Structuring pathology foundation models with domain knowledge.

Area of Science:

• Artificial intelligence
• Computational pathology
• Medical informatics

Background:

• Foundation models are increasingly used in medical research.
• Integrating structured knowledge into AI models can enhance performance.
• Pathology benchmarks often face challenges with rare disease classification.

Purpose of the Study:

• To introduce KEEP, a knowledge-guided vision-language foundation model.
• To leverage hierarchical disease knowledge for improved AI performance in pathology.
• To enhance zero-shot and few-shot learning capabilities for cancer classification.

Main Methods:

• Developed KEEP, a vision-language foundation model.
• Incorporated hierarchical disease knowledge using a structured disease graph during pre-training.
• Evaluated model performance on multiple pathology benchmarks.

Main Results:

• Knowledge-guided learning improved semantic representations.
• Achieved enhanced zero-shot and few-shot performance across pathology benchmarks.
• Demonstrated notable improvements in rare cancer classification.

Conclusions:

• KEEP effectively integrates hierarchical disease knowledge into foundation models.
• The model shows significant potential for advancing computational pathology and rare cancer diagnosis.
• Knowledge-guided learning is a promising approach for improving AI in medical applications.