Extra seminar: Transforming histopathology using generative AI-based virtual staining – challenges and opportunities – Leena Latonen

Tissue morphology is investigated through microscopy of histologically stained tissue. On their own, tissue sections are nearly invisible to the naked eye, but histological staining makes the structures and morphology visible through chemical reactions. These chemical reactions are often irreversible and, thus, a separate sample is needed for each staining whenever several stainings are required for a specific diagnosis. This is laborious, sample- and chemical-consuming, and renders single-cell level comparisons impossible, as each staining represents different cells originating from different sections.

Virtual staining has emerged as a potential option for chemical staining in histology through utilization of AI-based image-to-image translation. Particularly, using regular brightfield microscopy for imaging unstained tissue sections could enable widely adoptable, scalable, and cost-efficient workflows. We develop deep learning-based approaches for virtual hematoxylin and eosin (HE) and other stainings using unstained brightfield images and, for example, generative adversarial network (GAN) and diffusion model -based approaches. Our results show that virtual staining of unstained tissue is feasible and enables histological assessment of multiple tissue types, stainings, and even clinically relevant features. Virtual staining possesses enormous potential to save laboratory work, consumption, and costs, in addition to enabling faster cancer diagnostics for the benefit of patients in the future.

About Leena Latonen