust and generalizable cell segmentation methods leveraging foundation models.