LAMP is a pretraining framework that aggregates multi-scale information from pathology WSIs using a hierarchical transformer architecture with learnable cross-magnification attention mechanisms, guided by language supervision from pathology reports.
Key findings
LAMP introduces a novel architecture for language-guided mixed-magnification aggregation.
The framework dynamically weights features across scales based on semantic guidance from pathology reports.
LAMP enables adaptive magnification selection and region-level representation learning without pixel-level annotations.
The model is evaluated on downstream tasks including cancer subtyping, biomarker prediction, and survival analysis.
Limitations & open questions
The framework's performance in real-world clinical settings remains to be validated.
The model's ability to generalize across different types of pathology reports and conditions needs further investigation.