Source-linked AI summary
COVID-19 Image Data Collection
Joseph Paul Cohen, Paul Morrison, Lan Dao
TL;DR
The paper addresses the lack of a public COVID-19 chest-imaging collection designed for computational analysis. It assembles a public database from websites, publications, and extracted PDF images, with the intended use of developing tools for diagnosis, pneumonia characterization, progression monitoring, and outcome prediction.
Problem
Public datasets contained more typical chest X-rays, but no collection of COVID-19 chest X-rays or CT scans was designed for computational analysis.
Method
The authors assemble a public database of COVID-19 and other pneumonia cases using images and metadata collected from public websites, publications, and PDF extractions.
Results
The resulting dataset is intended to support tools for identifying COVID-19 characteristics, distinguishing pneumonia types, monitoring progression, and predicting outcomes.
Takeaways & Limitations
The collection provides data for studying COVID-19 radiological findings and developing computational approaches to diagnosis and patient management.
Abstract
from arXiv · showhide
This paper describes the initial COVID-19 open image data collection. It was created by assembling medical images from websites and publications and currently contains 123 frontal view X-rays.
1. Motivation
The paper addresses the lack of computationally usable COVID-19 chest-imaging data by introducing a public database assembled from public sources. The authors envision using it to develop tools for identifying COVID-19 characteristics and comparing pneumonia types.
- Motivation: There was no public collection of COVID-19 chest X-rays or CT scans designed for computational analysis.Large public datasets existed for more typical chest X-rays, but not for COVID-19 imaging.
- Contribution: The paper introduces a public database containing COVID-19, MERS, SARS, and ARDS pneumonia cases with chest X-ray or CT images.The images and data were collected from public sources to avoid infringing patient confidentiality.
- Motivation: The database is intended to support deep-learning systems that identify COVID-19 characteristics, distinguish pneumonia types, or predict survival.The authors specifically mention transfer learning as a likely approach.
- Release: The images and data were released through a public GitHub URL after being collected from already-public sources.
2. Expected outcome
The dataset is intended to support research on COVID-19 progression, radiological differences, diagnosis, triage, and patient outcomes. Proposed tools could also help monitor disease evolution and inform treatment understanding.
- Scientific uses: The dataset can support study of COVID-19 progression and variation in radiological findings from other pneumonias.
- Prediction and management: Tools could predict pneumonia type and outcome, including survival, to facilitate patient management.The text frames these capabilities as potential future developments.
- Triage: Tools could triage cases when physical tests are unavailable, particularly during shortages of polymerase chain reaction tests.
- Monitoring: Monitoring COVID-19-positive patients could help track disease evolution and improve understanding of disease dynamics and treatments.
3. Dataset
The dataset was assembled from websites, publications, and extracted PDF images while preserving image quality and collecting metadata attributes. Its source material includes a set of cited medical papers.
- Data collection: Images and metadata were compiled from websites including Radiopaedia.org, the Italian Society of Medical and Interventional Radiology, and Figure1.com.Images were also extracted from online publications, websites, or PDFs using pdfimages.
- Data collection: The collection process aimed to maintain image quality during extraction.
- Sources: The dataset drew medical images from the cited papers listed in the dataset section.
- Metadata: Metadata attributes are described in a dedicated table accompanying the dataset.