CellSighter: a neural network to classify cells in highly multiplexed images

Area of Science:

• Biomedical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Multiplexed imaging allows in situ measurement of multiple proteins, crucial for understanding tissue cell types and states.
• Automated cell classification in these complex images is essential for high-throughput analysis but remains a significant challenge.

Purpose of the Study:

• To develop and validate CellSighter, a deep learning pipeline designed to accelerate and enhance cell classification in multiplexed imaging data.
• To provide a robust and accurate method for identifying diverse cell types within tissue samples.

Main Methods:

• CellSighter utilizes a deep learning approach trained on expert-labeled multiplexed images.
• The pipeline analyzes protein expression levels and spatial features, including subcellular patterns, for classification.
• It incorporates techniques to reduce overfitting and can be trained with limited labeled examples.

Main Results:

• CellSighter achieves over 80% accuracy in classifying major cell types across different imaging platforms, nearing inter-observer concordance.
• The model demonstrates generalization capabilities and effectively learns from protein expression and spatial features.
• It provides prediction confidence scores, enabling expert oversight.

Conclusions:

• CellSighter significantly reduces manual effort and time required for cell classification in multiplexed imaging.
• The tool enhances accuracy and consistency in cell type identification across various datasets.
• It empowers researchers with a powerful, efficient, and reliable method for analyzing complex biological tissues.