What is XRayCovid-19?
XRayCovid-19 is an ongoing project that uses Artificial Intelligence to assist the health system in the COVID-19 diagnostic process. It is characterized by easy use; efficiency in response time and effectiveness in the result.
The aim of this project is to provide a free access aid-to-diagnosis tool to assist professionals working in the health system to face the current pandemic more efficiently.
This is the first project developed by ATISLabs linked to the Department of Computer Science and the Postgraduate Program in Digital Humanities, both based on the Nova Iguaçu Campus, Multidisciplinary Institute of the Federal Rural University of Rio de Janeiro
Why radiographs?
Chest radiographs are low-cost tests, routinely performed in emergency care units, public and private hospitals. They can be obtained in a few hours and are important for the diagnosis of varied pneumonias, and the characteristics of the image may suggest infection caused by COVID-19.
Why to implement such a method?
The difficulty of carrying out massive testing is now a worldwide problem. Possible under-ascertainment in infections and deaths leads to inaccurate estimates, which makes it difficult to adjust governmental interventions and hindering decision-making by different levels of government. In this way, new systems that assist in rapid diagnosis can help to generate more reliable statistics.
The rapid screening of pneumonia cases resulting from COVID-19 in relation to those generated by other causes can also collaborate in the organization of care in the face of the advance of the pandemic.
Providing answers to problems of this nature is a major challenge for the Artificial Intelligence area, and this experience may open up perspectives for new advances in the field of Diagnostic Medicine even after the end of the current crisis.
Technicalities
We apply advanced tools (Julia language, Deep Learning, Flux framework) and the best available models (Convolutional Networks) to generate system predictions. Our diagnostic time is 55ms and our average accuracy is 92%.