Histological AI-Assisted Analysis Outperforms Pathologist Assessment in Predicting Disease Outcomes in PSC Patients

Liver biopsy has shown prognostic significance in primary sclerosing cholangitis (PSC). However, histology is typically assessed manually based on coarse grading scales, making precise quantification difficult. We developed a neural network model to address the key histological features of liver tissue, including portal tracts, fibrosis, biliary epithelium, portal inflammation, and vasculature.

Our training data included 603 liver sections from patients with PSC (n = 478), other hepatobiliary diseases (n = 119), and normal controls (n = 6), ensuring heterogeneity of the sample material. The model was validated and evaluated for its ability to predict key clinical endpoints in PSC: liver transplantation (LT) due to end-stage liver disease (ESLD), and a composite endpoint comprising LT, cholangiocarcinoma (CCA), and liver-related death, allowing evaluation with all relevant clinical outcomes. The model reliably identified the targeted histological features. AI-derived portal inflammation and fibrosis demonstrated superior discriminatory performance compared with manual scoring, particularly for prediction of LT due to ESLD (AUC 0.81-0.82 vs. 0.86-0.87). AI-derived blood vessels and biliary epithelium were also prognostic for LT due to ESLD (AUC 0.80, 0.83). In multivariable analysis adjusted for age and sex, AI-derived portal inflammation remained independently associated with the composite endpoint (HR 2.51, 95% CI 1.04-6.05).

AI-measured portal inflammation is a robust predictor of clinical endpoints. Additionally, vascular area, fibrosis, and biliary epithelium showed prognostic value. AI-based measures outperformed manual scores for several parameters. Overall, multiple AI-derived parameters predicted LT due to ESLD, and comparable trends were observed for the composite endpoint. These findings highlight the importance of liver histology as a surrogate endpoint in patients with PSC.