General Pathologists Achieve Near-Specialist Diagnostic Performance Using Deep Learning-Based Virtual Staining for Donor Kidney Assessment: A Retrospective-Prospective Diagnostic Concordance Study.

Conventional rapid evaluation of donor kidney quality primarily relies on hematoxylin and eosin (H&E) staining, which inadequately visualizes collagen fibers, compromising the assessment of interstitial fibrosis and posing challenges for general pathologists (GPs) lacking specialized renal training. This study investigated whether artificial intelligence-based virtual staining could enhance donor kidney evaluation, particularly for interstitial fibrosis and chronic pathologies. Using 187 paired whole-slide images of H&E and Masson's trichrome-stained sections, we developed and validated a CycleGAN-based virtual staining model that transforms H&E images into virtual Masson's trichrome (VMT) representations. A renal pathologist (RP) and 2 GPs (GP.1 and GP.2) evaluated interstitial fibrosis and chronic changes using the Remuzzi scoring system, comparing results with and without virtual staining. Prospective validation was performed on 46 frozen sections. Results demonstrated that VMT effectively visualized interstitial collagen fibers, enabling more reliable fibrosis classification. Diagnostic accuracy significantly improved with virtual staining versus H&E alone (weighted kappa: GPs improved from 0.39-0.51 to 0.78-0.82; RP from 0.61 to 0.86), and interobserver agreement markedly increased (GPs: 55%-64% to 84%-88%; RP: 71%-90%). Notably, VMT bridged the expertise gap, allowing GPs to achieve near-specialist performance in fibrosis assessment (weighted kappa: GP.1 from 0.48 to 0.82; GP.2 from 0.41 to 0.84). Prospective validation confirmed these advantages, showing superior accuracy when combining VMT with H&E versus H&E alone (weighted kappa: GP.1 from 0.27 to 0.65; GP.2 from 0.34 to 0.68). The technology also enhanced glomerulosclerosis detection (GPs' kappa = 0.82-0.87) and transplant decision-making accuracy (kappa = 0.84-0.85), elevating GPs to specialist-level competence. In conclusion, deep learning-based virtual staining significantly improves the precision of donor kidney evaluations by GPs, approaching specialist-level performance. This technology offers an efficient, cost-effective solution for assessing fibrosis and chronic pathologies, potentially eliminating diagnostic disparities between GPs and specialist pathologists in transplantation medicine.