Dataset for Segmentation and Classification of Cardiac Implantable Electronic Devices in Chest X-Rays

Abstract

This dataset encompasses a comprehensive collection of 2,321 chest radiographs (originally DICOM, converted to PNG) and 11,072 smartphone images documenting various cardiac implantable electronic devices (CIEDs) such as implantable pacemakers, cardioverter defibrillators, cardiac resynchronization therapy devices, and cardiac monitors, collected from 896 patients at the Charité - Universitätsmedizin Berlin over a decade (January 2012 to January 2022). The dataset's primary objective is to advance the development and validation of automated techniques for precise segmentation and classification of CIEDs across both traditional and smartphone-generated imagery. It includes a diverse array of 27 CIED models from four manufacturers (Medtronic, Biotronic, Boston Scientific, Abbott Laboratories including St. Jude Medical), captured in anterior-posterior or posterior-anterior chest radiographs using five different smartphone brands. This dataset aims to underpin the creation of deep learning models for accurately identifying CIED types, manufacturers, and models. Designed to be a robust foundation for research and clinical applications, it facilitates the exploration of innovative machine learning solutions for CIED identification leveraging both original radiogrpahs and smartphone images, thereby addressing a critical need in cardiac care technology.

Background

Cardiac implantable electronic devices (CIEDs), which include implantable cardiac monitors (ICMs), pacemakers (PMs), and implantable cardioverter defibrillators (ICDs), are critical diagnostic and therapeutic tools used in the management of a variety of cardiac conditions. These devices are essential for electrical modulation, prevention of sudden cardiac death, and continuous cardiac monitoring [1]. As cardiovascular disease continues to impact global health, the need for CIEDs is increasing [2]. This increase underscores the importance of accurately identifying CIEDs in clinical scenarios, particularly in emergency settings where patient-provided information might be lacking.

While traditional identification algorithms have been somewhat effective, they often require labor-intensive and time-consuming manual interpretation. Artificial intelligence (AI) methods, particularly convolutional neural networks (CNNs), have shown promise in improving the speed and accuracy of this identification process [3, 4]. However, a major limitation has been the lack of publicly available, high-quality datasets and algorithms, which limits the external validation, transparency, and reproducibility of these methods [5, 6]. This gap is particularly glaring when it comes to identifying not only the manufacturer, but also the specific CIED model.

In response to these challenges, this dataset was curated with the goal of advancing the field by providing the first publicly available resource for developing deep learning models for accurate CIED segmentation and classification. The dataset includes both original images converted to PNG and smartphone images, serving the dual purpose of validating the algorithms in traditional settings and extending their applicability to point-of-care scenarios via smartphone applications.

Methods

Inclusion Criteria:

The study ran from January 2012 to January 2022, and included adult patients (≥18 years) with documented CIEDs on chest radiographs. Informed consent was not required. Eligible images included those from bedside anterior-posterior or standing posterior-anterior chest radiography.

Data Collection:

A comprehensive database search was performed to identify patients who met the eligibility criteria. Images from the same patient taken at different time points were included. Exclusion criteria were images with sagittal beam path, insufficient quality, incomplete device acquisition, or multiple devices of the same type on an image. Two reviewers independently extracted CIED manufacturer and model classifications for each DICOM image from the associated medical reports, with discrepancies resolved by consensus-based discussion. Rare types of CIEDs were classified as other.

Data Pre-processing:

The original DICOM images were anonymized and converted to PNG format. This removed all sensitive header information. Anonymization was then performed in the following steps.

First, rare CEID types were not explicitly classified, but grouped into a different category. This avoids identification of patients by their rare CEID type. Second, center cropping was used to remove potentially burned-in patient information, which is most often found in the corners of the image. Third, an object character recognition (OCR) algorithm was used to black out any image information remaining after center cropping. Finally, each image was again manually inspected for any remaining information, which was then manually removed.

Each of the converted images was then photographed up to five times with one of five different smartphones at different angles from a monitor display.

After re-matching the original and smartphone images using the patient ID, this ID was replaced with a new randomly generated ID. Finally, the patient ID, examination ID, and all other identifiers were deleted to prevent re-identification of the patients, even by the authors.

For segmentation model training, ground truth annotations were manually generated by two board-certified radiologists with 7-9 years of experience in image processing using 3D Slicer software.

Data Description

The dataset consists of a total of 13,393 images, radiographs converted to PNG and smartphone photographs. The images capture a range of CIEDs, including pacemakers, implantable cardioverter defibrillators, cardiac resynchronization therapy devices, and implantable cardiac monitors.

Image Types:

• Chest radiographs: These are high-resolution medical images obtained from the Imaging Department of Charité - University Medicine Berlin. These images serve as ground truth for CIED manufacturer and model identification..

