Structuring pathology foundation models with domain knowledge

Researchers have developed KEEP, a novel vision-language foundation model for computational pathology that integrates hierarchical disease knowledge into its pre-training process through a structured disease graph. Published in Cancer Cell, the study by Zhou et al. demonstrates that this knowledge-guided approach shapes more meaningful semantic representations, leading to significantly improved zero-shot and few-shot performance across multiple pathology benchmarks. The model shows particularly promising results in the classification of rare cancers, where training data is typically scarce and conventional AI models struggle. This work represents an important step toward more generalizable and clinically useful AI tools in pathology, though further clinical validation will be needed before deployment in diagnostic workflows.