Liver Biopsy Graph Neural Networks for Automated Histologic Scoring using the NASH CRN System

The prevalence of nonalcoholic fatty liver disease (NAFLD) is rising rapidly, resulting in a concurrent increase in its progressive form, nonalcoholic steatohepatitis (NASH), which can lead to cirrhosis1 . There are no currently approved therapeutics for NASH, with promising candidate drugs failing to meet surrogate clinical trial endpoints approved by the FDA that require pathologic review of liver biopsies1 . Inter- and intra-pathologist variability in grading and staging disease based on histological features leads to inconsistency which may impact results. “Black box” machine learning approaches using conventional neural networks (CNNs) can interpret NASH histology on digitized slides, but their application is limited by lack of interpretability. Graph Neural Networks (GNNs) are an emerging deep learning method that represent and characterize histologic features using graph representations and are well-suited to data types that can be modeled by a graph structure, such as fibrosis architecture. The GNN described here has been incorporated in to a PathAI Drug Development tool (DDT) which provides AI-based histological measurements of nonalcoholic steatohepatitis (AIM-NASH). This AIM-NASH DDT has been accepted by the FDA into the Biomarker Qualification Program to be evaluated for use in trial enrollment and to determine histologic-based endpoints.