Representation learning enables robust single cell phenotyping in whole slide liquid biopsy imaging

Area of Science:

• Biomedical Imaging
• Computational Biology
• Oncology

Background:

• Liquid biopsies offer promising biomarkers for cancer management, but challenges remain in identifying rare, heterogeneous, and plastic tumor-associated cells.
• Current analysis methods for circulating cells often rely on subjective manual review or engineered features, leading to variability and bias.

Purpose of the Study:

• To develop a deep contrastive learning framework for robust feature extraction from whole slide immunofluorescence microscopy images of circulating cells.
• To enable accurate identification, stratification, and enumeration of single circulating cells, including rare phenotypes, for improved cancer biomarker analysis.

Main Methods:

• A deep contrastive learning framework was applied to whole slide immunofluorescence microscopy images.
• The model extracted features for robust identification and stratification of single circulating cells.
• Performance was evaluated on cell phenotype classification, outlier detection, clustering, and automated enumeration of rare cell types.

Main Results:

• The learned features achieved 92.64% accuracy in classifying diverse cell phenotypes.
• Automated identification and enumeration of rare phenotypes reached an average F1-score of 0.93 on contrived samples and 0.858 on clinical circulating tumor cell phenotypes.
• The framework demonstrated improved performance in downstream tasks like outlier detection and clustering.

Conclusions:

• The deep contrastive learning framework provides a scalable and reproducible solution for analyzing tumor-associated cellular biomarkers in liquid biopsies.
• This approach has strong potential to enhance clinical prognosis and guide personalized cancer treatment strategies by accurately characterizing circulating cells.