Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep...

Area of Science:

• Computational biology
• Bioinformatics
• Medical imaging analysis

Background:

• Accurate cell segmentation is crucial for analyzing tissue imaging data.
• Existing datasets for training segmentation models are limited in size and scope.
• Automated extraction of cellular features from images remains a significant challenge.

Purpose of the Study:

• To develop a comprehensive dataset for training cell segmentation models.
• To create a deep learning-based algorithm for accurate and robust cell segmentation.
• To enable advanced analysis of cellular features and morphology changes in biological samples.

Main Methods:

• Construction of TissueNet, a dataset with over 1 million manually labeled cells.
• Training of Mesmer, a deep learning segmentation algorithm, using the TissueNet dataset.
• Adaptation of Mesmer for analyzing highly multiplexed datasets and quantifying cell morphology.

Main Results:

• Mesmer demonstrated superior accuracy and generalization across diverse tissue types and imaging platforms compared to previous methods.
• Mesmer achieved human-level performance in cell segmentation tasks.
• Automated extraction of subcellular protein localization and quantification of cell morphology changes during human gestation were enabled by Mesmer.

Conclusions:

• TissueNet and Mesmer represent a significant advancement in cell segmentation for tissue imaging analysis.
• The developed tools facilitate automated extraction of complex cellular features and biological insights.
• The release of code, data, and models fosters community-driven research in computational pathology and bioimage analysis.