Computational Pathology

Computational pathology turns digitised tissue slides into quantitative, reproducible evidence. Whole-slide images are gigapixel-sized, labels are scarce and expensive, and staining and scanners differ from lab to lab — exactly the conditions under which our methods for weakly supervised, robust and federated learning are needed.

Our work spans the full pipeline. With Roche we developed virtual staining that translates one stain into another with generative models, and with partners worldwide we organised the LYSTO benchmark and hackathon for lymphocyte assessment in immunohistochemistry. In ophthalmic pathology, INTERACT-CMIL grades conjunctival melanocytic lesions with interpretable multi-task learning. In MISTAI we fine-tune pathology foundation models for muscle-invasive bladder cancer together with Dnipro State Medical University, and earlier we contributed federated learning to the European BigPicture consortium for a pan-European pathology image platform.

Across these projects we ask how foundation models can be adapted with little labelled data, how models stay reliable across sites and stains, and how their predictions can be explained to pathologists.