Beyond the Failures: Rethinking Foundation Models in Pathology

TL;DR

This paper argues that foundation models underperform in pathology due to fundamental mismatches with biological data, advocating for models specifically designed for tissue analysis.

Contribution

It highlights the limitations of current foundation models in pathology and emphasizes the need for domain-specific model architectures tailored to biological images.

Findings

Current foundation models show low accuracy and instability in pathology.

Existing models inherit flaws from natural-image assumptions that do not hold for tissue.

Pathology requires models explicitly designed for biological tissue analysis.

Abstract

Despite their successes in vision and language, foundation models have stumbled in pathology, revealing low accuracy, instability, and heavy computational demands. These shortcomings stem not from tuning problems but from deeper conceptual mismatches: dense embeddings cannot represent the combinatorial richness of tissue, and current architectures inherit flaws in self-supervision, patch design, and noise-fragile pretraining. Biological complexity and limited domain innovation further widen the gap. The evidence is clear-pathology requires models explicitly designed for biological images rather than adaptations of large-scale natural-image methods whose assumptions do not hold for tissue.