VI Seminar #82: Foundation Models for Pathology: Capable and Scalable, but still Fragile.

Presented by Vilde Schulerud Bøe, PhD Candidate at the University of Oslo / SFI Visual Intelligence

Abstract

Deep learning has transformed computational pathology, enabling highly accurate models for tasks such as survival analysis and tumor detection. However, despite impressive progress, these models often struggle to maintain performance across domain shifts arising from variations in staining, scanners, institutions, and patient populations. This highlights a challenge of robustness and generalization.

In this talk, I will trace the evolution of foundation models (FMs) in computational pathology, from early approaches to today’s increasingly large-scale vision encoders. I will discuss how these models differ in design and behavior, revealing both strengths and a persistent tendency to fall short of being truly foundational in practice. I will then present our recent work exploring how lightweight adaptation techniques such as LoRA offer a practical path for model tuning, improving the robustness of FM features. Finally, I will discuss the future of computational pathology FMs and potential paths towards models that are not only accurate but reliably deployable in real-world clinical settings.