Class visualizations and activation atlases for computational pathology

The clinical promise of computational pathology increasingly depends on foundation model-based pipelines, yet the morphological concepts encoded by these systems remain poorly understood. We address this gap with a concept-level visualization framework using class visualizations (CVs) and activation atlases (AAs) in a pathology foundation-model setting across colorectal tissue and multi-organ cancer tasks. Four pathologists annotate hematoxylin and eosin-stained image patches as well as generated visualizations, complemented by attribution- and similarity-based metrics. CVs retain class-associated morphology for distinct tissue classes, with reduced separability and higher annotator variability in morphologically overlapping cancer classes. AAs expose layer-dependent organization of encoded concepts, with coherent regions for coarse tissue and cancer groupings and increasing dispersion and overlap at finer label granularities. Together, these findings show that concept-level visualization makes representations learned by pathology foundation models inspectable across diagnostic difficulty levels, with feature separation decreasing and ambiguity increasing as morphological complexity and expert disagreement rise.