Semantic Segmentation of Intralobular and Extralobular Tissue from Liver Scaffold H&E Images

Area of Science:

• Biomedical Engineering
• Computational Pathology
• Tissue Engineering

Background:

• Decellularized tissues are vital for tissue engineering, but their quality assessment relies on subjective analysis of H&E stained images.
• Automated segmentation of whole slide images into background, intralobular, and extralobular areas is essential for objective quality evaluation of liver scaffolds.
• Current semi-automatic segmentation methods lack the required accuracy for this specialized task.

Discussion:

• A novel two-stage segmentation method was developed to address the limitations of existing techniques.
• The first stage utilizes hand-crafted pixel descriptors for initial classification on partially annotated data.
• The second stage employs a U-Net inspired Convolutional Neural Network (CNN) for refined segmentation, trained on the outputs of the first stage.

Key Insights:

• The proposed two-stage method effectively segments liver scaffolds despite limited training data.
• The CNN-based second stage, leveraging insights from the initial classification, significantly improves segmentation accuracy.
• The system achieved a high recognition accuracy of 90.70% with an optimized training setup.

Outlook:

• This automated segmentation tool has the potential to standardize and improve the quality assessment of decellularized liver scaffolds.
• Further development could enable objective texture analysis within the intralobular regions, critical for recellularization.
• The methodology may be adaptable to other tissue engineering applications requiring automated image segmentation.