The 3C dataset: A comprehensive dataset for COVID-19 cardiac complications diagnosis

Abstract

Coronavirus Disease 2019 (or commonly called COVID-19) infections have been associated with numerous severe cardiovascular complications, including myocardial infarction, myocarditis, and arrhythmias, which pose significant risks to long-term cardiac health. Despite extensive research on COVID-19–related cardiovascular diseases, there remains a shortage of curated, annotated medical imaging datasets to support studies on the early detection or prediction of post-COVID cardiac complications.

To address this gap, we introduce the 3C dataset (Cardiac-CT-COVID-19), a public benchmark database of cardiac computed tomography (CT) scans from 134 COVID-19-positive patients, including cases both with and without documented cardiovascular complications.

All scans were annotated by two radiologists with expertise in cardiothoracic imaging, incorporating (1) pixel-level delineations of cardiac structures and pathological findings (e.g., myocardial scarring and pericardial effusion), and (2) a normalized severity grading scale for cardiovascular manifestations. The dataset also includes demographic and clinical metadata to support holistic analysis. The 3C dataset is the first to offer detailed, expert-validated annotations that link COVID-19 infection to specific cardiac abnormalities observed in CT scans. The database is expected to provide an excellent benchmark for researchers working on tasks such as automated detection, prediction, and characterization of diverse cardiovascular manifestations secondary to COVID-19.

Specifications Table

Subject	Health Sciences, Medical Sciences & Pharmacology
Specific subject area	CT imaging for diagnosing COVID-19
Type of data	Image
Data collection	A total of 98,643 CT images were collected, representing image slices for different CT scans related to 134 subjects with confirmed COVID-19. Two senior radiologists analysed all the images and provided accurate annotations and labelling. Furthermore, we annotated the dataset images with a severity grading scale considering all cardiovascular conditions.
Data source location	Military Hospital of Instruction of Tunis, Tunisia.
Data accessibility	Repository name: 3C dataset: Cardiac-CT-Covid19 Data identification number: 10.5281/zenodo.10801492 Direct URL to data: https://doi.org/10.5281/zenodo.10801492 Instructions for accessing these data: the dataset is available upon reasonable request by contacting the first author at [narjes.benameur@istmt.utm.tn]. Access is conditional upon appropriate acknowledgment and citation of this article.
Related research article	Benameur, N., Sassi, A., Ouarda, W., Mahmoudi, R., Arous, Y., Mohammed, M. A., Amar, C. b., Labidi, S., & Mahjoubi, H. (2024). Deciphering the Complexities of COVID‐19‐Related Cardiac Complications: Enhancing Classification Accuracy with an Advanced Deep Learning Framework. International Journal of Imaging

1. Value of the Data

The availability of CT images and DICOM files of lungs and thoracic cavities from COVID-19 patients with or without cardiac problems provides a valuable resource for various ML-based medical research and applications. Some key aspects of the data's value are:

• •Disease Diagnosis, Prognosis, and Severity Assessment: The dataset enables the development of ML models that can assist in the automated diagnosis of COVID-19-related lung abnormalities and cardiac complications. By learning patterns from CT images, these models can not only identify health issues but also predict the severity of the disease, helping healthcare professionals to provide timely and targeted interventions [[1], [2], [3], [4], [5], [6]].
• •Treatment Planning and Monitoring: ML models can be used to analyze the data and recommend appropriate treatment plans for patients with cardiac problems. Additionally, these models can monitor disease progression and treatment effectiveness [7].
• •Risk Stratification: ML algorithms trained on this data can stratify patients into different risk groups based on lung and cardiac findings [8,9]. This can aid in prioritizing resources for patients at higher risk and optimizing healthcare delivery.
• •Data-driven Research: Researchers can use the anonymized data to conduct large-scale studies on the impact of COVID-19 on lung and cardiac health. This data can facilitate discoveries of new insights and correlations between different variables.
• •Benchmarking and Training Robust ML Models: This dataset can serve as a benchmark for evaluating the performance of new ML algorithms in medical image analysis. Additionally, the diverse CT images from different patients provide a rich source for training ML models that are more robust and capable of generalizing well to new, unseen cases [[10], [11]].
• •Developing Decision Support Systems: ML models can be integrated into decision support systems that provide evidence-based recommendations to healthcare professionals, enhancing diagnostic accuracy and treatment decisions.

