The worldwide implementation of a liver graft pool using marginal livers (ie, grafts with a high risk of technical complications and impaired function or with a risk of transmitting infection or malignancy to the recipient) has led to a growing interest in developing methods for accurate evaluation of graft quality. Liver steatosis is associated with a higher risk of primary nonfunction, early graft dysfunction, and poor graft survival rate. The present study aimed to analyze the value of artificial intelligence (AI) in the assessment of liver steatosis during procurement compared with liver biopsy evaluation. A total of 117 consecutive liver grafts from brain-dead donors were included and classified into 2 cohorts: ≥30 versus <30% hepatic steatosis. AI analysis required the presence of an intraoperative smartphone liver picture as well as a graft biopsy and donor data. First, a new algorithm arising from current visual recognition methods was developed, trained, and validated to obtain automatic liver graft segmentation from smartphone images. Second, a fully automated texture analysis and classification of the liver graft was performed by machine-learning algorithms. Automatic liver graft segmentation from smartphone images achieved an accuracy (Acc) of 98%, whereas the analysis of the liver graft features (cropped picture and donor data) showed an Acc of 89% in graft classification (≥30 versus <30%). This study demonstrates that AI has the potential to assess steatosis in a handy and noninvasive way to reliably identify potential nontransplantable liver grafts and to avoid improper graft utilization.
Copyright © 2020 by the American Association for the Study of Liver Diseases.