Abstract
BACKGROUND: Nonalcoholic fatty liver disease (NAFLD) affects 25% of the global population and is a leading cause of cirrhosis. Although liver biopsy remains the diagnostic gold standard, its clinical utility is limited by the labor-intensive Kleiner scoring system. Existing deep learning (DL) solutions face two critical barriers: pathologist-dependent annotations requiring equivalent time to manual assessment, and prohibitive costs for large-scale labeled datasets. The primary objective of this research was to establish a weakly supervised framework using multi-instance learning (MIL) for NAFLD assessment, aiming to reduce the annotation workload for pathologists while developing a clinically applicable diagnostic approach. METHODS: This study utilized a hepatocellular pathology dataset published by Heinemann et al. in July 2023 on the Open Science Framework (OSF) platform (accessible at osf.io/8e7hd). We established MMD-Net, a weakly-supervised framework integrating MIL with multi-task learning (MTL) to concurrently evaluate steatosis, inflammation, and ballooning. To quantitatively evaluate the effectiveness of our model performance, 5 common metrics were employed: accuracy, precision, recall, F1 score, and Cohen's κ. RESULTS: The system achieved exceptional agreement with ground truth, demonstrating quadratic weighted Cohen's κ coefficients of 0.932±0.004 (ballooning), 0.836±0.016 (inflammation), and 0.766±0.029 (steatosis), with mean κ=0.845±0.014. CONCLUSIONS: This approach establishes a new paradigm for standardized NAFLD histopathological assessment while eliminating the need for pixel-level annotations. By doing so, it charts a promising path for artificial intelligence (AI)-powered histopathological analysis to standardized NAFLD assessment in clinical practice.