Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

DeePSC 2D MRCP dataset for automated PSC diagnosis (internal 1.5 T, 3 T; external 3 T different vendor)

Link

https://doi.org/10.1148/ryai.220160

Indexing

Keywords: Neural Networks, Deep Learning, Liver Disease, MRI, Primary Sclerosing Cholangitis, MR Cholangiopancreatography

Content: GI, MR

RadLex: RID38812, RID10312, RID34712, RID16639, RID3386

SNOMED: 197441003

Author(s)

Haissam Ragab

Fabian Westhaeusser

Anne Ernst

Jin Yamamura

Patrick Fuhlert

Marina Zimmermann

Julia Sauerbeck

Farzad Shenas

Cansu Özden

Anna Weidmann

Gerhard Adam

Stefan Bonn

Christoph Schramm

Organization(s)

University Medical Center Hamburg-Eppendorf (UKE) – Department of Diagnostic and Interventional Radiology and Nuclear Medicine

Institute of Medical Systems Biology, Center for Biomedical AI, Center for Molecular Neurobiology, UKE

Department of Medicine, UKE

Hamburg Center for Translational Immunology (HCTI)

Martin Zeitz Center for Rare Diseases, UKE

Funding

F.W.: UKE M3I grant. S.B.: UKE R3 reduction of animal testing grant; LFF-FV 78; DFG-funded CRU 306. A.E., P.F., S.B., C.S.: LFF-FV 78 and DFG-funded CRU 306. M.Z.: DFG SFB 1192 project B8. C.S.: YAEL Foundation; Hannelore and Helmut Greve Foundation; LFF-FV 78; DFG-funded CRU 306.

Ethical review

Retrospective single-center study approved by the University Medical Center Hamburg-Eppendorf IRB (no. 2021–100723-B0-ff); informed consent waived.

Date

Published: 2023-04-19

References

[1] Ragab H, Westhaeusser F, Ernst A, et al.. "DeePSC: A Deep Learning Model for Automated Diagnosis of Primary Sclerosing Cholangitis at Two-dimensional MR Cholangiopancreatography". Radiology: Artificial Intelligence. 2023-05-01. doi:10.1148/ryai.220160. PMID: 37293347. PMCID: PMC10245178.

Dataset

Motivation

Automated classification of PSC-compatible findings on 2D MRCP to support clinical decision-making and address interreader variability.

Sampling

Patients with confirmed large-duct PSC per EASL guidelines and controls without immune-mediated liver/bile duct disease and without visible bile duct alterations, identified via PACS query.

Partitioning scheme

Internal datasets split by field strength (3 T and 1.5 T). For each, 39 patients were randomly assigned to unseen test sets with stratification by demographics; remaining patients used for training. External test set consisted of 37 3-T Siemens exams.

Missing information

Public data release, file formats, de-identification details, and per-partition demographics not specified.

Relationships between instances

Each MRCP exam comprises seven radial 2D images from different angular viewpoints. Some patients underwent both 1.5 T and 3 T exams during follow-up and can appear in both field-strength datasets; within each field-strength dataset, one exam per patient was included.

Noise

Motion artifacts and gastrointestinal fluid can affect images and model predictions; variability in maximum gray values across samples noted.

External data

An independent 3-T Siemens MAGNETOM Vida MRCP cohort (n=37) was used for external testing.

Confidentiality

Retrospective clinical MRCP data from 2002–2022 at a single academic medical center; IRB-approved; consent waived.

Sensitive data

Clinical MR images with associated health information.