+ abstract = {Whole-slide imaging has transformed histopathology into a data-intensive field, requiring robust and generalisable computational tools. Foundation models offer a promising approach for a range of downstream tasks with minimal labelled data. While recent work has shown their effectiveness for slide-level classification and retrieval, their potential for dense prediction tasks such as image segmentation remains underexplored. In this study, we present a comprehensive benchmark of 15 pathology-specific foundation models for histopathological image segmentation, evaluated across two distinct modalities: H&E-stained histology and Annexin A5-stained immunohistochemistry. To ensure a fair and architecture-neutral comparison, we freeze each foundation models encoder and pair it with a shared lightweight decoder, disentangling representation quality from model size. Results show that foundation model encoders can sometimes lead to strong segmentation performance without fine-tuning, but effectiveness varies significantly by model and modality. Our findings reveal that compact encoders can often outperform larger, more recent models, underscoring that model size and classification accuracy are poor predictors of segmentation capabilities.},
0 commit comments