Train and evaluate histology models on STHELAR with a leave-one-tissue-out (LOTO) protocol. Demonstrate that a domain-generalization (DG) approach reduces the performance collapse on unseen organs compared to a baseline (CellViT/ViT/UNet).