Automating Ground Truth Annotations for Gland Segmentation Through Immunohistochemistry.

Area of Science:

• Computational pathology
• Digital histopathology
• Machine learning in medicine

Background:

• Microscopic evaluation of colon glands is crucial for diagnosing inflammatory bowel disease and cancer.
• Deep learning models offer systematic, reproducible, and quantitative assessment of glandular tissue architecture.
• Manual annotation of histopathology slides for deep learning is time-consuming and expensive.

Purpose of the Study:

• To develop an automated method for generating ground truth annotations of colon glands in H&E-stained slides.
• To utilize immunohistochemistry (IHC) labels for transferring gland masks to H&E images for deep learning model training.
• To assess the performance of automated annotations compared to manual annotations and improve model generalizability.

Main Methods:

• Developed an image processing pipeline to transfer gland masks from KRT8/18, CDX2, or EPCAM IHC to coregistered H&E images.
• Trained deep learning models using automated IHC-derived annotations.
• Compared model performance against manual annotations on internal and public datasets.
• Proposed a data sampling technique for adapting models to new data sources.

Main Results:

• EPCAM IHC provided gland outlines closely matching manual annotations (Dice = 0.89) and were resilient to inflammation.
• Models trained with 10% of annotated cases achieved high performance on public datasets (Dice scores of 0.902 and 0.89).
• Automated annotations using cell type-specific IHC markers demonstrated comparable performance to manual annotations.

Conclusions:

• Automated gland annotation using IHC labels can effectively replace manual annotations in digital pathology.
• The proposed method facilitates the development of robust deep learning models for colon cancer and inflammatory bowel disease diagnosis.
• A simple data sampling technique enhances model adaptability across diverse datasets, improving generalization.