Characterization of tumour heterogeneity through segmentation-free representation learning on multiplexed imaging...

Area of Science:

• Computational pathology
• Tumor microenvironment analysis
• Biomarker discovery

Background:

• High-dimensional multiplexed imaging offers detailed insights into tumor tissue spatial organization.
• Characterizing tumor microenvironment heterogeneity is challenging due to data scale and complexity.

Purpose of the Study:

• To leverage self-supervised representation learning for distinguishing tumor microenvironment morphologies.
• To precisely characterize distinct microenvironment signatures using imaging mass cytometry data.

Main Methods:

• Trained a vision transformer using self-supervised masked image modeling on high-dimensional multiplexed mass cytometry images.
• Employed a segmentation-free, pixel-level approach to retain morphological and biomarker distribution information.
• Applied the model to a lung tumor dataset.

Main Results:

• Successfully distinguished morphological differences in tumor microenvironments.
• Identified and validated a specific monocytic signature within the tumor microenvironment.
• The identified monocytic signature was associated with poor prognosis.

Conclusions:

• Self-supervised representation learning is effective for analyzing complex imaging mass cytometry data.
• The developed vision transformer accurately characterizes tumor microenvironment heterogeneity.
• The monocytic signature represents a potential prognostic biomarker in lung cancer.