Benchmarking pathology foundation models: Adaptation strategies and scenarios

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology image analysis

Background:

• Foundation models show promise for pathology image analysis but face adaptation challenges with diverse data and limited availability.
• Adapting these models across different data sources, acquisition conditions, and downstream tasks requires robust benchmarking.

Purpose of the Study:

• To benchmark four pathology-specific foundation models across 20 datasets.
• To assess model performance in consistency and flexibility scenarios under varying data conditions.
• To provide guidance for deploying foundation models in clinical pathology settings.

Main Methods:

• Benchmarking four foundation models on 20 datasets using consistency and flexibility assessment scenarios.
• Evaluating five fine-tuning methods for adaptation to diverse datasets within classification tasks.
• Assessing five few-shot learning methods for performance in data-limited environments for slide-level survival prediction.

Main Results:

• Parameter-efficient fine-tuning demonstrated efficiency and effectiveness for adapting models to diverse datasets in classification tasks.
• Foundation model performance in survival prediction was influenced by feature aggregation and data characteristics.
• Few-shot learning methods modifying only during testing phase showed greater benefit for foundation models in data-limited scenarios.

Conclusions:

• Parameter-efficient fine-tuning is a viable strategy for adapting pathology foundation models.
• Feature aggregation and data characteristics are critical for survival prediction tasks.
• Specific few-shot learning approaches enhance foundation model utility in low-data settings, guiding clinical deployment.