Classification of the Stages of Nonalcoholic Steatohepatitis via Federated General Visual Representation Learning

Abstract

ABSTRACTThe excess buildup of fat within a patient's liver, even in the absence of any previous alcohol consumption, defines nonalcoholic fatty liver disease (NAFLD). Accurate classification of NAFLD subtypes, especially nonalcoholic steatohepatitis (NASH), through histology biopsies is vital for risk stratification and patient management, including liver transplantation decisions. Therapeutic options targeting liver molecular pathways are being explored, as approved therapies for NASH/NAFLD are currently lacking. Deep learning‐driven pathological diagnosis holds promise for precise assessments, and the application of the federated learning version of the Big Transfer (BiT) model in this study represents a significant advancement in NAFLD/NASH classification and staging using histopathological imaging data. The AUC performance results of federated learning settings of “mainclass” and “multiclass” scenarios are obtained respectively as 0.9299 and 0.9582. The AUC results of each “subclass” are achieved as 0.9376, 0.9435, 0.9867, and 0.9618 respectively, for ballooning, inflammation, steatosis, and fibrosis. The effectiveness of the federated learning configurations in the scenarios categorized as “main‐class” and “multi‐class” is reflected in their respective accuracy (ACC) performance outcomes, which are recorded as 0.7650 and 0.6653. Furthermore, the ACC achievements for each specific “sub‐class” namely, ballooning, inflammation, steatosis, and fibrosis are 0.9016, 0.8568, 0.9203, and 0.8815, respectively.