Pathomic analysis of 5-year surveillance biopsies as predictors of kidney allograft loss

Abstract

This study aimed to assess whether various pathomic features in automatically segmented glomeruli and tubules were associated with death-censored graft loss in kidney transplant recipients. We studied 906 patients who were transplanted between 2000 and 2013, returned to Mayo Clinic for a 5-year surveillance biopsy, and had available subsequent follow-up. Banff scores were obtained from the pathology reports and clinical characteristics from the patients’ charts. A multicompartment segmentation model segmented nonsclerotic glomeruli and tubules, and within these compartments, morphometric and textural (pathomic) features were computed. There were 70 death-censored graft losses that occurred during the 5 years of follow-up, and Cox models were used to assess the risk of death-censored graft failure with all Banff scores, with 2 composite pathomic features (glomeruli and tubules), and with Banff scores and pathomic features combined. The composite pathomic score model had similar performance to the Banff score model (c-statistics 0.845 vs 0.823, P = .47). A model combining the Banff scores and pathomic scores showed the best performance (c-statistic 0.897). We conclude that multicompartment segmentation models can identify pathomic features which can complement Banff scoring and are associated with future graft loss.

This study aimed to assess whether various pathomic features in automatically segmented glomeruli and tubules were associated with death-censored graft loss in kidney transplant recipients. We studied 906 patients who were transplanted between 2000 and 2013, returned to Mayo Clinic for a 5-year surveillance biopsy, and had available subsequent follow-up. Banff scores were obtained from the pathology reports and clinical characteristics from the patients’ charts. A multicompartment segmentation model segmented nonsclerotic glomeruli and tubules, and within these compartments, morphometric and textural (pathomic) features were computed. There were 70 death-censored graft losses that occurred during the 5 years of follow-up, and Cox models were used to assess the risk of death-censored graft failure with all Banff scores, with 2 composite pathomic features (glomeruli and tubules), and with Banff scores and pathomic features combined. The composite pathomic score model had similar performance to the Banff score model (c-statistics 0.845 vs 0.823, P = .47). A model combining the Banff scores and pathomic scores showed the best performance (c-statistic 0.897). We conclude that multicompartment segmentation models can identify pathomic features which can complement Banff scoring and are associated with future graft loss.