Foundation Models for Histopathology-Fanfare or Flair.

Area of Science:

• Histopathology
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Foundation models (FMs) show promise in various fields, but their performance in histopathology requires thorough assessment.
• General-purpose FMs, fine-tuned on internet data, are compared against specialized models trained on curated datasets.

Purpose of the Study:

• To evaluate the performance of current foundation models in histopathology.
• To compare general-purpose FMs (CLIP derivatives like PLIP, BiomedCLIP) with domain-specific histology models.

Main Methods:

• Models were evaluated on 8 datasets (4 internal Mayo Clinic, 4 public: PANDA, BRACS, CAMELYON16, DigestPath).
• Evaluation metrics included accuracy and macro-averaged F1 score (MV@5) at whole slide image and patch levels.
• Classification tasks were used for all model evaluations.

Main Results:

• Domain-specific models (DinoSSLPath, KimiaNet) outperformed general FMs in most tasks.
• DinoSSLPath excelled in internal colorectal cancer and liver image retrieval (MV@5 F1: 63%, 74%).
• KimiaNet led in breast and skin cancer tasks and performed well on CAMELYON16 (75%).

Conclusions:

• Specialized, domain-specific models demonstrate superior performance in histopathology.
• High-quality, multi-modal medical datasets with expert-validated alignment are essential for advancing biomedical vision-language FMs.
• Collaborative efforts in data curation are vital for clinical translation.