POLCOVID: a multicenter multiclass chest X-ray database (Poland, 2020-2021)

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• The SARS-CoV-2 pandemic strained healthcare systems globally, increasing diagnostic waiting times.
• Chest radiographs (CXR) are crucial for COVID-19 diagnosis, leading to a surge in AI detection tools.
• Existing AI tools often lack sufficient, well-annotated training data, particularly for diverse pneumonia types.

Purpose of the Study:

• Introduce the POLCOVID dataset, a comprehensive collection of chest X-ray images.
• Provide annotated data to support the development of AI for COVID-19 and pneumonia detection.
• Facilitate research in lung segmentation techniques using matched image and mask data.

Main Methods:

• Collected CXR images from 15 Polish hospitals, including COVID-19, other pneumonia, and healthy cases.
• Developed preprocessed images focused on lung areas and generated corresponding lung masks via segmentation models.
• Manually annotated lung masks for a subset of the POLCOVID dataset and four other public CXR collections.

Main Results:

• The POLCOVID dataset comprises original radiographs, preprocessed lung images, and segmentation masks.
• Manually created lung masks are available for enhanced accuracy in model training and validation.
• The dataset supports both direct diagnostic AI development and lung segmentation algorithm research.

Conclusions:

• The POLCOVID dataset addresses the need for high-quality, annotated CXR data for AI in respiratory disease diagnosis.
• It serves as a valuable resource for advancing AI-driven detection of COVID-19 and other pneumonias.
• The provision of lung masks facilitates the development of improved lung segmentation algorithms.