CellSeg3D, Self-supervised 3D cell segmentation for fluorescence microscopy.

Area of Science:

• Cell biology
• Neuroscience
• Bioimaging
• Machine learning

Background:

• Accurate 3D cell structure analysis is vital for biology and neuroscience.
• Generating 3D ground truth data for cell quantification is complex and data-intensive.
• Existing methods struggle with label-scarce biological contexts.

Purpose of the Study:

• To develop novel 3D models and a self-supervised learning method for cell nuclei quantification.
• To create accessible tools (CellSeg3D package) for 3D cell segmentation.
• To establish robust benchmarks using diverse 3D datasets.

Main Methods:

• Introduced a 3D transformer (SwinUNetR) and a 3D self-supervised learning method (WNet3D).
• Developed the CellSeg3D Python package with Jupyter Notebook and napari GUI plugin integration.
• Created a human-annotated mesoSPIM dataset for benchmarking.
• Benchmarked WNet3D against state-of-the-art supervised methods on four diverse 3D datasets.

Main Results:

• The self-supervised WNet3D model achieved performance comparable to supervised methods.
• WNet3D demonstrated effectiveness across varied datasets, including challenging, densely packed nuclei.
• CellSeg3D provides accessible tools for 3D cell segmentation.

Conclusions:

• WNet3D offers a powerful solution for 3D cell nuclei quantification in label-scarce scenarios.
• The CellSeg3D package democratizes access to advanced 3D segmentation tools.
• This work facilitates broader applications of 3D cell analysis in biological research.