Automating liver biopsy segmentation with a robust, open-source tool for pathology research: the HOTSPoT model

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Current artificial intelligence (AI) tools for liver pathology have limited applications and lack external validation.
• Automated analysis of whole slide images (WSIs) in liver pathology is crucial for improving diagnostic accuracy and efficiency.

Purpose of the Study:

• To introduce and validate HOTSPoT, an open-source, transformer-based AI model for automated segmentation of portal tracts in H&E-stained liver biopsy WSIs.
• To assess the performance and generalizability of HOTSPoT across a multi-institutional dataset.

Main Methods:

• Development of HOTSPoT, a transformer-based model for semantic segmentation.
• Training and validation on a multi-institutional dataset of 223 H&E-stained liver biopsy WSIs annotated by expert hepatopathologists.
• Evaluation of segmentation performance using Dice scores and Intersection over Union (IoU), and assessment of domain shift.

Main Results:

• HOTSPoT achieved high segmentation performance with mean Dice scores of 0.92 (train/val) and 0.91 (test), and mean IoUs of 0.86, 0.85, and 0.84.
• The model demonstrated minimal domain shift across different institutions.
• Automated portal tract quantification showed strong concordance with manual assessments (κ up to 0.90), and portal area correlated significantly with fibrosis stage (r=0.87, p<0.001).

Conclusions:

• HOTSPoT is a validated, open-source AI tool for accurate portal tract segmentation in liver pathology.
• The model facilitates automated quantification and shows potential for objective assessment of liver fibrosis.
• HOTSPoT's availability and integration capabilities (TorchScript, WSInfer, QuPath) promote its adoption in digital pathology workflows.