• Smartphone photos: A variety of smartphones were used to capture these images from different angles. The breakdown of smartphone types is as follows:

  • iPhone 13 mini: N=695 (5.2%)
  • Samsung Galaxy S4: N=1,675 (12.5%)
  • iPhone 8: N=2,890 (21.5%)
  • iPhone 13: N=2,902 (21.6%)
  • Samsung Galaxy S5: N=2,910 (21.7%)

Annotations:

Ground-truth annotations for CIED segmentation were manually created on each image by board-certified radiologists. These annotations serve as a reference standard for the evaluation of the segmentation algorithms. They are available as PNG in the `seg_masks` folder. In these images a pixel value of 0 represents the background, 1 represents a pacemaker or defibrillator and 2 represents an event monitor.

Variables and Labels:

• Manufacturer: The manufacturer of the CIED is labeled, providing the ability to train and validate manufacturer-specific algorithms.

• Device Model: Each CIED is labeled with its specific model, facilitating more granular analysis and classification. For details on the distribution of device types and models, please refer to the accompanying table.

File Format:

• Anonymized original images are provided in PNG format.
• Smartphone photos are also available in PNG format.

The table below presents an overview of the various devices included in the study. Bolded rows provide a summary of the total number of devices from each vendor.

Data Structure:

The data is arranged as follows (please also refer to the README.md).

Directory Structure

.
├── README.md  # Data description
├── galaxys4   # Folder with smartphone photos
│   ├── fe40d960c739bbbd3002d1759155bad92d3818b5ec88e459117e64c6df2a0fb2.png
│   ├── fe5b9ad8b3a0046eacc334578a49cabe0837d813a9b8f588538132045c38a2be.png
│   ├── ff76fce90f157c2cdf328e5e012698c26c563b1443a37156e24ac1f7be4c7c76.png
│   ├── ff88824115f5e1f3d237fc6769bca905e8fcf1a89765e1fffdbf0b6aeb1dd89a.png
|   .
│   └── ffeb08d7f946d32c9c7ff6a6235c012b38887b634e246fddfd2e19f9b92d6613.png
├── galaxys5
├── iphone13
├── iphone13mini
├── iphone8
├── originals  # original images
├── seg_masks  # segmentation masks
└── data.json  # meta data, containing labels and IDs.  

df = pd.read_json("data.json")
df.head()

Usage Notes

This dataset is designed to support multiple research applications in medical image analysis and artificial intelligence. Researchers can utilize it to develop and validate deep learning models for automated CIED identification in both clinical and point-of-care settings. The dataset enables three primary applications: device segmentation in chest radiographs, manufacturer classification, and specific model identification. The inclusion of smartphone images alongside traditional radiographs allows for the development of transfer learning approaches and mobile applications for bedside device identification.

Dataset limitations include the single-center origin of the data, which may affect generalizability across different healthcare settings. The dataset covers devices from four major manufacturers but excludes some smaller vendors. Additionally, the smartphone images were captured from monitor displays rather than printed films, which may influence image quality characteristics. To preserve patient anonymity, some rare device types are grouped under the "other" category, which may limit training for these specific models.

For practical implementation, we provide example code and model architectures at our GitHub repository [7]. This includes complete training pipelines for both segmentation and classification tasks, preprocessing scripts for both radiographs and smartphone images, and evaluation metrics for model performance assessment.

Ethics

The institutional review board-approved this retrospective single-center study (EA4/042/20). The need for informed consent was waived as the data was collected retrospectively and was anonymized.

Conflicts of Interest

None

References

1. Goldenberg I, Huang DT, Nielsen JC. The role of implantable cardioverter-defibrillators and sudden cardiac death prevention: indications, device selection, and outcome. Eur Heart J. 2020;41(21):2003-11.
2. Vaduganathan M, Mensah GA, Turco JV, Fuster V, Roth GA. The global burden of cardiovascular diseases and risk: a compass for future health. J Am Coll Cardiol. 2022;80(25):2361-71.
3. Howard JP, Fisher L, Shun-Shin MJ, Keene D, Arnold AD, Ahmad Y, et al. Cardiac rhythm device identification using neural networks. JACC Clin Electrophysiol. 2019;5(5):576-86.
4. Kim UH, Kim MY, Park EA, Lee W, Lim WH, Kim HL, et al. Deep learning-based algorithm for the detection and characterization of MRI safety of cardiac implantable electronic devices on chest radiographs. Korean J Radiol. 2021;22(11):1918.
5. Lauzier PT, Gomes DG, Weng W, Sadek MM, Zakutney T, Bernier ML, Birnie D. Detection and identification of cardiac implanted electronic devices in a large data set of chest radiographs using semi-supervised artificial intelligence methods. Heart Rhythm. 2023;20(4):642-3.
6. Weinreich M, Weinreich B, Chudow JJ, Krumerman T, Rahgozar K, Nag T, et al. Computer-aided detection and identification of implanted cardiac devices on chest radiography utilizing deep convolutional neural networks, a form of machine learning. J Am Coll Cardiol. 2019;73(9S1):307.
7. Cardiac Device Identification [Internet]. GitHub; 2024. Available from: github.com/AndreasZhukov/cied [Accessed 2024 Feb 3].