NVIDIA Inception partner DarwinAI developed a new AI model to detect COVID-19 in CT scans with 96% accuracy across a wide and diverse number of scenarios. The model, COVID-Net CT-2, was built using a number of large and diverse datasets created over several months with the University of Waterloo and is publicly available on GitHub.
Last year, the startup launched an open source neural network for COVID-19 detection called COVID-Net. Their new model builds upon the initiative with a more robust model, trained on the largest quantity and diversity of multinational patient cases in research literature. Their academic study detailing the construction and validation of the model can be found here.
“By building COVID-NET CT-2 from such rich and voluminous data, we’ve been able to achieve a new level of screening accuracy, with COVID-19 sensitivity and positive predictive value exceeding 96% across a wide and diverse number of scenarios,” said Sheldon Fernandez, CEO of Darwin-AI.
“Our XAI platform was instrumental in the construction of the original COVID-Net model. For this version, we have engaged two senior radiologists in Canada to validate the way in which COVID-Net CT-2 makes its decisions. Much to our delight, both confirmed the decision-making process of COVID-Net CT-2 is consistent with their own expert interpretations,” Fernandez added. “In addition to illustrating the emerging cooperation between our respective domains, their validation exemplifies the importance of our XAI technology in constructing transparent and trustworthy AI.”
Both DarwinAI and NVIDIA are enabling researchers to build neural networks to fight COVID-19 with the help of open source AI and publicly available pre-trained models. Medical imaging AI models for detecting COVID-19 in X-rays and CTs can be accessed through DarwinAI’s COVID-Net Initiative and the NVIDIA COVID-19 NGC Catalog.
Learn more about how AI, accelerated computing, and GPU technology are contributing to the worldwide battle against the novel coronavirus in the COVID-19 Research Hub.