Superior transplant recipient outcome prediction and pathology assessment using rapid deep learning applied to procurement kidney biopsies.

Evaluating the suitability of deceased organ donor kidneys relies on clinical, laboratory, and biopsy data. Pathologist quantitation of glomerulosclerosis is a critical histologic parameter used in the decision to transplant. This is an arduous task with modest reproducibility and inconsistent correlation with graft outcomes. Prior work shows that deep learning methods automate glomerulosclerosis quantitation with performance superior to on-call pathologists. Extending beyond this prior work, in this study we find that deep learning glomerulosclerosis quantitation better correlates with graft outcomes. A customized and updated deep learning model was used to analyze 691 procurement biopsies of transplanted kidneys with an average recipient follow up of 4.34 ± 1.90 years. The model evaluates a whole slide image in 26.5 +/- 6.3 s. Both pathologist and deep learning model quantitation of glomerulosclerosis significantly correlated with recipient glomerular filtration rate. A multivariate Cox model was developed using glomerulosclerosis, the Kidney Donor Profile Index, cold ischemia time, and recipient age, body mass index, and history of diabetes. Deep learning quantitation of glomerulosclerosis, but not pathologist quantitation, correlated with graft survival after adjusting for other clinical and laboratory parameters. These results show deep learning quantitation of glomerulosclerosis can perform at the speed of clinical care and shows superior performance for graft outcomes, potentially optimizing kidney transplant decisions.