Comparing natural image pretraining with digital pathology foundation models for whole slide image-based survival analysis

Due to their high information density, Whole Slide Images (WSIs) are considered invaluable tools for survival analysis. Many existing Multiple Instance Learning (MIL) frameworks for this task rely on a ResNet50 backbone pre-trained on natural images. In this work, we leverage recently introduced histopathology foundation models, such as UNI and Hibou, to improve the predictive prowess of existing MIL frameworks. Additionally, we demonstrate that deploying an ensemble of digital pathology foundation models yields higher accuracy, although the gains diminish when applied to more complex MIL architectures.