Development and Applications of Pathology Foundation Models in Cancer Diagnosis

Abstract

Pathology foundation models are a class of deep neural networks pre-trained on massive pathology image data. These models typically obtain generic representations through self-supervised learning to support a wide range of cancer diagnostic tasks. In recent years, such models have demonstrated significant performance improvements in pan-cancer detection, rare cancer identification, cancer subtype classification, and molecular prediction. This paper focuses on the cancer diagnosis scenario, systematically summarizes the structural design and data resources of representative pathology foundation models, and emphasizes the discussion of patch-level Transformer, graph-based encoders, and visual-language models based on image-text pairing, and analyzes their roles in clinical-related tasks. Furthermore, this paper points out the main challenges currently faced in this field in terms of long-tail generalization, evaluation bias, engineering reproduction threshold, and the reliability of multimodal output, and looks forward to the future development direction of pathology foundation models for reliable clinical deployment.