2. Background

The collected Cardiac-CT- COVID-19 (3C) dataset is specifically designed to detect and predict cardiac complications resulting from COVID-19 infection. Researchers can use the anonymized data to conduct large-scale studies on the impact of COVID-19 on both lung and cardiac health. Numerous studies have already demonstrated a wide range of complications associated with the virus, including arrythmia, myocardial infarction and takotsubo cardiomyopathy [12,13]. However, these studies present several limitations: the COVID-19 datasets that include cardiac complications are relatively small, and they often lack detailed information and specificity regarding the nature of these complications. To address these limitations, there is a clear need for a publicly available COVID-19 dataset that includes CT images along with detailed data on cardiac complications and their severity. In response to this need, the present study introduces an open dataset for COVID-19, including patients with myocardial infarction, myocarditis, and those without cardiac complications. The severity of all COVID-19 cases with cardiac complications has been quantified. In addition to its clinical value, the 3C dataset can be used in deep learning (DL) models to enhance reconstruction quality or even convert non-contrast CT scans into contrast-enhanced ones, reducing the need for multiple acquisitions during the same clinical work-up. It can also support segmentation of the left atrium and ventricle. A key specificity of the 3C dataset is its systematic classification of COVID-19 patients into four categories of pulmonary involvement (<10%, 10–25%, 25–50%, >50% of lung volume), providing a unique opportunity to build predictive models of lung severity. These labels enable supervised or semi-supervised algorithms to automatically assess disease severity and potentially predict outcomes by integrating cardiac and pulmonary features. Overall, the dataset offers a strong foundation for AI-based methods in anatomical segmentation, reconstruction, severity prediction, and risk stratification.

3. Data Description

The images were collected from the Military Hospital of Instruction of Tunis to discern cardiac damage resulting from two types of COVID-19 pulmonary abnormalities: pneumonia and parenchymal involvement. All COVID-19 cases were confirmed via real-time reverse transcription-polymerase chain reaction (RT-PCR).

Patient selection for the 3C dataset was based on well-defined inclusion and exclusion criteria to ensure a consistent and clinically relevant cohort. Inclusion criteria included confirmed COVID-19 infection, availability of high-quality thoracic CT scans, and complete clinical and demographic data, while poor-quality or incomplete CT scans, missing essential clinical information, age below 18 years, and pregnancy were the main exclusion criteria. Following these criteria, patients were further classified based on the presence or absence of documented cardiovascular complications, such as myocardial infarction, myocarditis, arrhythmia, or coronary artery disease. Fig. 1 presents the main criteria of selection and exclusion for our selected population.

Fig. 1.

Patient selection workflow for the 3C dataset.

The 3C dataset contains a total of 98,643 DICOM images representing image slices from different CT scans of 134 subjects with confirmed COVID-19, including 93,423 CT images of COVID-19 patients without cardiac complications and 5220 CT images with cardiac complications. The data is available from the corresponding author upon request, with a total size of 47 GB. Table 1 summarizes the main characteristics of the patient cohort.

Table 1.

Summary of Patient demographics in the 3C dataset.

Titles	Entire Dataset	COVID-19 subjects without cardiac complications	COVID-19 subjects with cardiac complications
Number of Patients	134	108	26
Age, years (range)	[18 - 89]	[18 - 89]	[34 - 87]
Sex (Male/Female)	85/49	68/40	17/9

4. Experimental Design, Materials and Methods

4.1. Image acquisition and extraction

The dataset received approval from the Radiology Department of the Military Hospital of Tunis (Tunisia). The clinical protocol for image acquisition involved several sequential steps. This included patient preparation, positioning on the scanning table, calibration of the CT scanner, X-ray imaging to capture cross-sectional views of the heart, optional contrast administration, data post-processing, and image interpretation by two experienced radiologists (with 10 and 7 years of CT image expertise) Fig. 2.

Fig. 2.

Examples of COVID-19 severity levels in the lungs of patients with myocardial infarction. (a) Axial slice of a thoracic CT scan showing bilateral ground-glass opacities with a peripheral distribution and interlobular reticulations, predominantly affecting the upper lobes. (b) CT image of a patient with moderate severity, demonstrating peripheral ground-glass opacities in both upper lobes. (c) CT scan of a patient with mild severity, showing thoracic distension with paraseptal emphysema.

The CT scans for combined cardiac and pulmonary evaluation were performed using a Siemens SOMATOM Definition AS 64-slice CT scanner. Acquisition parameters were set to ensure high-resolution imaging of both the heart and lungs. All details of the scanner settings and reconstruction protocols are included in the Table 2.

Table 2.

CT scan parameters for cardiac and pulmonary imaging.

