Deep learning and machine learning in Magnetic Resonance Cholangiopancreatography-based diagnosis and management of primary Sclerosing Cholangitis: A systematic review

Mohammadreza Elhaie, Abolfazl Koozari, Sadra Shamsoldin, Anis Aryafard, Qurain Turki Alqurain,
Deep learning and machine learning in Magnetic Resonance Cholangiopancreatography-based diagnosis and management of primary Sclerosing Cholangitis: A systematic review,
Journal of Radiation Research and Applied Sciences,
Volume 18, Issue 4,
2025,
101994,
ISSN 1687-8507,
https://doi.org/10.1016/j.jrras.2025.101994.
(https://www.sciencedirect.com/science/article/pii/S168785072500706X)
Abstract: Background
Primary Sclerosing Cholangitis (PSC) is a rare, progressive liver disease characterized by bile duct inflammation and fibrosis, often diagnosed using Magnetic Resonance Cholangiopancreatography (MRCP). The complexity of MRCP interpretation necessitates advanced tools like deep learning (DL) and machine learning (ML) to enhance diagnostic accuracy and prognostic prediction. This systematic review evaluates the application of DL and ML in MRCP-based PSC diagnosis and management.
Purpose
To synthesize evidence on the feasibility, accuracy, and clinical utility of DL and ML in MRCP for PSC, focusing on diagnostic performance, disease progression prediction, and clinical decision-making support.
Materials and methods
A systematic review was conducted following PRISMA guidelines, searching PubMed, Embase, Scopus, Web of Science, and Cochrane Library from inception to June 2025. Studies applying DL or ML to MRCP for PSC diagnosis or management were included, with outcomes focusing on diagnostic accuracy (sensitivity, specificity, AUC) and predictive performance. Quality was assessed using the QUADAS-2 tool, and data were synthesized narratively due to heterogeneity.
Results
A total of Six studies (n = 1840 patients) were included, utilizing convolutional neural networks (CNNs), MRCP + software, and algebraic topology-based ML. Diagnostic accuracy was high, with sensitivities of 80.0 %–95.0 % and specificities of 80.0 %–90.9 % for PSC detection, often outperforming radiologists. Prognostic models achieved AUROCs of 0.80–0.86 for predicting hepatobiliary complications, liver transplantation, or hepatic decompensation, surpassing traditional scores (e.g., Anali, Mayo). Clinical utility included early PSC and cholangiocarcinoma detection and risk stratification. Limitations included retrospective designs, small sample sizes, and lack of external validation.
Conclusion
DL and ML applied to MRCP show promise in enhancing PSC diagnosis and prognosis, with high accuracy and potential to reduce interobserver variability. However, methodological limitations and validation challenges necessitate prospective, multicenter studies to ensure generalizability and clinical adoption.
Keywords: Primary sclerosing cholangitis; Magnetic resonance cholangiopancreatography; Deep learning; Machine learning; Artificial intelligence; Diagnostic accuracy; Prognosis