Robustness of computational pathology foundation models under distribution shifts

Determine whether computational pathology foundation models trained on large and varied datasets (specifically UNI and CONCH) are robust to commonly encountered distribution shifts in real-world histopathology whole-slide images, or whether their downstream performance can still break down in practical deployment settings.

Background

Foundation models in computational pathology are proposed as general-purpose feature extractors that should perform well across diverse downstream tasks. However, whole-slide image data exhibits substantial variability due to site-specific staining and scanning procedures, creating real-world distribution shifts.

While deep learning models are known to be sensitive to distribution shifts in broader machine learning contexts, it has not been established whether training pathology foundation models on large and varied datasets ensures robustness to such shifts. This paper evaluates UNI and CONCH as frozen feature extractors within prostate cancer ISUP grade classification models to probe this question under image and label distribution shifts.