Automated cell annotation and classification on histopathology for spatial biomarker discovery

Area of Science:

• Computational pathology
• Biomarker discovery
• Digital pathology

Background:

• Histopathology with hematoxylin and eosin (H&E) staining is crucial for clinical diagnosis.
• Single-cell analysis of histopathology offers insights into disease and treatment response.
• Current methods rely on inefficient and error-prone human annotations.

Purpose of the Study:

• To develop an automated approach for cell annotation and classification on H&E stained images.
• To create a high-quality annotated dataset for training deep learning models.
• To investigate the link between immune cell spatial interactions and patient outcomes.

Main Methods:

• Multiplexed immunofluorescence (mIF) was used to define cell types based on protein markers.
• H&E images were co-registered with mIF data at the single-cell level.
• A deep learning model combining self-supervised learning and domain adaptation was trained for cell classification.

Main Results:

• A dataset of 1,127,252 cells was annotated using mIF.
• The deep learning model achieved 86%-89% accuracy in classifying four cell types on H&E images.
• The cell classification model demonstrated applicability to whole slide images.
• Spatial interactions of immune cells in the tumor microenvironment were linked to patient survival and response to immune checkpoint inhibitors.

Conclusions:

• The developed approach provides a scalable method for single-cell analysis of standard histopathology.
• This technique can facilitate the discovery of novel spatial biomarkers for precision oncology.
• Automated cell annotation overcomes limitations of manual annotation in digital pathology.