Deep learning-based method for grading histopathological liver fibrosis in rodent models of metabolic dysfunction-associated steatohepatitis.

INTRODUCTION: Metabolic dysfunction-associated steatohepatitis (MASH) is a significant liver disease that can lead to cirrhosis and liver cancer. Accurate assessment of liver fibrosis is crucial for diagnosis, prognosis, and informed treatment decision-making. Staging of liver fibrosis in MASH is based on Kleiner's score, which categorizes fibrosis based on its location within the liver as observed microscopically. This scoring system is part of a standard clinical research network and relies heavily on the expertise of pathologists. METHODS: This study utilized Sirius Red-stained whole slide images of liver tissue obtained from various MASH animal models to develop deep learning (DL) models for scoring liver fibrosis, with a focus on the criteria outlined in Kleiner's score. We created a trainable and testable dataset of whole-slide images of the liver, consisting of 999,711 patch images derived from 914 whole-slide images. The performance of the multi-class classification model was evaluated using the kappa statistic, area under the precision-recall curve (AUPRC), area under the receiver operating characteristic curve (AUROC), and Matthews correlation coefficient (MCC). RESULTS: To address challenges in clinical subclassification, a 5-class classification model was initially applied; the model achieved moderate agreement. A more refined 7-class model was subsequently developed, which outperformed the 5-class classification model. The enhanced subclassification significantly improved classification performance, as evidenced by the superior AUROC and AUPRC values of the 7-class model. DISCUSSION: This study highlights that DL models for scoring liver fibrosis can support expert pathologists in staging liver fibrosis in preclinical animal studies.