Kidney Allograft Rejection

This study developed a deep learning system using multi-instance learning to automatically detect and subtype kidney transplant rejection from H&E-stained whole-slide biopsy images. Trained on 906 slides from 302 biopsies, the model achieved strong performance in three-category rejection classification (AUC 0.798), outperforming transplant pathologists under routine assessment conditions. Additionally, prognostic models accurately predicted 1-year graft loss (AUC 0.936) and treatment response (AUC 0.756), demonstrating potential to support both diagnosis and clinical decision-making.