Deep learning identifies pathological abnormalities predictive of graft loss in kidney transplant biopsies

BackgroundInterstitial fibrosis, tubular atrophy, and inflammation are major contributors to renal allograft failure. Here we seek an objective, quantitative pathological assessment of these lesions to improve predictive utility. MethodsWe constructed a deep-learning-based pipeline recognizing normal vs. abnormal kidney tissue compartments and mononuclear leukocyte (MNL) infiltrates from Periodic acid-Schiff (PAS) stained slides of transplant biopsies (training: n=60, testing: n=33) that quantified pathological lesions specific for interstitium, tubules and MNL infiltration. The pipeline was applied to 789 whole slide images (WSI) from baseline (n=478, pre-implantation) and 12-month post-transplant (n=311) protocol biopsies in two independent cohorts (GoCAR: 404 patients, AUSCAD: 212 patients) of transplant recipients to correlate composite lesion features with graft loss. ResultsOur model accurately recognized kidney tissue compartments and MNLs. The digital features significantly correlated with Banff scores, but were more sensitive to subtle pathological changes below the thresholds in Banff scores. The Interstitial and Tubular Abnormality Score (ITAS) in baseline samples was highly predictive of 1-year graft loss (p=2.8e-05), while a Composite Damage Score (CDS) in 12-month post-transplant protocol biopsies predicted later graft loss (p=7.3e-05). ITAS and CDS outperformed Banff scores or clinical predictors with superior graft loss prediction accuracy. High/intermediate risk groups stratified by ITAS or CDS also demonstrated significantly higher incidence of eGFR decline and subsequent graft damage. ConclusionsThis deep-learning approach accurately detected and quantified pathological lesions from baseline or post-transplant biopsies, and demonstrated superior ability for prediction of posttransplant graft loss with potential application as a prevention, risk stratification or monitoring tool.