This research introduces a novel approach for creating universal embeddings in digital pathology using multimodal large language models, enabling efficient diagnosis across diverse pathologies without task-specific training.

• Addresses the sustainability challenge of traditional pathology models that require extensive labeled datasets
• Develops universal multimodal embeddings capable of supporting multiple downstream diagnostic tasks
• Demonstrates how MLLM-based pathology embeddings can reduce dependence on specialized training for each disease type
• Provides a scalable framework for pathology AI that can adapt to diverse tissue types and conditions

This advancement matters for healthcare by significantly reducing the time and resources needed for pathology AI development, potentially accelerating diagnosis while maintaining accuracy across a spectrum of diseases.

MLLM4PUE: Toward Universal Embeddings in Digital Pathology through Multimodal LLMs