Cellstitch: 3D cellular anisotropic image segmentation via optimal transport

Area of Science:

• Microscopy and Imaging
• Computational Biology
• Bioinformatics

Background:

• Spatial mapping of transcriptional states offers biological insights into tissue context.
• Accurate 3D cell segmentation is crucial for in situ analysis of diseases and development.
• Existing 3D segmentation methods face challenges with accuracy, training data, and image anisotropy.

Purpose of the Study:

• To develop a novel 3D cell segmentation pipeline addressing limitations of current methods.
• To enable accurate cell segmentation from 3D microscopic images, particularly those with anisotropy.
• To improve the analysis of cellular functions and interactions within tissues.

Main Methods:

• Developed CellStitch, a flexible pipeline utilizing an optimal transport (OT) approach for cell correspondence across layers.
• The method segments cells from 3D images without extensive 3D training data.
• Extended the pipeline to interpolate internal slices for recovering isotropic cell morphology from anisotropic images.

Main Results:

• CellStitch was evaluated on eight diverse 3D plant microscopic datasets with varying anisotropy.
• The pipeline significantly outperformed state-of-the-art methods on anisotropic images.
• Achieved comparable segmentation quality to competing methods on isotropic images, with detailed instance-level metrics.

Conclusions:

• The proposed optimal transport-based 3D segmentation pipeline demonstrates superior performance over existing methods.
• CellStitch provides high-fidelity recovery of 3D cell morphology from microscopic images, especially with non-zero anisotropy.
• This advancement facilitates more accurate in situ biological analysis and understanding of cellular structures.