One of the best tools available to humans in the fight against infectious diseases is technology. Not only for the obvious, in relation to therapeutic techniques that help the survival of patients; but also in many other dimensions. In the fight against COVID-19, massive population monitoring and testing, rapid detection of symptoms such as fever and even the detection of effective drugs can be more easily done with the help of artificial intelligence.
Now, a new method based on big data and deep learning makes it possible to efficiently discern between healthy patients, those ill with pneumonia and those with COVID-19 from chest X-rays. The system has been developed by researchers from CVBLab, from the Polytechnic University of Valencia.
Valery Naranjo, UPV professor and CVBLab director, explains that the model can reach a success rate of an average of 92% when differentiating between the different classes of patients, and is especially effective when it comes to detect cases of coronavirus infection : “The algorithm performs even better in predicting cases of the SARS-CoV-2 coronavirus; its success rate is slightly higher in relation to the rest of the cases, with a 97% success rate to determine if the X-ray is of a patient with COVID-19 ”.
CVBLab engineers have applied classification and segmentation techniques, based on deep learning algorithms (a set of machine learning algorithms) on a large number of X-ray images ( big data ).
In this sense, Valery Naranjo explains that there are many more X-rays of healthy patients and other pneumonia than of COVID-19, “because of how recent it is and because many databases are not released”, which is an added difficulty: ” The model we have developed solves this imbalance in patient classes, and allows us to offer reliable and robust results ”.
How does a machine learn?
Thanks to the neural networks used by this technology, prediction occurs instantly. Scientists only have to identify a single X-ray of each type of patient: healthy, with pneumonia and with COVID-19. Once the patterns with which the machine must work have been introduced, the AI is to learn by itself ( machine learning ) to massively compare many radiographs with the data initially introduced by the researchers. Starting from a series of layers formed by neural networks ( deep learning ), AI is becoming more and more sophisticated and develops an increasingly high percentage of effectiveness.
Furthermore, given a new radiographic image, apart from obtaining a prediction about its diagnosis, the most similar preceding cases are automatically provided from a large, constantly growing database. The most similar areas of lung disease in the case history are presented using a very intuitive heat map for the expert personnel who use it. Thus, the doctor has more data when making a decision.
“It’s like when they search an atlas, but automatically”, exemplifies Adrián Colomer, doctor in telecommunications and researcher at the CVBLab of the UPV.
