Densely Populated Cell and Organelles Segmentation with U-Net Ensembles

Area of Science:

• Biomedical imaging
• Computational biology
• Machine learning

Background:

• Segmenting complex cell structures like platelets in thrombi is challenging due to intertwined morphology.
• Existing single-network approaches struggle with fine details and long-range smoothing.

Purpose of the Study:

• To present a robust dual-network pipeline for accurate cell and organelle segmentation.
• To improve the segmentation of activated platelets within thrombi.
• To achieve state-of-the-art results in biological image segmentation challenges.

Main Methods:

• A dual-network pipeline combining fine-detail detection near membranes and long-range smoothing.
• Utilizing multi-axis neural network predictions to capture 3D correlations with 2D networks.
• Applying the watershed clustering algorithm enhanced by the dual-network approach.

Main Results:

• Significantly improved segmentation performance compared to single-network methods.
• Quantitative morphological measurements of hundreds of platelets consistent with manual segmentation.
• Achieved state-of-the-art results on the CREMI neuron segmentation challenge data.

Conclusions:

• The dual-network pipeline offers a robust solution for challenging cell and organelle segmentation tasks.
• This approach enhances the analysis of platelet morphology and contributes to advancements in bioimage analysis.
• The method demonstrates broad applicability, achieving top performance in a benchmark neuron segmentation challenge.