Learning to detect lymphocytes in immunohistochemistry with deep learning.

Area of Science:

• Oncology
• Immunology
• Computational Pathology

Background:

• Immune system evasion is a hallmark of cancer.
• Accurate quantification of immune cells, specifically lymphocytes like CD3+ and CD8+ cells, is crucial for understanding anti-tumor immune responses.
• Histopathology image analysis for immune cell detection is complex and benefits from automated solutions.

Purpose of the Study:

• To develop and evaluate deep learning algorithms for automatic detection of CD3+ and CD8+ lymphocytes in cancer histopathology images.
• To assess the performance of different deep learning methods across various tissue subcompartments (normal, immune clusters, artifacts).
• To compare the accuracy of automated lymphocyte detection with manual quantification by pathologists.

Main Methods:

• A dataset of 171,166 manually annotated CD3+ and CD8+ cells was created.
• Four deep learning methods were trained and tested on histopathology slides from breast, colon, and prostate cancers.
• Performance was evaluated across different image regions and compared against a pathologist observer study.

Main Results:

• The U-Net deep learning model achieved the highest performance with an F1-score of 0.78.
• U-Net demonstrated the highest agreement with manual evaluation (κ=0.72), surpassing the average pathologist agreement (κ=0.64).
• Performance varied depending on tissue subcompartments, highlighting the need for robust algorithms.

Conclusions:

• Deep learning, particularly U-Net, offers a promising automated approach for lymphocyte quantification in cancer histopathology.
• Automated detection shows potential for higher consistency and accuracy compared to manual assessment by pathologists.
• The developed methods and dataset are publicly available to facilitate further research and clinical application.