Pathology Foundation Models: Evolution, Current Landscape, Challenges and Opportunities from a Technical and Clinical...

Area of Science:

• Computational pathology
• Artificial intelligence in histopathology
• Foundation models in medicine

Background:

• Foundation models (PFMs) are revolutionizing computational pathology with scalable, task-agnostic representations of whole-slide images (WSIs).
• PFMs utilize self-supervised Vision Transformer architectures and large WSI datasets for broad generalization and few-shot learning.
• The field has evolved from earlier methods like CLAM and HIPT to large-scale models such as UNI, Virchow, Phikon, CONCH, GigaPath, H-Optimus, TITAN, and the Mayo Clinic Atlas.

Purpose of the Study:

• To review the development of PFMs in computational pathology.
• To critically evaluate the strengths and limitations of current PFMs.
• To outline priorities for the safe and effective clinical translation of PFMs.

Main Methods:

• Review of existing literature on pathology foundation models.
• Analysis of PFM performance across diagnostic and prognostic benchmarks.
• Evaluation of multimodal integration capabilities with genomics and clinical data.

Main Results:

• PFMs demonstrate impressive performance and enable multimodal data integration.
• Significant barriers include inconsistent institutional generalization, lagging interpretability, and slow workflow integration.
• Specific areas like cytopathology and rare tumor subtypes remain challenging for current models.

Conclusions:

• The next phase of PFM development necessitates rigorous benchmarking and pathologist-in-the-loop deployment.
• Multimodal fusion is crucial for advancing PFMs from research tools to clinically robust systems.
• Addressing limitations in generalization, interpretability, and workflow integration is key for clinical translation.