Label-free evaluation of transplant tissue samples via deep learning–based virtual staining

Histochemical analysis of biopsies is the diagnostic gold-standard for rejection and requires histochemical staining of adjacent tissue sections to produce H&E, MT, and EVG-stained slides, which is a laborious process prone to artifacts and batch-to-batch variability. Here, we introduce a deep learning-enabled virtual staining platform that digitally generates H&E, MT, and EVG-stained images from label-free autofluorescence images. Specialized neural networks produce virtual stains that match their histochemical counterparts, accurately depicting histological biomarkers of transplant rejection. In a blinded evaluation of specimens from ~60 transplant recipients, virtual staining yielded diagnostic concordance rates of 82.4% for lung and 91.7% for heart rejection assessments compared with conventional staining methods. Quantitative evaluation of nuclear, cytoplasmic, and extracellular staining quality confirmed the non-inferiority of virtual staining results. By delivering multi-stain outputs on the same tissue cross-section, virtual staining reduces turnaround time, reagent consumption, and tissue usage—delivering a reproducible, cost-effective solution for transplant pathology.