Application of artificial intelligence (AI) in assessment of renal allograft biopsies: Comparison with subjective Banff criteria for graft inflammation.

Abstract

Background & objectives: The Banff classification is used routinely in reporting of transplant biopsies. The variable reproducibility of Banff criteria limits its clinical value. This study evaluated the use of AI to quantify graft inflammation and correlated inflammatory cell density with Banff scores. Method(s): Digital images from 119 allograft biopsies were scored by two renal pathologists according to Banff criteria. Images were then analysed using the Visiopharm Integrator System platform. An APP was designed to quantify inflammation within normal and fibrotic cortex, and an auxiliary APP to exclude glomeruli. The model was trained through one million iterations. Result(s): The inter-observer agreement for i, ti and i-IFTA scores was moderate to good (kappa-scores 0.615, 0.566, 0.552 respectively). Banff scores correlated only weakly with the density of inflammatory cells assessed using AI. i-score: only for i0 vs i3 was there a significantly different inflammatory cell density. i-IFTA and ti scores: the density was significantly greater for Banff categories 1-3 than 0, but there were no significant differences between categories 1-3. Conclusion(s): he density of inflammatory cells correlates weakly with Banff scores, likely reflecting the fact that a pathologist will only recognise cortex as inflamed when there is greater than a threshold level of inflammatory cell density in a particular area. It remains to be ascertained whether inflammatory cell density assessed by AI or subjective Banff scores correlate better with clinical outcome and response to treatment.