Overview

Schema Version

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

Name

Cardiac Device Identification on Standard DICOM and Smartphone-based Chest Radiographs

Link

https://pubmed.ncbi.nlm.nih.gov/39017033/

Indexing

Keywords: Conventional radiography, Chest radiograph, Segmentation, Classification, CIED, Pacemaker, Defibrillator, Smartphone images, DICOM

Content: CH, CA

RadLex: RID5436, RID49475, RID5431, RID5434, RID13060

Author(s)

Felix Busch

Keno K. Bressem

Phillip Suwalski

Lena Hoffmann

Stefan M. Niehues

Denis Poddubnyy

Marcus R. Makowski

Hugo J. W. L. Aerts

Andrei Zhukov

Lisa C. Adams

Organization(s)

Charité–Universitätsmedizin Berlin

German Heart Center, Technical University of Munich

Department of Diagnostic and Interventional Radiology, Technical University of Munich

Mass General Brigham, Harvard Medical School

Dana-Farber Cancer Institute and Brigham and Women’s Hospital

Maastricht University

Version

1.0

Contact

ed.mut@messerb.onek

Funding

Authors declared no funding for this work; the study did not receive industry support.

Ethical review

Institutional review board–approved retrospective study (Charité–Berlin University Medicine, approval no. EA4/042/20); informed consent waived due to retrospective design.

Date

Updated: 2024-09-01

Published: 2024-07-17

Created: 2023-11-09

References

[1] Busch F, Bressem KK, Suwalski P, Hoffmann L, Niehues SM, Poddubnyy D, Makowski MR, Aerts HJWL, Zhukov A, Adams LC. "Open Access Data and Deep Learning for Cardiac Device Identification on Standard DICOM and Smartphone-based Chest Radiographs". Radiology: Artificial Intelligence. 2024 Sep;6(5):e230502.. 2024-07-17. doi:10.1148/ryai.230502. PMID: 39017033. PMCID: PMC11427927.

Model

Architecture

U-Net with ResNet-50 backbone for CIED segmentation; ResNet-50 classifier for manufacturer and exact model classification; implemented in Python (3.8.16) using FastAI (v2.4).

Availability

Public code and data: https://github.com/AndreasZhukov/cied; dataset to be made accessible on PhysioNet (per article).

Clinical benefit

Automates identification of CIED manufacturer and exact model on chest radiographs, supporting rapid device interrogation, programming, and MRI compatibility assessment, including in emergency care settings.

Clinical workflow phase

Clinical decision support; bedside and emergency triage support.

Decision threshold

For segmentation-based detection, threshold defined as ≥25% of the image crop containing the CIED vs <25% for detection accuracy calculation.

Degree of automation

Fully automated segmentation and classification; outputs intended to support, not replace, clinician decision-making.

Indications for use

Adults (≥18 years) with implanted cardiac devices (pacemakers, implantable cardioverter defibrillators, CRT devices, implantable cardiac monitors) imaged with chest radiography (DICOM or smartphone photographs of radiographs) in clinical settings including emergency care.

Input

Chest radiograph images (DICOM) and smartphone photographs of chest radiographs (converted to PNG for processing).

Instructions

Use segmentation model to localize CIED and crop largest connected region; feed cropped region to classification models. Training used augmentation (±30° rotation, horizontal flip), batch size 32, LR 1e-3, weight decay 1e-5, Adam optimizer, 50 epochs.

Limitations

Single-center dataset (897 patients) with uneven distribution across CIED models/manufacturers; limited geographic diversity of devices; classes with <20 images grouped as 'other'; smartphone photos taken from monitors may differ from real-world variability at other sites.

Output

CDEs: RDE2566, RDE2565

Description: - Segmentation mask of CIED on chest radiograph; - Classification labels for CIED manufacturer and exact device model.

Recommendation

Intended for research and development; supports clinical identification workflows when validated locally.

Reproducibility

Code and labeled dataset publicly available; training environment specified (Ubuntu 18.04, NVIDIA GTX 2080Ti, Python 3.8.16, FastAI 2.4); data splits reported (approx. 70.03%/18.52%/11.45%).

Sustainability

Training performed on a single NVIDIA GTX 2080Ti GPU; runtime/energy consumption not reported.

Use

Intended: Detection and diagnosis

Out-of-scope: Decision support, Detection and diagnosis

Excluded: Other

User

Intended: Physician, Researcher, Referring provider, Radiologist, Other

Out-of-scope: Layperson

Excluded: Layperson