Characterization of tumor heterogeneity through segmentation-free representation learning.

Area of Science:

• Computational biology
• Cancer research
• Artificial intelligence in pathology

Background:

• Tumor microenvironment (TME) interactions are complex and heterogeneous.
• High-dimensional multiplexed imaging offers insights into tumor spatial organization.
• Analyzing large, complex TME images for discovery remains a significant challenge.

Purpose of the Study:

• To introduce CANVAS, a self-supervised representation learning framework for TME discovery.
• To enable the identification and characterization of novel TME subtypes.
• To overcome limitations of traditional segmentation-based spatial analysis methods.

Main Methods:

• CANVAS, a vision transformer, utilizes self-supervised masked image modeling on high-dimensional multiplexed images.
• The framework is segmentation-free, operating directly on pixel-level data.
• It preserves local morphology and biomarker distribution for detailed analysis.

Main Results:

• CANVAS effectively distinguishes subtle morphological differences and biomarker patterns.
• The framework precisely separates and characterizes distinct TME signatures.
• Application to a lung tumor dataset identified a novel monocytic TME signature.

Conclusions:

• CANVAS provides a powerful, segmentation-free approach for TME analysis.
• The identified monocytic signature is associated with poor prognosis in lung cancer.
• This framework facilitates the discovery of new TME subtypes and their clinical relevance.