Overview

Recent developments in artificial intelligence (AI) are positioning computational pathology as a vital component in precision oncology. Traditional methods have relied on task-specific models, necessitating separate models for each unique task. This approach often requires extensive annotated datasets, leading to increased costs and lengthy development times, while also struggling with adaptability across different tasks.

Challenges with Task-Specific Models

• High dependency on large-scale annotated datasets.
• Increased costs and prolonged development cycles.
• Poor adaptability to real-world clinical needs.
• Inability to effectively address rare diseases and limited sample scenarios.

Introduction of Foundation Models

The emergence of foundation models (FMs) signifies a shift towards more generalized intelligence in computational pathology. These models are trained using self-supervised learning on extensive, multimodal pathology datasets, enhancing their representation learning and generalization capabilities.

Research Insights

A recent review led by Professor S. Kevin Zhou and Dr. Rui Yan from the University of Science and Technology of China, along with Dr. Fei Ren from the Institute of Computing Technology, China, explores the advancements and applications of pathological FMs in oncology. The study highlights key challenges and opportunities in this evolving field, published online on September 25, 2025, in the Chinese Medical Journal.

Types of Pathology Foundation Models

Current research on pathology foundation models can be categorized into three main types:

1. Pathology Image Foundation Models: These models extract essential visual features from whole slide images (WSIs) for tasks like cancer classification and biomarker prediction. Examples include GigaPath, UNI, and Virchow, which outperform traditional methods across various cancer types.
2. Pathology Image-Text Foundation Models: Integrating natural language processing, these models are pre-trained on both pathology images and textual reports, facilitating tasks such as diagnostic report generation and educational support. Notable models include PLIP, CONCH, and PathChat, which enable zero-shot learning for previously unseen cases.
3. Pathology Image-Gene Foundation Models: These models align pathology images with multi-omics data, enhancing tumor subtype classification and treatment response predictions. Examples like mSTAR, GiMP, and TANGLE provide insights into cancer heterogeneity and molecular mechanisms.

Clinical Implications

Foundation models significantly improve the efficiency and accuracy of pathological analyses while reducing reliance on specialized techniques like immunohistochemistry (IHC) and genomic testing. This advancement accelerates diagnostic processes, lowers healthcare costs, and enhances patient treatment experiences.

Challenges Ahead

Despite their potential, pathology foundation models face challenges in clinical deployment:

• Lack of extensive validation on multi-center, real-world datasets.
• “Black-box” nature hindering clinical trust and interpretability.
• Issues with feature redundancy and data heterogeneity in multimodal tasks.

Future Directions

Future research should focus on:

• Long-sequence modeling and high-dimensional feature fusion.
• Development of ethical guidelines.
• Intelligent cross-modal collaboration.

Foundation models are set to reshape computational pathology, laying the groundwork for intelligent, automated, and personalized decision-support systems in pathology. Their ongoing evolution is expected to significantly impact precision oncology and life science research.

Reference: Wang Y, Gu Y, Zhang X, Wang B, Wang R, Li X, Liu Y, Qu F, Ren F, Yan R, Zhou SK. Computational pathology in precision oncology: Evolution from task-specific models to foundation models. Chin Med J (Engl). 2025 Nov 20;138(22):2868-2878. doi: 10.1097/CM9.0000000000003790