Parameter	value
Scanner	Siemens SOMATOM Definition AS 64-Slice
Tube voltage (kVp)	120
Tube current (mAs)	200–270
Slice thickness	1
Reconstruction interval (mm)	0.5–1
Pitch	0.9–1.2
Reconstruction kernel	Soft tissue (cardiac), lung (pulmonary)
Contrast agent	1.5 ml/kg, 3–4 ml/s

The study adhered to the principles outlined in the Declaration of Helsinki, and approval was obtained from the Ethics Committee for the protocol. All COVID-19 data were classified into 4 classes according to the following lung severity level:

• •< 10% of lung volume: little damage
• •10%−25% of lung volume: moderate damage
• •25% −50% of lung volume: severe damage
• •> 50% of lung volume: critical pulmonary damage:

Furthermore, for CT data associated to patients with cardiovascular diseases, the type of complication was defined for each patient. Detailed clinical information is provided in Table 3.

Table 3.

Clinical and laboratory measurements of COVID-19 patients with and without cardiac complications.

Characteristics	COVID-19 subjects without cardiac complications (n = 108)	COVID-19 subjects with cardiac complications (n = 26)
Temperature (  °c)	37.3 [36.4–39,1]	38,2 [36.9–40.2]
Pulse (/min)	79 [75–89]	77 [74–90]
History of Coronary Artery Disease (CAD)	n = 11 (10.18%)	n = 3 (11.53%)
Hypertension	n = 43 (39.81%)	n = 12 (46.15%)
Type 1 - Diabetes	n = 17 (15.74%)	n = 5 (19.23%)
Type 2 - Diabetes	n = 37 (34.25%)	n = 6 (23.07%)
N-terminal pro–B-type Natriuretic Peptide (NT-proBNP), pg/ml	243[37.9–837]	[650.2–1472]
Type of pulmonary abnormalities: Pneumonia (%) Parenchymal involvement (%)	Pneumonia (35%) Parenchymal involvement (65%)	Pneumonia (44%) Parenchymal involvement (56%)
Type of cardiac complications:		Myocarditis (69%)
		Arrhythmia (22%)
		Myocardial Infarction (9%)

4.2. Data pre-processing: anonymization

The process of anonymization is a critical step in medical research to protect patient privacy while allowing researchers to use and share medical data. In this context, anonymization refers to the removal or de-identification of personal and sensitive information from DICOM (Digital Imaging and Communications in Medicine) files, which are the standard format for medical imaging data.

The steps involved in the anonymization process can be described as follows:

• 1.
  Identification of Sensitive Information:

  • ○
    The code identifies various DICOM tags that contain sensitive patient information. These tags include patient demographics, study details, institutional information, acquisition details, and other identifying information.
• 2.
  Removal of Sensitive Tags:

  • ○
    The identified sensitive DICOM tags are removed from the DICOM files to eliminate any personally identifiable information.
  • ○
    The tags are deleted from the DICOM objects to ensure that the sensitive information is not present in the anonymized files.
• 3.
  Removal of Private Tags:

  • ○
    Private tags, if present in the DICOM files, are also removed to eliminate any proprietary or institution-specific information.
• 4.
  Output Folder Structure:

  • ○
    An output folder structure is created to mirror the original directory structure of the CT files.
  • ○
    The anonymized CT files are saved in the corresponding output folders.
• 5.
  File Naming:

  • ○
    The anonymized CT files are given new names to further dissociate them from the original patient and study identifiers.
  • ○
    A naming scheme is used to ensure uniqueness and track the order of anonymized files.

By following these steps, the code anonymizes the CT files in the provided root directory. The anonymized files are then saved in the output folder with modified names and without sensitive patient information. Patient folders are anonymized using identifiers (P01, P02, …), while image series are labeled using DICOM series identifiers (e.g., S0001, S0002). These codes correspond to CT acquisition series. The resulting anonymized dataset can be used for research purposes while protecting patient privacy.

Limitations

The COVID-19 data collection was limited to the most frequently observed cardiac pathologies, even though other cardiac conditions may also occur following COVID-19 infection. Since the dataset primarily focuses on the cardiac effects of COVID-19, rarer heart diseases may not be fully represented. Furthermore, although the dataset includes 134 patients (98,643 CT slices), it may not adequately reflect the global population, as this limitation arises from the single-center design and the availability of imaging data during the study period.

Ethics Statement

All contributions to this study were approved by the Institutional Review Board of the Military Hospital of Instruction of Tunis, with the protocol number “04/2023/CLPP/HMPIT”, including a waiver of informed consent, and performed in accordance with data protection policy. All research has been carried out in accordance with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

Data Availability

• 3C dataset: Cardiac-CT-Covid193C dataset: Cardiac-CT-Covid19 (Original data).