CellBinDB: a large-scale multimodal annotated dataset for cell segmentation with benchmarking of universal models

Area of Science:

• Computational Biology
• Bioimage Analysis
• Digital Pathology

Background:

• Cell segmentation is vital for quantitative biological image analysis.
• Deep learning models excel but lack generalizability due to modality-specific tuning.
• A need exists for comprehensive datasets to train universal models and evaluate segmentation techniques.

Purpose of the Study:

• Introduce CellBinDB, a large-scale multimodal annotated dataset.
• Facilitate training of universal cell segmentation models.
• Benchmark current cell segmentation technologies.

Main Methods:

• Created CellBinDB with over 1,000 annotated images from human and mouse tissues.
• Included diverse staining: DAPI, ssDNA, H&E, and multiplex immunofluorescence.
• Benchmarked 8 state-of-the-art cell segmentation methods using the dataset.

Main Results:

• CellBinDB covers >30 normal and diseased tissue types.
• Complex cell shapes were found to reduce segmentation accuracy.
• Higher image gradients were correlated with improved boundary detection.

Conclusions:

• CellBinDB supports development of versatile cell segmentation solutions.
• Insights gained can refine segmentation strategies for diverse imaging scenarios.
• The dataset aids in advancing quantitative biological image analysis.