COVID-Net: An open source convolutional neural network for COVID-19 detection via chest radiography


DarwinAI Open

DarwinAI leverages Explainable AI (XAI) to quickly develop a high-performing and transparent machine learning model used by thousands of researchers, clinical scientists, and citizen scientists around the world

About The Project

Researchers from DarwinAI and the University of Waterloo developed a high-performing convolutional neural network for COVID-19 detection via chest radiography. By leveraging explainability in a human-machine collaborative design strategy, the research team was able to build a model with a high level of accuracy and transparency in under a week. In doing so, they demonstrated that our breakthrough XAI technology has tremendous capacity to benefit healthcare institutions and their patients by accelerating scalable development with much improved model design and performance.

week to create
DarwinAI researchers
COVID-19 sensitivity

Difference Makers

  • Transparency creates trust by showing reasons for diagnosis
  • Low compute requirements enable neural network operation in edge devices
  • Explainability enabled rapid prototyping and optimization

Helping Healthcare Workers in a Pandemic

Recognizing that one of the largest bottlenecks in triage and diagnosis was the need for experts to interpret radiography images, on March 22, 2020, we released COVID-Net: an open source convolutional neural network for COVID-19 detection via chest radiography. In tandem, we also released COVIDx, a dataset which has grown from nearly 6,000 chest radiography images at launch to many times that number. Since that time, we have been privileged to collaborate with a long list of researchers and medical institutions, and COVID-Net has expanded into several specialized derivatives while the COVIDx dataset has grown into one of the world’s largest collections of open COVID-19 chest radiography images.

COVID-Net employs a diverse collection of architectural structures that result in a high-performance model

image description

Building a High-Performing, Purpose-Specific Network

Given the urgency of the initiative, we endeavored to build COVID-Net much faster than typical deep learning initiatives, which can often take several months.

Rather than simply treating AI as a tool to be leveraged, our researchers employed a human-machine design strategy. This philosophy reimagines AI as a collaborator that learns from a developer’s needs and that subsequently proposes multiple design approaches with different trade-offs—thereby enabling a rapid and iterative approach to model building.

  1. Data collection

    To begin the process we constructed a dataset, termed COVIDx, using publicly available sources and images from our collaborators.

  2. Principled network design prototyping

    We constructed a prototype based on human-driven design principles and best practices. Essentially, the prototype provides the initial scaffolding of the model while leaving final architectural decisions to the machine-driven aspect of the process.

  3. Machine-driven design exploration

    Given data and human-defined requirements, the platform guides a design exploration that learns and identifies the optimal macro-and micro-architectures with which the final model can be constructed.

  4. Machine-guided design audit

    We leverage explainability to audit the constructed COVID-Net model—validating consistency between COVID-Net’s behavior and radiologist interpretation, and auditing for erroneous behavior due to inappropriate visual cues (e.g., embedded text markup, imaging artifacts, etc.).

Partnering for Shared Success

COVID-Net, as one of the most popular open source initiatives for COVID-19 clinical decision support, is currently being used by thousands of researchers, clinical scientists, and citizen scientists around the world. Dozens of hospitals around the world are collaborating on this initiative and a number of world-leading technology companies have made significant contributions, including RedHat, Lockheed Martin, Intel, and ARM.

COVID-Net is also one of the most highly cited works in the area of computer-assisted COVID-19 clinical decision support.

Beyond COVID-Net

Since its release—and with the contributions of many other researchers—COVID-Net has been extended and fine-tuned into a range of neural networks including: COVID-Net CT and COVID-Net CT-2, which provide diagnostic assistance with CT scans; COVID-Net S, a neural network for assessing COVID-19 severity; and COVID-Net Clinical ICU, an explainable and trustworthy AI for predicting COVID-19 ICU admission. Our experience with COVID-Net also inspired the creation of the broader DarwinAI Open initiative.

Deep Learning AI developers interested in leveraging our open source models can access them all on GitHub.