Subspecialty-specific foundation model for intelligent gastrointestinal pathology

Area of Science:

• Computational pathology
• Gastrointestinal (GI) disease research
• Artificial intelligence in medicine

Background:

• Conventional histopathology for GI diseases lacks objectivity and reproducibility.
• Existing computational pathology models are broad but lack deep clinical utility in real-world scenarios.

Purpose of the Study:

• To develop Digepath, a disease-specialized foundation model for high-impact GI pathology.
• To improve the clinical utility and diagnostic accuracy of computational pathology in GI diseases.

Main Methods:

• Two-stage iterative optimization: pretraining on 353 million multi-scale patches from 210,043 H&E slides.
• Fine-tuning on 471,443 expert-annotated regions, balancing tumor and non-tumor samples.
• Integration into an agent-based clinical reasoning framework for end-to-end diagnostic workflows.

Main Results:

• Digepath achieved state-of-the-art performance on 32 out of 33 systematic downstream tasks in GI pathology.
• Demonstrated robust generalization across diverse GI pathology tasks, including diagnosis, molecular profiling, and survival prognosis.
• Successfully integrated into a clinical reasoning framework, supporting intelligent diagnostic workflows.

Conclusions:

• Digepath represents a significant advancement in disease-specialized foundation models for GI pathology.
• The model shows promise for real-world deployment, enhancing diagnostic capabilities and clinical workflows.
• Specialized models can overcome limitations of broad benchmarks, offering deeper clinical utility.