Quantifying Fibrosis for Staging Metabolic Dysfunction-Associated Steatotic Liver Disease from Hispathology Images using Deep Learning

Sammanfattning

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a significant and growing clinical challenge that can progress to end-stage liver disease and liver-related mortality. This progression is closely linked to the accumulation of liver fibrosis, making accurate fibrosis staging critical for prognosis. While fibrosis primarily involves the deposition of collagen fibers around vessels, current machine learning-based methods often overlook the spatial relationship between these microstructures or rely on simplistic spatial proximity rules. This limitation restricts the predictive performance of automated staging tools. To address this, we propose a biologically informed, topology-aware graph construction framework where nodes and edges explicitly model vessel-collagen architecture. From these graphs, biologically meaningful feature embeddings are extracted and used to train an advanced graph neural network (GNN) model known as Graph isomorphism networks with edge features (GINE). Using a dataset of 120 slides, covering the full spectrum of fibrosis, we systematically compare our approach with alternative graph construction strategies and benchmark it against a baseline patch-based model and a state-of-the-art (SOTA) graph-based model. Our results demonstrate that, despite challenges in vessel segmentation, the proposed graph construction captures biologically consistent patterns, including a transition from a sparse to an interconnected fibrotic network, and yields improved classification performance compared to purely spatial-based graphs, doubling the accuracy from 0.24 to 0.48. Furthermore, while the baseline and SOTA models struggle with subtle transitions of intermediate fibrosis stages, our framework yields a significantly higher quadratic weighted kappa of 0.44 (an increase of 0.16 over SOTA) and accuracy of 0.37 (an increase of 0.11 over SOTA), demonstrating consistent performance across all stages. These findings highlight the importance of incorporating domain-specific biological knowledge into model design to enhance predictive performance in MASLD fibrosis staging.