COVID-Net: larger dataset, new models, and COVID-RiskNet
Note: since this post was published there’ve been a number of developments to the project. See here for the latest.
Dear Colleague,
The response to our release of COVID-Net one week ago has been overwhelming.
Here are the latest updates with the project:
1.) Thanks to the support and feedback from the research community, the COVIDx dataset now consists of 16,756 chest X-Rays across 13,645 patient cases. Instructions and scripts on generating this enhanced collection can be found on our GitHub repo.
We’d like to thank the following groups for making their data available as it accelerates efforts such as ours: the Radiological Society of North America and others involved in the RSNA Pneumonia Detection Challenge project, and Dr. Joseph Paul Cohen and the team at MILA involved in the COVID-19 image data collection project.
Additionally, a special thank you to the AI team at the City of London for their feedback on improving the COVIDx dataset.
2.) We’ve released two models based on the dataset above: COVIDNet-Large and COVIDNet-Small. The former provides higher detection sensitivity for COVID-19 detection, whereas the latter achieves a stronger balance between computational efficiency and performance (more suitable for edge-based deployment). Both are available at the GitHub repo.
Note these networks are intended to be used as reference models that can be built upon and enhanced as new data becomes available to seed new deep learning innovations in the fight against the COVID-19 pandemic. They are not yet intended as production-ready models, and we are working continuously to improve them as new data becomes available. COVID-Net is not intended for self-diagnosis. Please seek assistance from your local health authorities if you are presenting COVID-related symptoms.
3.) Our team is hard at work on COVID-RiskNet, a neural network tailored for COVID-19 risk stratification. Risk stratification can assist clinicians in customizing care based on risk levels by enabling personalized care planning and improving patient population management. Stay tuned as it will be made available shortly.
In order to improve the quality of predictions made by COVID-Net, we’d like to enhance our dataset to include at least 500 chest X-ray images of COVID-positive patients.
Please email us at info@darwinai.ca if you have data to share or you would like access to our explainability platform.
In Solidarity,
The DarwinAI team
