Immunocto: A massive immune cell database auto-generated for histopathology.

Area of Science:

• Computational pathology
• Cancer immunology
• Bioinformatics

Background:

• Studying the tumor immune microenvironment (TIME) is vital for cancer prognosis and immunotherapy response.
• Traditional methods for TIME analysis are labor-intensive and require specialized expertise.
• Automated analysis of histopathology images offers a scalable solution for TIME characterization.

Purpose of the Study:

• To develop an automated workflow for generating single-cell contours and labels from H&E and multiplexed immunofluorescence (IF) stained tissue sections.
• To create Immunocto, a large-scale, publicly available database of human cells for TIME research.
• To demonstrate the utility of Immunocto in training deep learning models for immune cell detection.

Main Methods:

• Utilized the Segment Anything Model (SAM) for automated cell segmentation and labeling.
• Combined H&E stained slides with multiplexed IF markers for dually stained tissue sections.
• Generated a database of over 6.8 million cells, including 2.2 million immune cells across four subtypes.

Main Results:

• Created Immunocto, a massive database containing 6,848,454 cells, with 2,282,818 immune cells (CD4+, CD8+, CD20+, macrophages).
• Each cell entry includes a 64x64 pixel H&E image, nucleus mask, and label.
• Deep learning models trained on Immunocto achieved state-of-the-art performance in lymphocyte detection.

Conclusions:

• The developed workflow enables robust, automated generation of single-cell data from routine histopathology slides.
• Immunocto provides a valuable resource for computational pathology, facilitating TIME studies and model training.
• The approach highlights the benefits of integrating matched H&E and IF data for advancing cancer research.