Artificial Intelligence Quantitation of Steatosis on Frozen Section Preimplant Slides Correlates with Some Liver Transplant Outcomes

Abstract

Introduction.

Increased steatosis on preimplant liver frozen section is associated with delayed graft function and primary nonfunction. Efforts to standardize histologic assessment have proven difficult. Frozen section artifact and lipopeliosis complicate the detection of steatosis. We aimed to develop and validate an AI model to recognize large droplet fat and fat induced artifact (FIA)/lipopeliosis on preimplantation frozen section and to correlate the AI results with post-transplant clinical parameters.

Methods.

The model was applied to 161 consecutive liver transplant specimens with preimplant slides. Results were correlated with traditional and Banff histologic assessment and clinical parameters.

Results.

By traditional assessment, steatosis ranged from 0%–40%. The AI model identified a range of 0 to 15.9% steatosis. There was no difference in patient survival by any measures of steatosis. AI steatosis correlated with increased risk of early allograft dysfunction (OR = 1.63, P < .001), respiratory failure (OR = 1.21, P = .003), and more advanced fibrosis (OR = 1.18, P = .030), but was not correlated with graft or patient survival. FIA/lipopeliosis were identified in a range of 0 to 6.42%. In univariate analysis the percentage of FIA/lipopeliosis correlated with both graft and patient survival (P = .044 and P = .009, respectively), but was not associated with increased risk of early allograft dysfunction, respiratory failure, or advanced fibrosis.

Conclusions.

We developed an AI model that quantitates large droplet fat and FIA/lipopeliosis on frozen section slides and found a correlation with post-transplant outcomes. Further studies on larger, multi-institutional cohorts with higher fat containing donors are necessary to determine the role this model may have in organ acceptance decisions.