Evaluating Vision and Pathology Foundation Models for Computational Pathology: A Comprehensive Benchmark Study.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Precision diagnostics

Background:

• Advancing precision medicine requires robust AI foundation models for pathology.
• Current AI models' performance and generalizability across diverse histopathology data are under-examined.

Purpose of the Study:

• To benchmark 31 AI foundation models for computational pathology.
• To evaluate model performance across diverse datasets and tasks.

Main Methods:

• Benchmarked 31 AI foundation models: vision models (VM), vision-language models (VLM), pathology-specific VM (Path-VM), and pathology-specific VLM (Path-VLM).
• Evaluated models on 41 tasks from TCGA, CPTAC, external, and out-of-domain datasets.
• Assessed performance based on disease detection, classification, and prognostic insights.

Main Results:

• Virchow2, a pathology foundation model, achieved the highest performance across multiple datasets.
• Path-VM models outperformed Path-VLM and general VM, ranking highly across tasks.
• Model and data size did not consistently correlate with improved performance.

Conclusions:

• Virchow2 demonstrates high effectiveness in diverse histopathological evaluations.
• Path-VM models show strong potential, but further research is needed to enhance generalizability.
• Fusion models integrating top performers offer superior generalization across diverse tissues and tasks.