Why Foundation Models in Pathology Are Failing (and What Comes Next)

The article says pathology foundation models are failing for a conceptual reason, not just an engineering one. Scaling up model size and data has not produced reliable gains on the clinical tasks that matter most, especially across institutions and scanners.

It argues that self-supervised pretraining learns pixel statistics more than diagnostic structure, and that models trained on one hospital often lose performance and stay overconfident when moved elsewhere. The text points to weakly supervised learning, task-specific models, geometric-aware architectures, and hybrid feature-extractor setups as better fits for pathology. It also insists that clinical deployment needs interpretability and real-world validation, not benchmark performance alone.

Reading notes

#

• Foundation models in pathology hit a performance ceiling on clinically important tasks, and larger parameter counts do not reliably improve downstream results.
• Self-supervised objectives on histology slides mainly teach pixel reconstruction, while pathology needs structural information such as glandular relationships, nuclei morphology, and tissue architecture.
• Models trained at one institution often lose AUC when tested at another, while still reporting high confidence, which creates a domain miscalibration and safety problem.
• Architectural complexity and overfitting can increase vulnerability to small perturbations, which matters in settings with stain drift, scanner variation, and equipment degradation.
• The field was pulled toward scaling because that formula worked in NLP and general vision, and because academic, commercial, and regulatory incentives favored foundation models.
• No pathology foundation models had FDA approval as of early 2024, and the article says none had randomized controlled trial evidence or proven patient outcome benefits.
• Weakly supervised multiple instance learning achieved strong detection results, including 0.99 AUC on prostate cancer detection, and generalizes better than fully supervised models trained on small curated datasets.
• Geometric-aware approaches such as spherical convolutions and persistent homology are presented as alternatives that encode tissue structure more directly.
• Hybrid FM+MIL systems can use foundation models as feature extractors and let task-specific aggregation handle diagnosis.
• The article recommends stopping the assumption that scale equals capability, making interpretability mandatory, validating in real clinical conditions, and investing in domain-specific innovation.
• Foundation models are still described as useful for feature extraction, data synthesis, and downstream ensembles, but not as universal end-to-end diagnostic systems.