Universal consensus 3D segmentation of cells from 2D segmented stacks.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Cell Biology

Background:

• Deep learning has advanced 2D cell segmentation in microscopy but 3D segmentation remains challenging due to complex annotation requirements.
• Manual labeling for 3D cell segmentation is time-consuming and often ambiguous, hindering the development of generalized models.
• Current 3D segmentation methods struggle with dense cell populations and complex cellular morphologies.

Purpose of the Study:

• To develop a computational framework for accurate 3D cell segmentation from existing 2D segmentation methods.
• To create a toolbox, u-Segment3D, that translates 2D instance cell masks into 3D consensus segmentations without requiring 3D training data.
• To demonstrate the versatility and performance of u-Segment3D across diverse biological samples.

Main Methods:

• Developed a theory and toolbox, u-Segment3D, for 2D-to-3D cell segmentation.
• The method is compatible with any 2D instance segmentation tool that outputs pixel-based masks.
• u-Segment3D enhances 2D segmentations to generate a 3D consensus instance segmentation.

Main Results:

• Successfully translated and enhanced 2D instance segmentations to 3D consensus segmentations across 11 diverse datasets (>70,000 cells).
• Demonstrated effectiveness on single cells, cell aggregates, and tissue samples.
• Achieved competitive or superior performance compared to native 3D segmentation methods, especially in crowded or complex cellular environments.

Conclusions:

• u-Segment3D offers a powerful, data-efficient solution for 3D cell segmentation in biological studies.
• The toolbox overcomes the limitations of manual 3D annotation, accelerating microscopy-based research.
• This 2D-to-3D approach provides a generalized and robust method for analyzing cellular structures in three dimensions.