[Paper Review] COVID-19 Image Data Collection

Joseph Cohen, Paul Morrison+1 more|arXiv (Cornell University)|Mar 25, 2020

COVID-19 diagnosis using AI17 references818 citations

TL;DR

This paper presents the initial open COVID-19 chest X-ray dataset (COVID-Chest X-ray Dataset), aggregating 123 frontal X-ray images from public sources for AI research.

ABSTRACT

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.

Motivation & Objective

• Motivate the creation of an open COVID-19 chest radiograph dataset to aid diagnostic tool development.
• Provide a public resource to train and test deep learning models for distinguishing COVID-19 from other pneumonias.
• Facilitate studies on disease progression, outcomes, and triage in the context of limited PCR testing.

Proposed method

• Aggregate public X-ray/CT images of COVID-19 and related conditions from sources like Radiopaedia, Figure1, and published papers.
• Extract and preserve image quality from PDFs and websites using pdfimages and manual curation.
• Define a metadata schema (Patient ID, Age, Sex, View, Modality, Finding, Survival, Date, Location, License, etc.).
• Publish the dataset publicly at a GitHub URL and reference contributing sources to maintain provenance.

Experimental results

Research questions

• RQ1 Can a publicly available COVID-19 chest radiograph dataset enable training and evaluation of pneumonia/disease classification models?
• RQ2 How do radiographic findings of COVID-19 differ from other pneumonias or ARDS in available images?
• RQ3 Can this dataset support modeling of disease progression and patient survival using radiographic data?

Key findings

• The dataset described contains 123 frontal chest X-ray images as of March 25, 2020.
• Images and metadata are collected from public sources to avoid patient confidentiality issues.
• The metadata schema includes attributes such as Patient ID, Offset (days since symptoms), Sex, Age, Finding, Survival, View, Modality, Date, Location, Filename, License, and notes.
• Initial use case emphasizes training/deploying deep learning models to identify COVID-19 characteristics and to predict outcomes.
• The paper situates the dataset as a resource to study progression and compare COVID-19 radiographic patterns with other pneumonia types.