Radiomics‐Based Prognostication in Primary Sclerosing Cholangitis: A Proof‐of‐Concept Study

Background and aim: Risk assessment in primary sclerosing cholangitis (PSC) by magnetic resonance imaging (MRI) relies on semi-quantitative analysis, which can result in interpretation variability. Radiomics may offer a quantitative approach for risk stratification. This study aims to explore and validate MRI-derived radiomic features to identify high-risk PSC patients. Methods: In this prospective study (January 2019-December 2022), consecutive PSC patients undergoing routine gadoxetate disodium-enhanced MRI were recruited. Using PyRadiomics, whole liver parenchyma features were extracted from five MRI sequences according to the Image Biomarker Standardisation Initiative (IBSI). Patients were categorised into risk groups based on the Mayo risk score (MRS) and liver stiffness measurement (LSM). Features associated with high-risk patients were selected and validated in an independent cohort. A survival analysis was conducted in the combined cohort to assess the prognostic value of the radiomic features for clinical events. Results: One hundred and two PSC patients were enrolled in this study. Five radiomics features were associated with high risk in the training cohort. In the validation setting, GLRLM-Run Entropy in the fat-saturation T2 weighted imaging (FS-T2W) sequence was the only significant feature, with an odds ratio of 3.90 (CI 1.46-10.42, p = 0.007) for MRS and 2.97 (CI 1.33-6.66, p = 0.008) for LSM. Its prognostic potential on clinical outcome was confirmed by Cox regression analysis in the combined cohort (hazard ratio per 0.1 increase = 1.480, CI 1.226-1.786), showing excellent predictive performance (C-index = 0.857). Conclusions: GLRLM-Run Entropy in FS-T2W is a novel radiomics-based biomarker for risk stratification in PSC patients. It is quantitative, standardised, easy to compute and cost-free, positioning it as a potential key innovation in PSC radiology-based biomarkers. Trial registration: Clinicaltrial.gov ID: NC705618145.