Structured crowdsourcing enables convolutional segmentation of histology images

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Deep learning models require extensive annotated datasets for high performance in semantic image segmentation.
• Histology image annotation is labor-intensive, requiring expert knowledge and posing challenges for data sharing.

Purpose of the Study:

• To address the need for large annotated datasets in histology image analysis.
• To evaluate inter-participant variability in tissue annotation.
• To develop accurate deep learning models for histology image segmentation and classification.

Main Methods:

• Recruited 25 participants (pathologists to medical students) to annotate 151 breast cancer whole-slide images.
• Systematically evaluated inter-participant discordance across different tissue types.
• Generated over 20,000 annotated tissue regions using expert feedback.
• Trained fully convolutional networks (FCNs) on the curated dataset.

Main Results:

• Low inter-participant discordance was observed for tumor and stroma regions.
• Higher discordance was noted for subjective or rare tissue classes.
• Trained FCNs achieved high accuracy in segmentation (mean AUC=0.945).
• The large-scale annotation dataset significantly improved image classification accuracy.

Conclusions:

• A large, curated dataset of annotated histology images can be generated through systematic evaluation and expert feedback.
• Deep learning models trained on this dataset demonstrate high accuracy for semantic segmentation and classification of histology images.
• The dataset is publicly available, facilitating further research in computational pathology.