CLEAR-IT, a framework for contrastive learning to capture the immune composition of tumor microenvironments

Area of Science:

• Computational pathology
• Cancer biology
• Machine learning for medical imaging

Background:

• Accurate cell phenotyping in the tumor microenvironment is crucial for cancer research.
• Current methods often require precise cell segmentation, limiting scalability.

Purpose of the Study:

• Introduce Contrastive Learning Enabled Accurate Registration of Immune and Tumor cells (CLEAR-IT), a self-supervised framework for cell-level feature learning.
• Develop a label-efficient and segmentation-light approach for scalable cell phenotyping.

Main Methods:

• CLEAR-IT utilizes self-supervised learning on multiplexed images, focusing on cell locations to learn features.
• The framework employs contrastive learning for robust feature extraction.
• Evaluated performance using linear evaluation, hyperparameter optimization, and cross-modality testing.

Main Results:

• CLEAR-IT encoders demonstrate strong performance, improving with optimization and maintaining accuracy with up to 90% fewer labels.
• CLEAR-IT features enhance state-of-the-art classifiers, achieving comparable accuracy with less labeled data.
• Prognostic modeling using CLEAR-IT identified survival-associated tissue features generalizing across cohorts and modalities.

Conclusions:

• CLEAR-IT offers a scalable, label-efficient solution for cell phenotyping in digital pathology.
• The framework significantly reduces reliance on extensive cell segmentation and labeling.
• CLEAR-IT enhances existing workflows for tumor microenvironment analysis and prognostic prediction.