Pathology Foundation Models

Pathology foundation models learn visual patterns from digital tissue slides, often using self-supervised learning on image patches and whole-slide images, or aligning images with text and molecular data. Researchers adapt these representations with frozen-feature probes, fine-tuning, and multiple-instance learning for tasks such as tumor classification, cell segmentation, biomarker prediction, and survival analysis. Evaluation spans benchmarks such as HEST and TCGA, external hospital cohorts, and increasingly prospective studies and pathologist reader evaluations. Models can transfer across many tasks, but performance may shift with tissue type, staining, or institution, while biological interpretability and the trade-off between model size, efficiency, and clinical reliability remain open challenges.